multimodalart HF staff commited on
Commit
5a510e7
1 Parent(s): 20bf7e1

Upload folder using huggingface_hub

Browse files
.github/workflows/static-check.yaml ADDED
@@ -0,0 +1,26 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ name: Pylint
2
+
3
+ on: [push, pull_request]
4
+
5
+ jobs:
6
+ static-check:
7
+ runs-on: ${{ matrix.os }}
8
+ strategy:
9
+ matrix:
10
+ os: [ubuntu-22.04]
11
+ python-version: ["3.10"]
12
+ steps:
13
+ - uses: actions/checkout@v3
14
+ - name: Set up Python ${{ matrix.python-version }}
15
+ uses: actions/setup-python@v3
16
+ with:
17
+ python-version: ${{ matrix.python-version }}
18
+ - name: Install dependencies
19
+ run: |
20
+ python -m pip install --upgrade pylint
21
+ python -m pip install --upgrade isort
22
+ python -m pip install -r requirements.txt
23
+ - name: Analysing the code with pylint
24
+ run: |
25
+ isort $(git ls-files '*.py') --check-only --diff
26
+ pylint $(git ls-files '*.py')
.gitignore ADDED
@@ -0,0 +1,170 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # running cache
2
+ mlruns/
3
+
4
+ # Test directories
5
+ test_data/
6
+ pretrained_models/
7
+
8
+ # Poetry project
9
+ poetry.lock
10
+
11
+ # Byte-compiled / optimized / DLL files
12
+ __pycache__/
13
+ *.py[cod]
14
+ *$py.class
15
+
16
+ # C extensions
17
+ *.so
18
+
19
+ # Distribution / packaging
20
+ .Python
21
+ build/
22
+ develop-eggs/
23
+ dist/
24
+ downloads/
25
+ eggs/
26
+ .eggs/
27
+ lib/
28
+ lib64/
29
+ parts/
30
+ sdist/
31
+ var/
32
+ wheels/
33
+ share/python-wheels/
34
+ *.egg-info/
35
+ .installed.cfg
36
+ *.egg
37
+ MANIFEST
38
+
39
+ # PyInstaller
40
+ # Usually these files are written by a python script from a template
41
+ # before PyInstaller builds the exe, so as to inject date/other infos into it.
42
+ *.manifest
43
+ *.spec
44
+
45
+ # Installer logs
46
+ pip-log.txt
47
+ pip-delete-this-directory.txt
48
+
49
+ # Unit test / coverage reports
50
+ htmlcov/
51
+ .tox/
52
+ .nox/
53
+ .coverage
54
+ .coverage.*
55
+ .cache
56
+ nosetests.xml
57
+ coverage.xml
58
+ *.cover
59
+ *.py,cover
60
+ .hypothesis/
61
+ .pytest_cache/
62
+ cover/
63
+
64
+ # Translations
65
+ *.mo
66
+ *.pot
67
+
68
+ # Django stuff:
69
+ *.log
70
+ local_settings.py
71
+ db.sqlite3
72
+ db.sqlite3-journal
73
+
74
+ # Flask stuff:
75
+ instance/
76
+ .webassets-cache
77
+
78
+ # Scrapy stuff:
79
+ .scrapy
80
+
81
+ # Sphinx documentation
82
+ docs/_build/
83
+
84
+ # PyBuilder
85
+ .pybuilder/
86
+ target/
87
+
88
+ # Jupyter Notebook
89
+ .ipynb_checkpoints
90
+
91
+ # IPython
92
+ profile_default/
93
+ ipython_config.py
94
+
95
+ # pyenv
96
+ # For a library or package, you might want to ignore these files since the code is
97
+ # intended to run in multiple environments; otherwise, check them in:
98
+ # .python-version
99
+
100
+ # pipenv
101
+ # According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control.
102
+ # However, in case of collaboration, if having platform-specific dependencies or dependencies
103
+ # having no cross-platform support, pipenv may install dependencies that don't work, or not
104
+ # install all needed dependencies.
105
+ #Pipfile.lock
106
+
107
+ # poetry
108
+ # Similar to Pipfile.lock, it is generally recommended to include poetry.lock in version control.
109
+ # This is especially recommended for binary packages to ensure reproducibility, and is more
110
+ # commonly ignored for libraries.
111
+ # https://python-poetry.org/docs/basic-usage/#commit-your-poetrylock-file-to-version-control
112
+ #poetry.lock
113
+
114
+ # pdm
115
+ # Similar to Pipfile.lock, it is generally recommended to include pdm.lock in version control.
116
+ #pdm.lock
117
+ # pdm stores project-wide configurations in .pdm.toml, but it is recommended to not include it
118
+ # in version control.
119
+ # https://pdm.fming.dev/#use-with-ide
120
+ .pdm.toml
121
+
122
+ # PEP 582; used by e.g. github.com/David-OConnor/pyflow and github.com/pdm-project/pdm
123
+ __pypackages__/
124
+
125
+ # Celery stuff
126
+ celerybeat-schedule
127
+ celerybeat.pid
128
+
129
+ # SageMath parsed files
130
+ *.sage.py
131
+
132
+ # Environments
133
+ .env
134
+ .venv
135
+ env/
136
+ venv/
137
+ ENV/
138
+ env.bak/
139
+ venv.bak/
140
+
141
+ # Spyder project settings
142
+ .spyderproject
143
+ .spyproject
144
+
145
+ # Rope project settings
146
+ .ropeproject
147
+
148
+ # mkdocs documentation
149
+ /site
150
+
151
+ # mypy
152
+ .mypy_cache/
153
+ .dmypy.json
154
+ dmypy.json
155
+
156
+ # Pyre type checker
157
+ .pyre/
158
+
159
+ # pytype static type analyzer
160
+ .pytype/
161
+
162
+ # Cython debug symbols
163
+ cython_debug/
164
+
165
+ # IDE
166
+ .idea/
167
+ .vscode/
168
+ data
169
+ pretrained_models
170
+ test_data
.pre-commit-config.yaml ADDED
@@ -0,0 +1,17 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ repos:
2
+ - repo: local
3
+ hooks:
4
+ - id: isort
5
+ name: isort
6
+ language: system
7
+ types: [python]
8
+ pass_filenames: false
9
+ entry: isort
10
+ args: ["."]
11
+ - id: pylint
12
+ name: pylint
13
+ language: system
14
+ types: [python]
15
+ pass_filenames: false
16
+ entry: pylint
17
+ args: ["**/*.py"]
.pylintrc ADDED
@@ -0,0 +1,633 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [MAIN]
2
+
3
+ # Analyse import fallback blocks. This can be used to support both Python 2 and
4
+ # 3 compatible code, which means that the block might have code that exists
5
+ # only in one or another interpreter, leading to false positives when analysed.
6
+ analyse-fallback-blocks=no
7
+
8
+ # Clear in-memory caches upon conclusion of linting. Useful if running pylint
9
+ # in a server-like mode.
10
+ clear-cache-post-run=no
11
+
12
+ # Load and enable all available extensions. Use --list-extensions to see a list
13
+ # all available extensions.
14
+ #enable-all-extensions=
15
+
16
+ # In error mode, messages with a category besides ERROR or FATAL are
17
+ # suppressed, and no reports are done by default. Error mode is compatible with
18
+ # disabling specific errors.
19
+ #errors-only=
20
+
21
+ # Always return a 0 (non-error) status code, even if lint errors are found.
22
+ # This is primarily useful in continuous integration scripts.
23
+ #exit-zero=
24
+
25
+ # A comma-separated list of package or module names from where C extensions may
26
+ # be loaded. Extensions are loading into the active Python interpreter and may
27
+ # run arbitrary code.
28
+ extension-pkg-allow-list=
29
+
30
+ # A comma-separated list of package or module names from where C extensions may
31
+ # be loaded. Extensions are loading into the active Python interpreter and may
32
+ # run arbitrary code. (This is an alternative name to extension-pkg-allow-list
33
+ # for backward compatibility.)
34
+ extension-pkg-whitelist=cv2
35
+
36
+ # Return non-zero exit code if any of these messages/categories are detected,
37
+ # even if score is above --fail-under value. Syntax same as enable. Messages
38
+ # specified are enabled, while categories only check already-enabled messages.
39
+ fail-on=
40
+
41
+ # Specify a score threshold under which the program will exit with error.
42
+ fail-under=10
43
+
44
+ # Interpret the stdin as a python script, whose filename needs to be passed as
45
+ # the module_or_package argument.
46
+ #from-stdin=
47
+
48
+ # Files or directories to be skipped. They should be base names, not paths.
49
+ ignore=CVS
50
+
51
+ # Add files or directories matching the regular expressions patterns to the
52
+ # ignore-list. The regex matches against paths and can be in Posix or Windows
53
+ # format. Because '\\' represents the directory delimiter on Windows systems,
54
+ # it can't be used as an escape character.
55
+ ignore-paths=
56
+
57
+ # Files or directories matching the regular expression patterns are skipped.
58
+ # The regex matches against base names, not paths. The default value ignores
59
+ # Emacs file locks
60
+ ignore-patterns=^\.#
61
+
62
+ # List of module names for which member attributes should not be checked
63
+ # (useful for modules/projects where namespaces are manipulated during runtime
64
+ # and thus existing member attributes cannot be deduced by static analysis). It
65
+ # supports qualified module names, as well as Unix pattern matching.
66
+ ignored-modules=cv2
67
+
68
+ # Python code to execute, usually for sys.path manipulation such as
69
+ # pygtk.require().
70
+ init-hook='import sys; sys.path.append(".")'
71
+
72
+ # Use multiple processes to speed up Pylint. Specifying 0 will auto-detect the
73
+ # number of processors available to use, and will cap the count on Windows to
74
+ # avoid hangs.
75
+ jobs=1
76
+
77
+ # Control the amount of potential inferred values when inferring a single
78
+ # object. This can help the performance when dealing with large functions or
79
+ # complex, nested conditions.
80
+ limit-inference-results=100
81
+
82
+ # List of plugins (as comma separated values of python module names) to load,
83
+ # usually to register additional checkers.
84
+ load-plugins=
85
+
86
+ # Pickle collected data for later comparisons.
87
+ persistent=yes
88
+
89
+ # Minimum Python version to use for version dependent checks. Will default to
90
+ # the version used to run pylint.
91
+ py-version=3.10
92
+
93
+ # Discover python modules and packages in the file system subtree.
94
+ recursive=no
95
+
96
+ # Add paths to the list of the source roots. Supports globbing patterns. The
97
+ # source root is an absolute path or a path relative to the current working
98
+ # directory used to determine a package namespace for modules located under the
99
+ # source root.
100
+ source-roots=
101
+
102
+ # When enabled, pylint would attempt to guess common misconfiguration and emit
103
+ # user-friendly hints instead of false-positive error messages.
104
+ suggestion-mode=yes
105
+
106
+ # Allow loading of arbitrary C extensions. Extensions are imported into the
107
+ # active Python interpreter and may run arbitrary code.
108
+ unsafe-load-any-extension=no
109
+
110
+ # In verbose mode, extra non-checker-related info will be displayed.
111
+ #verbose=
112
+
113
+
114
+ [BASIC]
115
+
116
+ # Naming style matching correct argument names.
117
+ argument-naming-style=snake_case
118
+
119
+ # Regular expression matching correct argument names. Overrides argument-
120
+ # naming-style. If left empty, argument names will be checked with the set
121
+ # naming style.
122
+ #argument-rgx=
123
+
124
+ # Naming style matching correct attribute names.
125
+ attr-naming-style=snake_case
126
+
127
+ # Regular expression matching correct attribute names. Overrides attr-naming-
128
+ # style. If left empty, attribute names will be checked with the set naming
129
+ # style.
130
+ #attr-rgx=
131
+
132
+ # Bad variable names which should always be refused, separated by a comma.
133
+ bad-names=foo,
134
+ bar,
135
+ baz,
136
+ toto,
137
+ tutu,
138
+ tata
139
+
140
+ # Bad variable names regexes, separated by a comma. If names match any regex,
141
+ # they will always be refused
142
+ bad-names-rgxs=
143
+
144
+ # Naming style matching correct class attribute names.
145
+ class-attribute-naming-style=any
146
+
147
+ # Regular expression matching correct class attribute names. Overrides class-
148
+ # attribute-naming-style. If left empty, class attribute names will be checked
149
+ # with the set naming style.
150
+ #class-attribute-rgx=
151
+
152
+ # Naming style matching correct class constant names.
153
+ class-const-naming-style=UPPER_CASE
154
+
155
+ # Regular expression matching correct class constant names. Overrides class-
156
+ # const-naming-style. If left empty, class constant names will be checked with
157
+ # the set naming style.
158
+ #class-const-rgx=
159
+
160
+ # Naming style matching correct class names.
161
+ class-naming-style=PascalCase
162
+
163
+ # Regular expression matching correct class names. Overrides class-naming-
164
+ # style. If left empty, class names will be checked with the set naming style.
165
+ #class-rgx=
166
+
167
+ # Naming style matching correct constant names.
168
+ const-naming-style=UPPER_CASE
169
+
170
+ # Regular expression matching correct constant names. Overrides const-naming-
171
+ # style. If left empty, constant names will be checked with the set naming
172
+ # style.
173
+ #const-rgx=
174
+
175
+ # Minimum line length for functions/classes that require docstrings, shorter
176
+ # ones are exempt.
177
+ docstring-min-length=-1
178
+
179
+ # Naming style matching correct function names.
180
+ function-naming-style=snake_case
181
+
182
+ # Regular expression matching correct function names. Overrides function-
183
+ # naming-style. If left empty, function names will be checked with the set
184
+ # naming style.
185
+ #function-rgx=
186
+
187
+ # Good variable names which should always be accepted, separated by a comma.
188
+ good-names=i,
189
+ j,
190
+ k,
191
+ ex,
192
+ Run,
193
+ _
194
+
195
+ # Good variable names regexes, separated by a comma. If names match any regex,
196
+ # they will always be accepted
197
+ good-names-rgxs=
198
+
199
+ # Include a hint for the correct naming format with invalid-name.
200
+ include-naming-hint=no
201
+
202
+ # Naming style matching correct inline iteration names.
203
+ inlinevar-naming-style=any
204
+
205
+ # Regular expression matching correct inline iteration names. Overrides
206
+ # inlinevar-naming-style. If left empty, inline iteration names will be checked
207
+ # with the set naming style.
208
+ #inlinevar-rgx=
209
+
210
+ # Naming style matching correct method names.
211
+ method-naming-style=snake_case
212
+
213
+ # Regular expression matching correct method names. Overrides method-naming-
214
+ # style. If left empty, method names will be checked with the set naming style.
215
+ #method-rgx=
216
+
217
+ # Naming style matching correct module names.
218
+ module-naming-style=snake_case
219
+
220
+ # Regular expression matching correct module names. Overrides module-naming-
221
+ # style. If left empty, module names will be checked with the set naming style.
222
+ #module-rgx=
223
+
224
+ # Colon-delimited sets of names that determine each other's naming style when
225
+ # the name regexes allow several styles.
226
+ name-group=
227
+
228
+ # Regular expression which should only match function or class names that do
229
+ # not require a docstring.
230
+ no-docstring-rgx=^_
231
+
232
+ # List of decorators that produce properties, such as abc.abstractproperty. Add
233
+ # to this list to register other decorators that produce valid properties.
234
+ # These decorators are taken in consideration only for invalid-name.
235
+ property-classes=abc.abstractproperty
236
+
237
+ # Regular expression matching correct type alias names. If left empty, type
238
+ # alias names will be checked with the set naming style.
239
+ #typealias-rgx=
240
+
241
+ # Regular expression matching correct type variable names. If left empty, type
242
+ # variable names will be checked with the set naming style.
243
+ #typevar-rgx=
244
+
245
+ # Naming style matching correct variable names.
246
+ variable-naming-style=snake_case
247
+
248
+ # Regular expression matching correct variable names. Overrides variable-
249
+ # naming-style. If left empty, variable names will be checked with the set
250
+ # naming style.
251
+ variable-rgx=(_?[a-z][A-Za-z0-9]{0,30})|([A-Z0-9]{1,30})
252
+
253
+
254
+ [CLASSES]
255
+
256
+ # Warn about protected attribute access inside special methods
257
+ check-protected-access-in-special-methods=no
258
+
259
+ # List of method names used to declare (i.e. assign) instance attributes.
260
+ defining-attr-methods=__init__,
261
+ __new__,
262
+ setUp,
263
+ asyncSetUp,
264
+ __post_init__
265
+
266
+ # List of member names, which should be excluded from the protected access
267
+ # warning.
268
+ exclude-protected=_asdict,_fields,_replace,_source,_make,os._exit
269
+
270
+ # List of valid names for the first argument in a class method.
271
+ valid-classmethod-first-arg=cls
272
+
273
+ # List of valid names for the first argument in a metaclass class method.
274
+ valid-metaclass-classmethod-first-arg=mcs
275
+
276
+
277
+ [DESIGN]
278
+
279
+ # List of regular expressions of class ancestor names to ignore when counting
280
+ # public methods (see R0903)
281
+ exclude-too-few-public-methods=
282
+
283
+ # List of qualified class names to ignore when counting class parents (see
284
+ # R0901)
285
+ ignored-parents=
286
+
287
+ # Maximum number of arguments for function / method.
288
+ max-args=7
289
+
290
+ # Maximum number of attributes for a class (see R0902).
291
+ max-attributes=20
292
+
293
+ # Maximum number of boolean expressions in an if statement (see R0916).
294
+ max-bool-expr=5
295
+
296
+ # Maximum number of branch for function / method body.
297
+ max-branches=12
298
+
299
+ # Maximum number of locals for function / method body.
300
+ max-locals=15
301
+
302
+ # Maximum number of parents for a class (see R0901).
303
+ max-parents=7
304
+
305
+ # Maximum number of public methods for a class (see R0904).
306
+ max-public-methods=20
307
+
308
+ # Maximum number of return / yield for function / method body.
309
+ max-returns=6
310
+
311
+ # Maximum number of statements in function / method body.
312
+ max-statements=300
313
+
314
+ # Minimum number of public methods for a class (see R0903).
315
+ min-public-methods=1
316
+
317
+
318
+ [EXCEPTIONS]
319
+
320
+ # Exceptions that will emit a warning when caught.
321
+ overgeneral-exceptions=builtins.BaseException,builtins.Exception
322
+
323
+
324
+ [FORMAT]
325
+
326
+ # Expected format of line ending, e.g. empty (any line ending), LF or CRLF.
327
+ expected-line-ending-format=
328
+
329
+ # Regexp for a line that is allowed to be longer than the limit.
330
+ ignore-long-lines=^\s*(# )?<?https?://\S+>?$
331
+
332
+ # Number of spaces of indent required inside a hanging or continued line.
333
+ indent-after-paren=4
334
+
335
+ # String used as indentation unit. This is usually " " (4 spaces) or "\t" (1
336
+ # tab).
337
+ indent-string=' '
338
+
339
+ # Maximum number of characters on a single line.
340
+ max-line-length=150
341
+
342
+ # Maximum number of lines in a module.
343
+ max-module-lines=2000
344
+
345
+ # Allow the body of a class to be on the same line as the declaration if body
346
+ # contains single statement.
347
+ single-line-class-stmt=no
348
+
349
+ # Allow the body of an if to be on the same line as the test if there is no
350
+ # else.
351
+ single-line-if-stmt=no
352
+
353
+
354
+ [IMPORTS]
355
+
356
+ # List of modules that can be imported at any level, not just the top level
357
+ # one.
358
+ allow-any-import-level=
359
+
360
+ # Allow explicit reexports by alias from a package __init__.
361
+ allow-reexport-from-package=no
362
+
363
+ # Allow wildcard imports from modules that define __all__.
364
+ allow-wildcard-with-all=no
365
+
366
+ # Deprecated modules which should not be used, separated by a comma.
367
+ deprecated-modules=
368
+
369
+ # Output a graph (.gv or any supported image format) of external dependencies
370
+ # to the given file (report RP0402 must not be disabled).
371
+ ext-import-graph=
372
+
373
+ # Output a graph (.gv or any supported image format) of all (i.e. internal and
374
+ # external) dependencies to the given file (report RP0402 must not be
375
+ # disabled).
376
+ import-graph=
377
+
378
+ # Output a graph (.gv or any supported image format) of internal dependencies
379
+ # to the given file (report RP0402 must not be disabled).
380
+ int-import-graph=
381
+
382
+ # Force import order to recognize a module as part of the standard
383
+ # compatibility libraries.
384
+ known-standard-library=
385
+
386
+ # Force import order to recognize a module as part of a third party library.
387
+ known-third-party=enchant
388
+
389
+ # Couples of modules and preferred modules, separated by a comma.
390
+ preferred-modules=
391
+
392
+
393
+ [LOGGING]
394
+
395
+ # The type of string formatting that logging methods do. `old` means using %
396
+ # formatting, `new` is for `{}` formatting.
397
+ logging-format-style=old
398
+
399
+ # Logging modules to check that the string format arguments are in logging
400
+ # function parameter format.
401
+ logging-modules=logging
402
+
403
+
404
+ [MESSAGES CONTROL]
405
+
406
+ # Only show warnings with the listed confidence levels. Leave empty to show
407
+ # all. Valid levels: HIGH, CONTROL_FLOW, INFERENCE, INFERENCE_FAILURE,
408
+ # UNDEFINED.
409
+ confidence=HIGH,
410
+ CONTROL_FLOW,
411
+ INFERENCE,
412
+ INFERENCE_FAILURE,
413
+ UNDEFINED
414
+
415
+ # Disable the message, report, category or checker with the given id(s). You
416
+ # can either give multiple identifiers separated by comma (,) or put this
417
+ # option multiple times (only on the command line, not in the configuration
418
+ # file where it should appear only once). You can also use "--disable=all" to
419
+ # disable everything first and then re-enable specific checks. For example, if
420
+ # you want to run only the similarities checker, you can use "--disable=all
421
+ # --enable=similarities". If you want to run only the classes checker, but have
422
+ # no Warning level messages displayed, use "--disable=all --enable=classes
423
+ # --disable=W".
424
+ disable=too-many-arguments,
425
+ too-many-locals,
426
+ too-many-branches,
427
+ protected-access
428
+
429
+
430
+ # Enable the message, report, category or checker with the given id(s). You can
431
+ # either give multiple identifier separated by comma (,) or put this option
432
+ # multiple time (only on the command line, not in the configuration file where
433
+ # it should appear only once). See also the "--disable" option for examples.
434
+ enable=
435
+
436
+
437
+ [METHOD_ARGS]
438
+
439
+ # List of qualified names (i.e., library.method) which require a timeout
440
+ # parameter e.g. 'requests.api.get,requests.api.post'
441
+ timeout-methods=requests.api.delete,requests.api.get,requests.api.head,requests.api.options,requests.api.patch,requests.api.post,requests.api.put,requests.api.request
442
+
443
+
444
+ [MISCELLANEOUS]
445
+
446
+ # List of note tags to take in consideration, separated by a comma.
447
+ notes=FIXME,
448
+ XXX
449
+
450
+ # Regular expression of note tags to take in consideration.
451
+ notes-rgx=
452
+
453
+
454
+ [REFACTORING]
455
+
456
+ # Maximum number of nested blocks for function / method body
457
+ max-nested-blocks=5
458
+
459
+ # Complete name of functions that never returns. When checking for
460
+ # inconsistent-return-statements if a never returning function is called then
461
+ # it will be considered as an explicit return statement and no message will be
462
+ # printed.
463
+ never-returning-functions=sys.exit,argparse.parse_error
464
+
465
+ # Let 'consider-using-join' be raised when the separator to join on would be
466
+ # non-empty (resulting in expected fixes of the type: ``"- " + " -
467
+ # ".join(items)``)
468
+ # suggest-join-with-non-empty-separator=yes
469
+
470
+
471
+ [REPORTS]
472
+
473
+ # Python expression which should return a score less than or equal to 10. You
474
+ # have access to the variables 'fatal', 'error', 'warning', 'refactor',
475
+ # 'convention', and 'info' which contain the number of messages in each
476
+ # category, as well as 'statement' which is the total number of statements
477
+ # analyzed. This score is used by the global evaluation report (RP0004).
478
+ evaluation=max(0, 0 if fatal else 10.0 - ((float(5 * error + warning + refactor + convention) / statement) * 10))
479
+
480
+ # Template used to display messages. This is a python new-style format string
481
+ # used to format the message information. See doc for all details.
482
+ msg-template=
483
+
484
+ # Set the output format. Available formats are: text, parseable, colorized,
485
+ # json2 (improved json format), json (old json format) and msvs (visual
486
+ # studio). You can also give a reporter class, e.g.
487
+ # mypackage.mymodule.MyReporterClass.
488
+ #output-format=
489
+
490
+ # Tells whether to display a full report or only the messages.
491
+ reports=no
492
+
493
+ # Activate the evaluation score.
494
+ score=yes
495
+
496
+
497
+ [SIMILARITIES]
498
+
499
+ # Comments are removed from the similarity computation
500
+ ignore-comments=yes
501
+
502
+ # Docstrings are removed from the similarity computation
503
+ ignore-docstrings=yes
504
+
505
+ # Imports are removed from the similarity computation
506
+ ignore-imports=yes
507
+
508
+ # Signatures are removed from the similarity computation
509
+ ignore-signatures=yes
510
+
511
+ # Minimum lines number of a similarity.
512
+ min-similarity-lines=4
513
+
514
+
515
+ [SPELLING]
516
+
517
+ # Limits count of emitted suggestions for spelling mistakes.
518
+ max-spelling-suggestions=4
519
+
520
+ # Spelling dictionary name. No available dictionaries : You need to install
521
+ # both the python package and the system dependency for enchant to work.
522
+ spelling-dict=
523
+
524
+ # List of comma separated words that should be considered directives if they
525
+ # appear at the beginning of a comment and should not be checked.
526
+ spelling-ignore-comment-directives=fmt: on,fmt: off,noqa:,noqa,nosec,isort:skip,mypy:
527
+
528
+ # List of comma separated words that should not be checked.
529
+ spelling-ignore-words=
530
+
531
+ # A path to a file that contains the private dictionary; one word per line.
532
+ spelling-private-dict-file=
533
+
534
+ # Tells whether to store unknown words to the private dictionary (see the
535
+ # --spelling-private-dict-file option) instead of raising a message.
536
+ spelling-store-unknown-words=no
537
+
538
+
539
+ [STRING]
540
+
541
+ # This flag controls whether inconsistent-quotes generates a warning when the
542
+ # character used as a quote delimiter is used inconsistently within a module.
543
+ check-quote-consistency=no
544
+
545
+ # This flag controls whether the implicit-str-concat should generate a warning
546
+ # on implicit string concatenation in sequences defined over several lines.
547
+ check-str-concat-over-line-jumps=no
548
+
549
+
550
+ [TYPECHECK]
551
+
552
+ # List of decorators that produce context managers, such as
553
+ # contextlib.contextmanager. Add to this list to register other decorators that
554
+ # produce valid context managers.
555
+ contextmanager-decorators=contextlib.contextmanager
556
+
557
+ # List of members which are set dynamically and missed by pylint inference
558
+ # system, and so shouldn't trigger E1101 when accessed. Python regular
559
+ # expressions are accepted.
560
+ generated-members=
561
+
562
+ # Tells whether to warn about missing members when the owner of the attribute
563
+ # is inferred to be None.
564
+ ignore-none=yes
565
+
566
+ # This flag controls whether pylint should warn about no-member and similar
567
+ # checks whenever an opaque object is returned when inferring. The inference
568
+ # can return multiple potential results while evaluating a Python object, but
569
+ # some branches might not be evaluated, which results in partial inference. In
570
+ # that case, it might be useful to still emit no-member and other checks for
571
+ # the rest of the inferred objects.
572
+ ignore-on-opaque-inference=yes
573
+
574
+ # List of symbolic message names to ignore for Mixin members.
575
+ ignored-checks-for-mixins=no-member,
576
+ not-async-context-manager,
577
+ not-context-manager,
578
+ attribute-defined-outside-init
579
+
580
+ # List of class names for which member attributes should not be checked (useful
581
+ # for classes with dynamically set attributes). This supports the use of
582
+ # qualified names.
583
+ ignored-classes=optparse.Values,thread._local,_thread._local,argparse.Namespace
584
+
585
+ # Show a hint with possible names when a member name was not found. The aspect
586
+ # of finding the hint is based on edit distance.
587
+ missing-member-hint=yes
588
+
589
+ # The minimum edit distance a name should have in order to be considered a
590
+ # similar match for a missing member name.
591
+ missing-member-hint-distance=1
592
+
593
+ # The total number of similar names that should be taken in consideration when
594
+ # showing a hint for a missing member.
595
+ missing-member-max-choices=1
596
+
597
+ # Regex pattern to define which classes are considered mixins.
598
+ mixin-class-rgx=.*[Mm]ixin
599
+
600
+ # List of decorators that change the signature of a decorated function.
601
+ signature-mutators=
602
+
603
+
604
+ [VARIABLES]
605
+
606
+ # List of additional names supposed to be defined in builtins. Remember that
607
+ # you should avoid defining new builtins when possible.
608
+ additional-builtins=
609
+
610
+ # Tells whether unused global variables should be treated as a violation.
611
+ allow-global-unused-variables=yes
612
+
613
+ # List of names allowed to shadow builtins
614
+ allowed-redefined-builtins=
615
+
616
+ # List of strings which can identify a callback function by name. A callback
617
+ # name must start or end with one of those strings.
618
+ callbacks=cb_,
619
+ _cb
620
+
621
+ # A regular expression matching the name of dummy variables (i.e. expected to
622
+ # not be used).
623
+ dummy-variables-rgx=_+$|(_[a-zA-Z0-9_]*[a-zA-Z0-9]+?$)|dummy|^ignored_|^unused_
624
+
625
+ # Argument names that match this expression will be ignored.
626
+ ignored-argument-names=_.*|^ignored_|^unused_
627
+
628
+ # Tells whether we should check for unused import in __init__ files.
629
+ init-import=no
630
+
631
+ # List of qualified module names which can have objects that can redefine
632
+ # builtins.
633
+ redefining-builtins-modules=six.moves,past.builtins,future.builtins,builtins,io
README.md CHANGED
@@ -1,12 +1,209 @@
1
- ---
2
- title: Hallo
3
- emoji: 👁
4
- colorFrom: green
5
- colorTo: green
6
- sdk: gradio
7
- sdk_version: 4.36.1
8
- app_file: app.py
9
- pinned: false
10
- ---
11
-
12
- Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ <h1 align='Center'>Hallo: Hierarchical Audio-Driven Visual Synthesis for Portrait Image Animation</h1>
2
+
3
+ <div align='Center'>
4
+ <a href='https://github.com/xumingw' target='_blank'>Mingwang Xu</a><sup>1*</sup>&emsp;
5
+ <a href='https://github.com/crystallee-ai' target='_blank'>Hui Li</a><sup>1*</sup>&emsp;
6
+ <a href='https://github.com/subazinga' target='_blank'>Qingkun Su</a><sup>1*</sup>&emsp;
7
+ <a href='https://github.com/NinoNeumann' target='_blank'>Hanlin Shang</a><sup>1</sup>&emsp;
8
+ <a href='https://github.com/AricGamma' target='_blank'>Liwei Zhang</a><sup>1</sup>&emsp;
9
+ <a href='https://github.com/cnexah' target='_blank'>Ce Liu</a><sup>3</sup>&emsp;
10
+ </div>
11
+ <div align='center'>
12
+ <a href='https://jingdongwang2017.github.io/' target='_blank'>Jingdong Wang</a><sup>2</sup>&emsp;
13
+ <a href='https://yoyo000.github.io/' target='_blank'>Yao Yao</a><sup>4</sup>&emsp;
14
+ <a href='https://sites.google.com/site/zhusiyucs/home' target='_blank'>Siyu Zhu</a><sup>1</sup>&emsp;
15
+ </div>
16
+
17
+ <div align='Center'>
18
+ <sup>1</sup>Fudan University&emsp; <sup>2</sup>Baidu Inc&emsp; <sup>3</sup>ETH Zurich&emsp; <sup>4</sup>Nanjing University
19
+ </div>
20
+
21
+ <br>
22
+ <div align='Center'>
23
+ <a href='https://github.com/fudan-generative-vision/hallo'><img src='https://img.shields.io/github/stars/fudan-generative-vision/hallo?style=social'></a>
24
+ <a href='https://fudan-generative-vision.github.io/hallo/#/'><img src='https://img.shields.io/badge/Project-HomePage-Green'></a>
25
+ <a href='https://arxiv.org/pdf/2406.08801'><img src='https://img.shields.io/badge/Paper-Arxiv-red'></a>
26
+ <a href='https://huggingface.co/fudan-generative-ai/hallo'><img src='https://img.shields.io/badge/%F0%9F%A4%97%20HuggingFace-Model-yellow'></a>
27
+ <a href='assets/wechat.jpeg'><img src='https://badges.aleen42.com/src/wechat.svg'></a>
28
+ </div>
29
+
30
+ <br>
31
+
32
+ # Showcase
33
+
34
+
35
+ https://github.com/fudan-generative-vision/hallo/assets/17402682/294e78ef-c60d-4c32-8e3c-7f8d6934c6bd
36
+
37
+
38
+ # Framework
39
+
40
+ ![abstract](assets/framework_1.jpg)
41
+ ![framework](assets/framework_2.jpg)
42
+
43
+ # News
44
+
45
+ - **`2024/06/15`**: 🎉🎉🎉 Release the first version on [GitHub](https://github.com/fudan-generative-vision/hallo).
46
+ - **`2024/06/15`**: ✨✨✨ Release some images and audios for inference testing on [Huggingface](https://huggingface.co/datasets/fudan-generative-ai/hallo_inference_samples).
47
+
48
+ # Installation
49
+
50
+ - System requirement: Ubuntu 20.04/Ubuntu 22.04, Cuda 12.1
51
+ - Tested GPUs: A100
52
+
53
+ Create conda environment:
54
+
55
+ ```bash
56
+ conda create -n hallo python=3.10
57
+ conda activate hallo
58
+ ```
59
+
60
+ Install packages with `pip`
61
+
62
+ ```bash
63
+ pip install -r requirements.txt
64
+ pip install .
65
+ ```
66
+
67
+ Besides, ffmpeg is also need:
68
+ ```bash
69
+ apt-get install ffmpeg
70
+ ```
71
+
72
+ # Inference
73
+
74
+ The inference entrypoint script is `scripts/inference.py`. Before testing your cases, there are two preparations need to be completed:
75
+
76
+ 1. [Download all required pretrained models](#download-pretrained-models).
77
+ 2. [Run inference](#run-inference).
78
+
79
+ ## Download pretrained models
80
+
81
+ You can easily get all pretrained models required by inference from our [HuggingFace repo](https://huggingface.co/fudan-generative-ai/hallo).
82
+
83
+ Clone the the pretrained models into `${PROJECT_ROOT}/pretrained_models` directory by cmd below:
84
+
85
+ ```shell
86
+ git lfs install
87
+ git clone https://huggingface.co/fudan-generative-ai/hallo pretrained_models
88
+ ```
89
+
90
+ Or you can download them separately from their source repo:
91
+
92
+ - [hallo](https://huggingface.co/fudan-generative-ai/hallo/tree/main/hallo): Our checkpoints consist of denoising UNet, face locator, image & audio proj.
93
+ - [audio_separator](https://huggingface.co/huangjackson/Kim_Vocal_2): Kim\_Vocal\_2 MDX-Net vocal removal model by [KimberleyJensen](https://github.com/KimberleyJensen). (_Thanks to runwayml_)
94
+ - [insightface](https://github.com/deepinsight/insightface/tree/master/python-package#model-zoo): 2D and 3D Face Analysis placed into `pretrained_models/face_analysis/models/`. (_Thanks to deepinsight_)
95
+ - [face landmarker](https://storage.googleapis.com/mediapipe-models/face_landmarker/face_landmarker/float16/1/face_landmarker.task): Face detection & mesh model from [mediapipe](https://ai.google.dev/edge/mediapipe/solutions/vision/face_landmarker#models) placed into `pretrained_models/face_analysis/models`.
96
+ - [motion module](https://github.com/guoyww/AnimateDiff/blob/main/README.md#202309-animatediff-v2): motion module from [AnimateDiff](https://github.com/guoyww/AnimateDiff). (_Thanks to guoyww_).
97
+ - [sd-vae-ft-mse](https://huggingface.co/stabilityai/sd-vae-ft-mse): Weights are intended to be used with the diffusers library. (_Thanks to stablilityai_)
98
+ - [StableDiffusion V1.5](https://huggingface.co/runwayml/stable-diffusion-v1-5): Initialized and fine-tuned from Stable-Diffusion-v1-2. (_Thanks to runwayml_)
99
+ - [wav2vec](https://huggingface.co/facebook/wav2vec2-base-960h): wav audio to vector model from [Facebook](https://huggingface.co/facebook/wav2vec2-base-960h).
100
+
101
+ Finally, these pretrained models should be organized as follows:
102
+
103
+ ```text
104
+ ./pretrained_models/
105
+ |-- audio_separator/
106
+ | `-- Kim_Vocal_2.onnx
107
+ |-- face_analysis/
108
+ | `-- models/
109
+ | |-- face_landmarker_v2_with_blendshapes.task # face landmarker model from mediapipe
110
+ | |-- 1k3d68.onnx
111
+ | |-- 2d106det.onnx
112
+ | |-- genderage.onnx
113
+ | |-- glintr100.onnx
114
+ | `-- scrfd_10g_bnkps.onnx
115
+ |-- motion_module/
116
+ | `-- mm_sd_v15_v2.ckpt
117
+ |-- sd-vae-ft-mse/
118
+ | |-- config.json
119
+ | `-- diffusion_pytorch_model.safetensors
120
+ |-- stable-diffusion-v1-5/
121
+ | |-- feature_extractor/
122
+ | | `-- preprocessor_config.json
123
+ | |-- model_index.json
124
+ | |-- unet/
125
+ | | |-- config.json
126
+ | | `-- diffusion_pytorch_model.safetensors
127
+ | `-- v1-inference.yaml
128
+ `-- wav2vec/
129
+ |-- wav2vec2-base-960h/
130
+ | |-- config.json
131
+ | |-- feature_extractor_config.json
132
+ | |-- model.safetensors
133
+ | |-- preprocessor_config.json
134
+ | |-- special_tokens_map.json
135
+ | |-- tokenizer_config.json
136
+ | `-- vocab.json
137
+ ```
138
+
139
+ ## Run inference
140
+
141
+ Simply to run the `scripts/inference.py` and pass `source_image` and `driving_audio` as input:
142
+
143
+ ```bash
144
+ python scripts/inference.py --source_image your_image.png --driving_audio your_audio.wav
145
+ ```
146
+
147
+ Animation results will be saved as `${PROJECT_ROOT}/.cache/output.mp4` by default. You can pass `--output` to specify the output file name.
148
+
149
+ For more options:
150
+
151
+ ```shell
152
+ usage: inference.py [-h] [-c CONFIG] [--source_image SOURCE_IMAGE] [--driving_audio DRIVING_AUDIO] [--output OUTPUT] [--pose_weight POSE_WEIGHT]
153
+ [--face_weight FACE_WEIGHT] [--lip_weight LIP_WEIGHT] [--face_expand_ratio FACE_EXPAND_RATIO]
154
+
155
+ options:
156
+ -h, --help show this help message and exit
157
+ -c CONFIG, --config CONFIG
158
+ --source_image SOURCE_IMAGE
159
+ source image
160
+ --driving_audio DRIVING_AUDIO
161
+ driving audio
162
+ --output OUTPUT output video file name
163
+ --pose_weight POSE_WEIGHT
164
+ weight of pose
165
+ --face_weight FACE_WEIGHT
166
+ weight of face
167
+ --lip_weight LIP_WEIGHT
168
+ weight of lip
169
+ --face_expand_ratio FACE_EXPAND_RATIO
170
+ face region
171
+ ```
172
+
173
+ # Roadmap
174
+
175
+ | Status | Milestone | ETA |
176
+ | :----: | :---------------------------------------------------------------------------------------------------- | :--------: |
177
+ | ✅ | **[Inference source code meet everyone on GitHub](https://github.com/fudan-generative-vision/hallo)** | 2024-06-15 |
178
+ | ✅ | **[Pretrained models on Huggingface](https://huggingface.co/fudan-generative-ai/hallo)** | 2024-06-15 |
179
+ | 🚀🚀🚀 | **[Traning: data preparation and training scripts]()** | 2024-06-25 |
180
+
181
+ # Citation
182
+
183
+ If you find our work useful for your research, please consider citing the paper:
184
+
185
+ ```
186
+ @misc{xu2024hallo,
187
+ title={Hallo: Hierarchical Audio-Driven Visual Synthesis for Portrait Image Animation},
188
+ author={Mingwang Xu and Hui Li and Qingkun Su and Hanlin Shang and Liwei Zhang and Ce Liu and Jingdong Wang and Yao Yao and Siyu zhu},
189
+ year={2024},
190
+ eprint={2406.08801},
191
+ archivePrefix={arXiv},
192
+ primaryClass={cs.CV}
193
+ }
194
+ ```
195
+
196
+ # Opportunities available
197
+
198
+ Multiple research positions are open at the **Generative Vision Lab, Fudan University**! Include:
199
+
200
+ - Research assistant
201
+ - Postdoctoral researcher
202
+ - PhD candidate
203
+ - Master students
204
+
205
+ Interested individuals are encouraged to contact us at [siyuzhu@fudan.edu.cn](mailto://siyuzhu@fudan.edu.cn) for further information.
206
+
207
+ # Social Risks and Mitigations
208
+
209
+ The development of portrait image animation technologies driven by audio inputs poses social risks, such as the ethical implications of creating realistic portraits that could be misused for deepfakes. To mitigate these risks, it is crucial to establish ethical guidelines and responsible use practices. Privacy and consent concerns also arise from using individuals' images and voices. Addressing these involves transparent data usage policies, informed consent, and safeguarding privacy rights. By addressing these risks and implementing mitigations, the research aims to ensure the responsible and ethical development of this technology.
assets/framework.png ADDED
assets/framework_1.jpg ADDED
assets/framework_2.jpg ADDED
assets/wechat.jpeg ADDED
configs/inference/.gitkeep ADDED
File without changes
configs/inference/default.yaml ADDED
@@ -0,0 +1,90 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ source_image: ./default.png
2
+ driving_audio: default.wav
3
+
4
+ weight_dtype: fp16
5
+
6
+ data:
7
+ n_motion_frames: 2
8
+ n_sample_frames: 16
9
+ source_image:
10
+ width: 512
11
+ height: 512
12
+ driving_audio:
13
+ sample_rate: 16000
14
+ export_video:
15
+ fps: 25
16
+
17
+ inference_steps: 40
18
+ cfg_scale: 3.5
19
+
20
+ audio_ckpt_dir: ./pretrained_models/hallo
21
+
22
+ base_model_path: ./pretrained_models/stable-diffusion-v1-5
23
+
24
+ motion_module_path: ./pretrained_models/motion_module/mm_sd_v15_v2.ckpt
25
+
26
+ face_analysis:
27
+ model_path: ./pretrained_models/face_analysis
28
+
29
+ wav2vec:
30
+ model_path: ./pretrained_models/wav2vec/wav2vec2-base-960h
31
+ features: all
32
+
33
+ audio_separator:
34
+ model_path: ./pretrained_models/audio_separator/Kim_Vocal_2.onnx
35
+
36
+ vae:
37
+ model_path: ./pretrained_models/sd-vae-ft-mse
38
+
39
+ save_path: ./.cache
40
+
41
+ face_expand_ratio: 1.1
42
+ pose_weight: 1.1
43
+ face_weight: 1.1
44
+ lip_weight: 1.1
45
+
46
+ unet_additional_kwargs:
47
+ use_inflated_groupnorm: true
48
+ unet_use_cross_frame_attention: false
49
+ unet_use_temporal_attention: false
50
+ use_motion_module: true
51
+ use_audio_module: true
52
+ motion_module_resolutions:
53
+ - 1
54
+ - 2
55
+ - 4
56
+ - 8
57
+ motion_module_mid_block: true
58
+ motion_module_decoder_only: false
59
+ motion_module_type: Vanilla
60
+ motion_module_kwargs:
61
+ num_attention_heads: 8
62
+ num_transformer_block: 1
63
+ attention_block_types:
64
+ - Temporal_Self
65
+ - Temporal_Self
66
+ temporal_position_encoding: true
67
+ temporal_position_encoding_max_len: 32
68
+ temporal_attention_dim_div: 1
69
+ audio_attention_dim: 768
70
+ stack_enable_blocks_name:
71
+ - "up"
72
+ - "down"
73
+ - "mid"
74
+ stack_enable_blocks_depth: [0,1,2,3]
75
+
76
+
77
+ enable_zero_snr: true
78
+
79
+ noise_scheduler_kwargs:
80
+ beta_start: 0.00085
81
+ beta_end: 0.012
82
+ beta_schedule: "linear"
83
+ clip_sample: false
84
+ steps_offset: 1
85
+ ### Zero-SNR params
86
+ prediction_type: "v_prediction"
87
+ rescale_betas_zero_snr: True
88
+ timestep_spacing: "trailing"
89
+
90
+ sampler: DDIM
configs/unet/unet.yaml ADDED
@@ -0,0 +1,44 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ unet_additional_kwargs:
2
+ use_inflated_groupnorm: true
3
+ unet_use_cross_frame_attention: false
4
+ unet_use_temporal_attention: false
5
+ use_motion_module: true
6
+ use_audio_module: true
7
+ motion_module_resolutions:
8
+ - 1
9
+ - 2
10
+ - 4
11
+ - 8
12
+ motion_module_mid_block: true
13
+ motion_module_decoder_only: false
14
+ motion_module_type: Vanilla
15
+ motion_module_kwargs:
16
+ num_attention_heads: 8
17
+ num_transformer_block: 1
18
+ attention_block_types:
19
+ - Temporal_Self
20
+ - Temporal_Self
21
+ temporal_position_encoding: true
22
+ temporal_position_encoding_max_len: 32
23
+ temporal_attention_dim_div: 1
24
+ audio_attention_dim: 768
25
+ stack_enable_blocks_name:
26
+ - "up"
27
+ - "down"
28
+ - "mid"
29
+ stack_enable_blocks_depth: [0,1,2,3]
30
+
31
+ enable_zero_snr: true
32
+
33
+ noise_scheduler_kwargs:
34
+ beta_start: 0.00085
35
+ beta_end: 0.012
36
+ beta_schedule: "linear"
37
+ clip_sample: false
38
+ steps_offset: 1
39
+ ### Zero-SNR params
40
+ prediction_type: "v_prediction"
41
+ rescale_betas_zero_snr: True
42
+ timestep_spacing: "trailing"
43
+
44
+ sampler: DDIM
hallo/__init__.py ADDED
File without changes
hallo/animate/__init__.py ADDED
File without changes
hallo/animate/face_animate.py ADDED
@@ -0,0 +1,442 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # pylint: disable=R0801
2
+ """
3
+ This module is responsible for animating faces in videos using a combination of deep learning techniques.
4
+ It provides a pipeline for generating face animations by processing video frames and extracting face features.
5
+ The module utilizes various schedulers and utilities for efficient face animation and supports different types
6
+ of latents for more control over the animation process.
7
+
8
+ Functions and Classes:
9
+ - FaceAnimatePipeline: A class that extends the DiffusionPipeline class from the diffusers library to handle face animation tasks.
10
+ - __init__: Initializes the pipeline with the necessary components (VAE, UNets, face locator, etc.).
11
+ - prepare_latents: Generates or loads latents for the animation process, scaling them according to the scheduler's requirements.
12
+ - prepare_extra_step_kwargs: Prepares extra keyword arguments for the scheduler step, ensuring compatibility with different schedulers.
13
+ - decode_latents: Decodes the latents into video frames, ready for animation.
14
+
15
+ Usage:
16
+ - Import the necessary packages and classes.
17
+ - Create a FaceAnimatePipeline instance with the required components.
18
+ - Prepare the latents for the animation process.
19
+ - Use the pipeline to generate the animated video.
20
+
21
+ Note:
22
+ - This module is designed to work with the diffusers library, which provides the underlying framework for face animation using deep learning.
23
+ - The module is intended for research and development purposes, and further optimization and customization may be required for specific use cases.
24
+ """
25
+
26
+ import inspect
27
+ from dataclasses import dataclass
28
+ from typing import Callable, List, Optional, Union
29
+
30
+ import numpy as np
31
+ import torch
32
+ from diffusers import (DDIMScheduler, DiffusionPipeline,
33
+ DPMSolverMultistepScheduler,
34
+ EulerAncestralDiscreteScheduler, EulerDiscreteScheduler,
35
+ LMSDiscreteScheduler, PNDMScheduler)
36
+ from diffusers.image_processor import VaeImageProcessor
37
+ from diffusers.utils import BaseOutput
38
+ from diffusers.utils.torch_utils import randn_tensor
39
+ from einops import rearrange, repeat
40
+ from tqdm import tqdm
41
+
42
+ from hallo.models.mutual_self_attention import ReferenceAttentionControl
43
+
44
+
45
+ @dataclass
46
+ class FaceAnimatePipelineOutput(BaseOutput):
47
+ """
48
+ FaceAnimatePipelineOutput is a custom class that inherits from BaseOutput and represents the output of the FaceAnimatePipeline.
49
+
50
+ Attributes:
51
+ videos (Union[torch.Tensor, np.ndarray]): A tensor or numpy array containing the generated video frames.
52
+
53
+ Methods:
54
+ __init__(self, videos: Union[torch.Tensor, np.ndarray]): Initializes the FaceAnimatePipelineOutput object with the generated video frames.
55
+ """
56
+ videos: Union[torch.Tensor, np.ndarray]
57
+
58
+ class FaceAnimatePipeline(DiffusionPipeline):
59
+ """
60
+ FaceAnimatePipeline is a custom DiffusionPipeline for animating faces.
61
+
62
+ It inherits from the DiffusionPipeline class and is used to animate faces by
63
+ utilizing a variational autoencoder (VAE), a reference UNet, a denoising UNet,
64
+ a face locator, and an image processor. The pipeline is responsible for generating
65
+ and animating face latents, and decoding the latents to produce the final video output.
66
+
67
+ Attributes:
68
+ vae (VaeImageProcessor): Variational autoencoder for processing images.
69
+ reference_unet (nn.Module): Reference UNet for mutual self-attention.
70
+ denoising_unet (nn.Module): Denoising UNet for image denoising.
71
+ face_locator (nn.Module): Face locator for detecting and cropping faces.
72
+ image_proj (nn.Module): Image projector for processing images.
73
+ scheduler (Union[DDIMScheduler, PNDMScheduler, LMSDiscreteScheduler,
74
+ EulerDiscreteScheduler, EulerAncestralDiscreteScheduler,
75
+ DPMSolverMultistepScheduler]): Diffusion scheduler for
76
+ controlling the noise level.
77
+
78
+ Methods:
79
+ __init__(self, vae, reference_unet, denoising_unet, face_locator,
80
+ image_proj, scheduler): Initializes the FaceAnimatePipeline
81
+ with the given components and scheduler.
82
+ prepare_latents(self, batch_size, num_channels_latents, width, height,
83
+ video_length, dtype, device, generator=None, latents=None):
84
+ Prepares the initial latents for video generation.
85
+ prepare_extra_step_kwargs(self, generator, eta): Prepares extra keyword
86
+ arguments for the scheduler step.
87
+ decode_latents(self, latents): Decodes the latents to produce the final
88
+ video output.
89
+ """
90
+ def __init__(
91
+ self,
92
+ vae,
93
+ reference_unet,
94
+ denoising_unet,
95
+ face_locator,
96
+ image_proj,
97
+ scheduler: Union[
98
+ DDIMScheduler,
99
+ PNDMScheduler,
100
+ LMSDiscreteScheduler,
101
+ EulerDiscreteScheduler,
102
+ EulerAncestralDiscreteScheduler,
103
+ DPMSolverMultistepScheduler,
104
+ ],
105
+ ) -> None:
106
+ super().__init__()
107
+
108
+ self.register_modules(
109
+ vae=vae,
110
+ reference_unet=reference_unet,
111
+ denoising_unet=denoising_unet,
112
+ face_locator=face_locator,
113
+ scheduler=scheduler,
114
+ image_proj=image_proj,
115
+ )
116
+
117
+ self.vae_scale_factor: int = 2 ** (len(self.vae.config.block_out_channels) - 1)
118
+
119
+ self.ref_image_processor = VaeImageProcessor(
120
+ vae_scale_factor=self.vae_scale_factor, do_convert_rgb=True,
121
+ )
122
+
123
+ @property
124
+ def _execution_device(self):
125
+ if self.device != torch.device("meta") or not hasattr(self.unet, "_hf_hook"):
126
+ return self.device
127
+ for module in self.unet.modules():
128
+ if (
129
+ hasattr(module, "_hf_hook")
130
+ and hasattr(module._hf_hook, "execution_device")
131
+ and module._hf_hook.execution_device is not None
132
+ ):
133
+ return torch.device(module._hf_hook.execution_device)
134
+ return self.device
135
+
136
+ def prepare_latents(
137
+ self,
138
+ batch_size: int, # Number of videos to generate in parallel
139
+ num_channels_latents: int, # Number of channels in the latents
140
+ width: int, # Width of the video frame
141
+ height: int, # Height of the video frame
142
+ video_length: int, # Length of the video in frames
143
+ dtype: torch.dtype, # Data type of the latents
144
+ device: torch.device, # Device to store the latents on
145
+ generator: Optional[torch.Generator] = None, # Random number generator for reproducibility
146
+ latents: Optional[torch.Tensor] = None # Pre-generated latents (optional)
147
+ ):
148
+ """
149
+ Prepares the initial latents for video generation.
150
+
151
+ Args:
152
+ batch_size (int): Number of videos to generate in parallel.
153
+ num_channels_latents (int): Number of channels in the latents.
154
+ width (int): Width of the video frame.
155
+ height (int): Height of the video frame.
156
+ video_length (int): Length of the video in frames.
157
+ dtype (torch.dtype): Data type of the latents.
158
+ device (torch.device): Device to store the latents on.
159
+ generator (Optional[torch.Generator]): Random number generator for reproducibility.
160
+ latents (Optional[torch.Tensor]): Pre-generated latents (optional).
161
+
162
+ Returns:
163
+ latents (torch.Tensor): Tensor of shape (batch_size, num_channels_latents, width, height)
164
+ containing the initial latents for video generation.
165
+ """
166
+ shape = (
167
+ batch_size,
168
+ num_channels_latents,
169
+ video_length,
170
+ height // self.vae_scale_factor,
171
+ width // self.vae_scale_factor,
172
+ )
173
+ if isinstance(generator, list) and len(generator) != batch_size:
174
+ raise ValueError(
175
+ f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
176
+ f" size of {batch_size}. Make sure the batch size matches the length of the generators."
177
+ )
178
+
179
+ if latents is None:
180
+ latents = randn_tensor(
181
+ shape, generator=generator, device=device, dtype=dtype
182
+ )
183
+ else:
184
+ latents = latents.to(device)
185
+
186
+ # scale the initial noise by the standard deviation required by the scheduler
187
+ latents = latents * self.scheduler.init_noise_sigma
188
+ return latents
189
+
190
+ def prepare_extra_step_kwargs(self, generator, eta):
191
+ """
192
+ Prepares extra keyword arguments for the scheduler step.
193
+
194
+ Args:
195
+ generator (Optional[torch.Generator]): Random number generator for reproducibility.
196
+ eta (float): The eta (η) parameter used with the DDIMScheduler.
197
+ It corresponds to η in the DDIM paper (https://arxiv.org/abs/2010.02502) and should be between [0, 1].
198
+
199
+ Returns:
200
+ dict: A dictionary containing the extra keyword arguments for the scheduler step.
201
+ """
202
+ # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
203
+ # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
204
+ # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
205
+ # and should be between [0, 1]
206
+
207
+ accepts_eta = "eta" in set(
208
+ inspect.signature(self.scheduler.step).parameters.keys()
209
+ )
210
+ extra_step_kwargs = {}
211
+ if accepts_eta:
212
+ extra_step_kwargs["eta"] = eta
213
+
214
+ # check if the scheduler accepts generator
215
+ accepts_generator = "generator" in set(
216
+ inspect.signature(self.scheduler.step).parameters.keys()
217
+ )
218
+ if accepts_generator:
219
+ extra_step_kwargs["generator"] = generator
220
+ return extra_step_kwargs
221
+
222
+ def decode_latents(self, latents):
223
+ """
224
+ Decode the latents to produce a video.
225
+
226
+ Parameters:
227
+ latents (torch.Tensor): The latents to be decoded.
228
+
229
+ Returns:
230
+ video (torch.Tensor): The decoded video.
231
+ video_length (int): The length of the video in frames.
232
+ """
233
+ video_length = latents.shape[2]
234
+ latents = 1 / 0.18215 * latents
235
+ latents = rearrange(latents, "b c f h w -> (b f) c h w")
236
+ # video = self.vae.decode(latents).sample
237
+ video = []
238
+ for frame_idx in tqdm(range(latents.shape[0])):
239
+ video.append(self.vae.decode(
240
+ latents[frame_idx: frame_idx + 1]).sample)
241
+ video = torch.cat(video)
242
+ video = rearrange(video, "(b f) c h w -> b c f h w", f=video_length)
243
+ video = (video / 2 + 0.5).clamp(0, 1)
244
+ # we always cast to float32 as this does not cause significant overhead and is compatible with bfloa16
245
+ video = video.cpu().float().numpy()
246
+ return video
247
+
248
+
249
+ @torch.no_grad()
250
+ def __call__(
251
+ self,
252
+ ref_image,
253
+ face_emb,
254
+ audio_tensor,
255
+ face_mask,
256
+ pixel_values_full_mask,
257
+ pixel_values_face_mask,
258
+ pixel_values_lip_mask,
259
+ width,
260
+ height,
261
+ video_length,
262
+ num_inference_steps,
263
+ guidance_scale,
264
+ num_images_per_prompt=1,
265
+ eta: float = 0.0,
266
+ motion_scale: Optional[List[torch.Tensor]] = None,
267
+ generator: Optional[Union[torch.Generator,
268
+ List[torch.Generator]]] = None,
269
+ output_type: Optional[str] = "tensor",
270
+ return_dict: bool = True,
271
+ callback: Optional[Callable[[
272
+ int, int, torch.FloatTensor], None]] = None,
273
+ callback_steps: Optional[int] = 1,
274
+ **kwargs,
275
+ ):
276
+ # Default height and width to unet
277
+ height = height or self.unet.config.sample_size * self.vae_scale_factor
278
+ width = width or self.unet.config.sample_size * self.vae_scale_factor
279
+
280
+ device = self._execution_device
281
+
282
+ do_classifier_free_guidance = guidance_scale > 1.0
283
+
284
+ # Prepare timesteps
285
+ self.scheduler.set_timesteps(num_inference_steps, device=device)
286
+ timesteps = self.scheduler.timesteps
287
+
288
+ batch_size = 1
289
+
290
+ # prepare clip image embeddings
291
+ clip_image_embeds = face_emb
292
+ clip_image_embeds = clip_image_embeds.to(self.image_proj.device, self.image_proj.dtype)
293
+
294
+ encoder_hidden_states = self.image_proj(clip_image_embeds)
295
+ uncond_encoder_hidden_states = self.image_proj(torch.zeros_like(clip_image_embeds))
296
+
297
+ if do_classifier_free_guidance:
298
+ encoder_hidden_states = torch.cat([uncond_encoder_hidden_states, encoder_hidden_states], dim=0)
299
+
300
+ reference_control_writer = ReferenceAttentionControl(
301
+ self.reference_unet,
302
+ do_classifier_free_guidance=do_classifier_free_guidance,
303
+ mode="write",
304
+ batch_size=batch_size,
305
+ fusion_blocks="full",
306
+ )
307
+ reference_control_reader = ReferenceAttentionControl(
308
+ self.denoising_unet,
309
+ do_classifier_free_guidance=do_classifier_free_guidance,
310
+ mode="read",
311
+ batch_size=batch_size,
312
+ fusion_blocks="full",
313
+ )
314
+
315
+ num_channels_latents = self.denoising_unet.in_channels
316
+
317
+ latents = self.prepare_latents(
318
+ batch_size * num_images_per_prompt,
319
+ num_channels_latents,
320
+ width,
321
+ height,
322
+ video_length,
323
+ clip_image_embeds.dtype,
324
+ device,
325
+ generator,
326
+ )
327
+
328
+ # Prepare extra step kwargs.
329
+ extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
330
+
331
+ # Prepare ref image latents
332
+ ref_image_tensor = rearrange(ref_image, "b f c h w -> (b f) c h w")
333
+ ref_image_tensor = self.ref_image_processor.preprocess(ref_image_tensor, height=height, width=width) # (bs, c, width, height)
334
+ ref_image_tensor = ref_image_tensor.to(dtype=self.vae.dtype, device=self.vae.device)
335
+ ref_image_latents = self.vae.encode(ref_image_tensor).latent_dist.mean
336
+ ref_image_latents = ref_image_latents * 0.18215 # (b, 4, h, w)
337
+
338
+
339
+ face_mask = face_mask.unsqueeze(1).to(dtype=self.face_locator.dtype, device=self.face_locator.device) # (bs, f, c, H, W)
340
+ face_mask = repeat(face_mask, "b f c h w -> b (repeat f) c h w", repeat=video_length)
341
+ face_mask = face_mask.transpose(1, 2) # (bs, c, f, H, W)
342
+ face_mask = self.face_locator(face_mask)
343
+ face_mask = torch.cat([torch.zeros_like(face_mask), face_mask], dim=0) if do_classifier_free_guidance else face_mask
344
+
345
+ pixel_values_full_mask = (
346
+ [torch.cat([mask] * 2) for mask in pixel_values_full_mask]
347
+ if do_classifier_free_guidance
348
+ else pixel_values_full_mask
349
+ )
350
+ pixel_values_face_mask = (
351
+ [torch.cat([mask] * 2) for mask in pixel_values_face_mask]
352
+ if do_classifier_free_guidance
353
+ else pixel_values_face_mask
354
+ )
355
+ pixel_values_lip_mask = (
356
+ [torch.cat([mask] * 2) for mask in pixel_values_lip_mask]
357
+ if do_classifier_free_guidance
358
+ else pixel_values_lip_mask
359
+ )
360
+ pixel_values_face_mask_ = []
361
+ for mask in pixel_values_face_mask:
362
+ pixel_values_face_mask_.append(
363
+ mask.to(device=self.denoising_unet.device, dtype=self.denoising_unet.dtype))
364
+ pixel_values_face_mask = pixel_values_face_mask_
365
+ pixel_values_lip_mask_ = []
366
+ for mask in pixel_values_lip_mask:
367
+ pixel_values_lip_mask_.append(
368
+ mask.to(device=self.denoising_unet.device, dtype=self.denoising_unet.dtype))
369
+ pixel_values_lip_mask = pixel_values_lip_mask_
370
+ pixel_values_full_mask_ = []
371
+ for mask in pixel_values_full_mask:
372
+ pixel_values_full_mask_.append(
373
+ mask.to(device=self.denoising_unet.device, dtype=self.denoising_unet.dtype))
374
+ pixel_values_full_mask = pixel_values_full_mask_
375
+
376
+
377
+ uncond_audio_tensor = torch.zeros_like(audio_tensor)
378
+ audio_tensor = torch.cat([uncond_audio_tensor, audio_tensor], dim=0)
379
+ audio_tensor = audio_tensor.to(dtype=self.denoising_unet.dtype, device=self.denoising_unet.device)
380
+
381
+ # denoising loop
382
+ num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
383
+ with self.progress_bar(total=num_inference_steps) as progress_bar:
384
+ for i, t in enumerate(timesteps):
385
+ # Forward reference image
386
+ if i == 0:
387
+ self.reference_unet(
388
+ ref_image_latents.repeat(
389
+ (2 if do_classifier_free_guidance else 1), 1, 1, 1
390
+ ),
391
+ torch.zeros_like(t),
392
+ encoder_hidden_states=encoder_hidden_states,
393
+ return_dict=False,
394
+ )
395
+ reference_control_reader.update(reference_control_writer)
396
+
397
+ # expand the latents if we are doing classifier free guidance
398
+ latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
399
+ latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
400
+
401
+ noise_pred = self.denoising_unet(
402
+ latent_model_input,
403
+ t,
404
+ encoder_hidden_states=encoder_hidden_states,
405
+ mask_cond_fea=face_mask,
406
+ full_mask=pixel_values_full_mask,
407
+ face_mask=pixel_values_face_mask,
408
+ lip_mask=pixel_values_lip_mask,
409
+ audio_embedding=audio_tensor,
410
+ motion_scale=motion_scale,
411
+ return_dict=False,
412
+ )[0]
413
+
414
+ # perform guidance
415
+ if do_classifier_free_guidance:
416
+ noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
417
+ noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
418
+
419
+ # compute the previous noisy sample x_t -> x_t-1
420
+ latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
421
+
422
+ # call the callback, if provided
423
+ if i == len(timesteps) - 1 or (i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0:
424
+ progress_bar.update()
425
+ if callback is not None and i % callback_steps == 0:
426
+ step_idx = i // getattr(self.scheduler, "order", 1)
427
+ callback(step_idx, t, latents)
428
+
429
+ reference_control_reader.clear()
430
+ reference_control_writer.clear()
431
+
432
+ # Post-processing
433
+ images = self.decode_latents(latents) # (b, c, f, h, w)
434
+
435
+ # Convert to tensor
436
+ if output_type == "tensor":
437
+ images = torch.from_numpy(images)
438
+
439
+ if not return_dict:
440
+ return images
441
+
442
+ return FaceAnimatePipelineOutput(videos=images)
hallo/animate/face_animate_static.py ADDED
@@ -0,0 +1,481 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # pylint: disable=R0801
2
+ """
3
+ This module is responsible for handling the animation of faces using a combination of deep learning models and image processing techniques.
4
+ It provides a pipeline to generate realistic face animations by incorporating user-provided conditions such as facial expressions and environments.
5
+ The module utilizes various schedulers and utilities to optimize the animation process and ensure efficient performance.
6
+
7
+ Functions and Classes:
8
+ - StaticPipelineOutput: A class that represents the output of the animation pipeline, c
9
+ ontaining properties and methods related to the generated images.
10
+ - prepare_latents: A function that prepares the initial noise for the animation process,
11
+ scaling it according to the scheduler's requirements.
12
+ - prepare_condition: A function that processes the user-provided conditions
13
+ (e.g., facial expressions) and prepares them for use in the animation pipeline.
14
+ - decode_latents: A function that decodes the latent representations of the face animations into
15
+ their corresponding image formats.
16
+ - prepare_extra_step_kwargs: A function that prepares additional parameters for each step of
17
+ the animation process, such as the generator and eta values.
18
+
19
+ Dependencies:
20
+ - numpy: A library for numerical computing.
21
+ - torch: A machine learning library based on PyTorch.
22
+ - diffusers: A library for image-to-image diffusion models.
23
+ - transformers: A library for pre-trained transformer models.
24
+
25
+ Usage:
26
+ - To create an instance of the animation pipeline, provide the necessary components such as
27
+ the VAE, reference UNET, denoising UNET, face locator, and image processor.
28
+ - Use the pipeline's methods to prepare the latents, conditions, and extra step arguments as
29
+ required for the animation process.
30
+ - Generate the face animations by decoding the latents and processing the conditions.
31
+
32
+ Note:
33
+ - The module is designed to work with the diffusers library, which is based on
34
+ the paper "Diffusion Models for Image-to-Image Translation" (https://arxiv.org/abs/2102.02765).
35
+ - The face animations generated by this module should be used for entertainment purposes
36
+ only and should respect the rights and privacy of the individuals involved.
37
+ """
38
+ import inspect
39
+ from dataclasses import dataclass
40
+ from typing import Callable, List, Optional, Union
41
+
42
+ import numpy as np
43
+ import torch
44
+ from diffusers import DiffusionPipeline
45
+ from diffusers.image_processor import VaeImageProcessor
46
+ from diffusers.schedulers import (DDIMScheduler, DPMSolverMultistepScheduler,
47
+ EulerAncestralDiscreteScheduler,
48
+ EulerDiscreteScheduler, LMSDiscreteScheduler,
49
+ PNDMScheduler)
50
+ from diffusers.utils import BaseOutput, is_accelerate_available
51
+ from diffusers.utils.torch_utils import randn_tensor
52
+ from einops import rearrange
53
+ from tqdm import tqdm
54
+ from transformers import CLIPImageProcessor
55
+
56
+ from hallo.models.mutual_self_attention import ReferenceAttentionControl
57
+
58
+ if is_accelerate_available():
59
+ from accelerate import cpu_offload
60
+ else:
61
+ raise ImportError("Please install accelerate via `pip install accelerate`")
62
+
63
+
64
+ @dataclass
65
+ class StaticPipelineOutput(BaseOutput):
66
+ """
67
+ StaticPipelineOutput is a class that represents the output of the static pipeline.
68
+ It contains the images generated by the pipeline as a union of torch.Tensor and np.ndarray.
69
+
70
+ Attributes:
71
+ images (Union[torch.Tensor, np.ndarray]): The generated images.
72
+ """
73
+ images: Union[torch.Tensor, np.ndarray]
74
+
75
+
76
+ class StaticPipeline(DiffusionPipeline):
77
+ """
78
+ StaticPipelineOutput is a class that represents the output of the static pipeline.
79
+ It contains the images generated by the pipeline as a union of torch.Tensor and np.ndarray.
80
+
81
+ Attributes:
82
+ images (Union[torch.Tensor, np.ndarray]): The generated images.
83
+ """
84
+ _optional_components = []
85
+
86
+ def __init__(
87
+ self,
88
+ vae,
89
+ reference_unet,
90
+ denoising_unet,
91
+ face_locator,
92
+ imageproj,
93
+ scheduler: Union[
94
+ DDIMScheduler,
95
+ PNDMScheduler,
96
+ LMSDiscreteScheduler,
97
+ EulerDiscreteScheduler,
98
+ EulerAncestralDiscreteScheduler,
99
+ DPMSolverMultistepScheduler,
100
+ ],
101
+ ):
102
+ super().__init__()
103
+
104
+ self.register_modules(
105
+ vae=vae,
106
+ reference_unet=reference_unet,
107
+ denoising_unet=denoising_unet,
108
+ face_locator=face_locator,
109
+ scheduler=scheduler,
110
+ imageproj=imageproj,
111
+ )
112
+ self.vae_scale_factor = 2 ** (
113
+ len(self.vae.config.block_out_channels) - 1)
114
+ self.clip_image_processor = CLIPImageProcessor()
115
+ self.ref_image_processor = VaeImageProcessor(
116
+ vae_scale_factor=self.vae_scale_factor, do_convert_rgb=True
117
+ )
118
+ self.cond_image_processor = VaeImageProcessor(
119
+ vae_scale_factor=self.vae_scale_factor,
120
+ do_convert_rgb=True,
121
+ do_normalize=False,
122
+ )
123
+
124
+ def enable_vae_slicing(self):
125
+ """
126
+ Enable VAE slicing.
127
+
128
+ This method enables slicing for the VAE model, which can help improve the performance of decoding latents when working with large images.
129
+ """
130
+ self.vae.enable_slicing()
131
+
132
+ def disable_vae_slicing(self):
133
+ """
134
+ Disable vae slicing.
135
+
136
+ This function disables the vae slicing for the StaticPipeline object.
137
+ It calls the `disable_slicing()` method of the vae model.
138
+ This is useful when you want to use the entire vae model for decoding latents
139
+ instead of slicing it for better performance.
140
+ """
141
+ self.vae.disable_slicing()
142
+
143
+ def enable_sequential_cpu_offload(self, gpu_id=0):
144
+ """
145
+ Offloads selected models to the GPU for increased performance.
146
+
147
+ Args:
148
+ gpu_id (int, optional): The ID of the GPU to offload models to. Defaults to 0.
149
+ """
150
+ device = torch.device(f"cuda:{gpu_id}")
151
+
152
+ for cpu_offloaded_model in [self.unet, self.text_encoder, self.vae]:
153
+ if cpu_offloaded_model is not None:
154
+ cpu_offload(cpu_offloaded_model, device)
155
+
156
+ @property
157
+ def _execution_device(self):
158
+ if self.device != torch.device("meta") or not hasattr(self.unet, "_hf_hook"):
159
+ return self.device
160
+ for module in self.unet.modules():
161
+ if (
162
+ hasattr(module, "_hf_hook")
163
+ and hasattr(module._hf_hook, "execution_device")
164
+ and module._hf_hook.execution_device is not None
165
+ ):
166
+ return torch.device(module._hf_hook.execution_device)
167
+ return self.device
168
+
169
+ def decode_latents(self, latents):
170
+ """
171
+ Decode the given latents to video frames.
172
+
173
+ Parameters:
174
+ latents (torch.Tensor): The latents to be decoded. Shape: (batch_size, num_channels_latents, video_length, height, width).
175
+
176
+ Returns:
177
+ video (torch.Tensor): The decoded video frames. Shape: (batch_size, num_channels_latents, video_length, height, width).
178
+ """
179
+ video_length = latents.shape[2]
180
+ latents = 1 / 0.18215 * latents
181
+ latents = rearrange(latents, "b c f h w -> (b f) c h w")
182
+ # video = self.vae.decode(latents).sample
183
+ video = []
184
+ for frame_idx in tqdm(range(latents.shape[0])):
185
+ video.append(self.vae.decode(
186
+ latents[frame_idx: frame_idx + 1]).sample)
187
+ video = torch.cat(video)
188
+ video = rearrange(video, "(b f) c h w -> b c f h w", f=video_length)
189
+ video = (video / 2 + 0.5).clamp(0, 1)
190
+ # we always cast to float32 as this does not cause significant overhead and is compatible with bfloa16
191
+ video = video.cpu().float().numpy()
192
+ return video
193
+
194
+ def prepare_extra_step_kwargs(self, generator, eta):
195
+ """
196
+ Prepare extra keyword arguments for the scheduler step.
197
+
198
+ Since not all schedulers have the same signature, this function helps to create a consistent interface for the scheduler.
199
+
200
+ Args:
201
+ generator (Optional[torch.Generator]): A random number generator for reproducibility.
202
+ eta (float): The eta parameter used with the DDIMScheduler. It should be between 0 and 1.
203
+
204
+ Returns:
205
+ dict: A dictionary containing the extra keyword arguments for the scheduler step.
206
+ """
207
+ # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
208
+ # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
209
+ # and should be between [0, 1]
210
+
211
+ accepts_eta = "eta" in set(
212
+ inspect.signature(self.scheduler.step).parameters.keys()
213
+ )
214
+ extra_step_kwargs = {}
215
+ if accepts_eta:
216
+ extra_step_kwargs["eta"] = eta
217
+
218
+ # check if the scheduler accepts generator
219
+ accepts_generator = "generator" in set(
220
+ inspect.signature(self.scheduler.step).parameters.keys()
221
+ )
222
+ if accepts_generator:
223
+ extra_step_kwargs["generator"] = generator
224
+ return extra_step_kwargs
225
+
226
+ def prepare_latents(
227
+ self,
228
+ batch_size,
229
+ num_channels_latents,
230
+ width,
231
+ height,
232
+ dtype,
233
+ device,
234
+ generator,
235
+ latents=None,
236
+ ):
237
+ """
238
+ Prepares the initial latents for the diffusion pipeline.
239
+
240
+ Args:
241
+ batch_size (int): The number of images to generate in one forward pass.
242
+ num_channels_latents (int): The number of channels in the latents tensor.
243
+ width (int): The width of the latents tensor.
244
+ height (int): The height of the latents tensor.
245
+ dtype (torch.dtype): The data type of the latents tensor.
246
+ device (torch.device): The device to place the latents tensor on.
247
+ generator (Optional[torch.Generator], optional): A random number generator
248
+ for reproducibility. Defaults to None.
249
+ latents (Optional[torch.Tensor], optional): Pre-computed latents to use as
250
+ initial conditions for the diffusion process. Defaults to None.
251
+
252
+ Returns:
253
+ torch.Tensor: The prepared latents tensor.
254
+ """
255
+ shape = (
256
+ batch_size,
257
+ num_channels_latents,
258
+ height // self.vae_scale_factor,
259
+ width // self.vae_scale_factor,
260
+ )
261
+ if isinstance(generator, list) and len(generator) != batch_size:
262
+ raise ValueError(
263
+ f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
264
+ f" size of {batch_size}. Make sure the batch size matches the length of the generators."
265
+ )
266
+
267
+ if latents is None:
268
+ latents = randn_tensor(
269
+ shape, generator=generator, device=device, dtype=dtype
270
+ )
271
+ else:
272
+ latents = latents.to(device)
273
+
274
+ # scale the initial noise by the standard deviation required by the scheduler
275
+ latents = latents * self.scheduler.init_noise_sigma
276
+ return latents
277
+
278
+ def prepare_condition(
279
+ self,
280
+ cond_image,
281
+ width,
282
+ height,
283
+ device,
284
+ dtype,
285
+ do_classififer_free_guidance=False,
286
+ ):
287
+ """
288
+ Prepares the condition for the face animation pipeline.
289
+
290
+ Args:
291
+ cond_image (torch.Tensor): The conditional image tensor.
292
+ width (int): The width of the output image.
293
+ height (int): The height of the output image.
294
+ device (torch.device): The device to run the pipeline on.
295
+ dtype (torch.dtype): The data type of the tensor.
296
+ do_classififer_free_guidance (bool, optional): Whether to use classifier-free guidance or not. Defaults to False.
297
+
298
+ Returns:
299
+ Tuple[torch.Tensor, torch.Tensor]: A tuple of processed condition and mask tensors.
300
+ """
301
+ image = self.cond_image_processor.preprocess(
302
+ cond_image, height=height, width=width
303
+ ).to(dtype=torch.float32)
304
+
305
+ image = image.to(device=device, dtype=dtype)
306
+
307
+ if do_classififer_free_guidance:
308
+ image = torch.cat([image] * 2)
309
+
310
+ return image
311
+
312
+ @torch.no_grad()
313
+ def __call__(
314
+ self,
315
+ ref_image,
316
+ face_mask,
317
+ width,
318
+ height,
319
+ num_inference_steps,
320
+ guidance_scale,
321
+ face_embedding,
322
+ num_images_per_prompt=1,
323
+ eta: float = 0.0,
324
+ generator: Optional[Union[torch.Generator,
325
+ List[torch.Generator]]] = None,
326
+ output_type: Optional[str] = "tensor",
327
+ return_dict: bool = True,
328
+ callback: Optional[Callable[[
329
+ int, int, torch.FloatTensor], None]] = None,
330
+ callback_steps: Optional[int] = 1,
331
+ **kwargs,
332
+ ):
333
+ # Default height and width to unet
334
+ height = height or self.unet.config.sample_size * self.vae_scale_factor
335
+ width = width or self.unet.config.sample_size * self.vae_scale_factor
336
+
337
+ device = self._execution_device
338
+
339
+ do_classifier_free_guidance = guidance_scale > 1.0
340
+
341
+ # Prepare timesteps
342
+ self.scheduler.set_timesteps(num_inference_steps, device=device)
343
+ timesteps = self.scheduler.timesteps
344
+
345
+ batch_size = 1
346
+
347
+ image_prompt_embeds = self.imageproj(face_embedding)
348
+ uncond_image_prompt_embeds = self.imageproj(
349
+ torch.zeros_like(face_embedding))
350
+
351
+ if do_classifier_free_guidance:
352
+ image_prompt_embeds = torch.cat(
353
+ [uncond_image_prompt_embeds, image_prompt_embeds], dim=0
354
+ )
355
+
356
+ reference_control_writer = ReferenceAttentionControl(
357
+ self.reference_unet,
358
+ do_classifier_free_guidance=do_classifier_free_guidance,
359
+ mode="write",
360
+ batch_size=batch_size,
361
+ fusion_blocks="full",
362
+ )
363
+ reference_control_reader = ReferenceAttentionControl(
364
+ self.denoising_unet,
365
+ do_classifier_free_guidance=do_classifier_free_guidance,
366
+ mode="read",
367
+ batch_size=batch_size,
368
+ fusion_blocks="full",
369
+ )
370
+
371
+ num_channels_latents = self.denoising_unet.in_channels
372
+ latents = self.prepare_latents(
373
+ batch_size * num_images_per_prompt,
374
+ num_channels_latents,
375
+ width,
376
+ height,
377
+ face_embedding.dtype,
378
+ device,
379
+ generator,
380
+ )
381
+ latents = latents.unsqueeze(2) # (bs, c, 1, h', w')
382
+ # latents_dtype = latents.dtype
383
+
384
+ # Prepare extra step kwargs.
385
+ extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
386
+
387
+ # Prepare ref image latents
388
+ ref_image_tensor = self.ref_image_processor.preprocess(
389
+ ref_image, height=height, width=width
390
+ ) # (bs, c, width, height)
391
+ ref_image_tensor = ref_image_tensor.to(
392
+ dtype=self.vae.dtype, device=self.vae.device
393
+ )
394
+ ref_image_latents = self.vae.encode(ref_image_tensor).latent_dist.mean
395
+ ref_image_latents = ref_image_latents * 0.18215 # (b, 4, h, w)
396
+
397
+ # Prepare face mask image
398
+ face_mask_tensor = self.cond_image_processor.preprocess(
399
+ face_mask, height=height, width=width
400
+ )
401
+ face_mask_tensor = face_mask_tensor.unsqueeze(2) # (bs, c, 1, h, w)
402
+ face_mask_tensor = face_mask_tensor.to(
403
+ device=device, dtype=self.face_locator.dtype
404
+ )
405
+ mask_fea = self.face_locator(face_mask_tensor)
406
+ mask_fea = (
407
+ torch.cat(
408
+ [mask_fea] * 2) if do_classifier_free_guidance else mask_fea
409
+ )
410
+
411
+ # denoising loop
412
+ num_warmup_steps = len(timesteps) - \
413
+ num_inference_steps * self.scheduler.order
414
+ with self.progress_bar(total=num_inference_steps) as progress_bar:
415
+ for i, t in enumerate(timesteps):
416
+ # 1. Forward reference image
417
+ if i == 0:
418
+ self.reference_unet(
419
+ ref_image_latents.repeat(
420
+ (2 if do_classifier_free_guidance else 1), 1, 1, 1
421
+ ),
422
+ torch.zeros_like(t),
423
+ encoder_hidden_states=image_prompt_embeds,
424
+ return_dict=False,
425
+ )
426
+
427
+ # 2. Update reference unet feature into denosing net
428
+ reference_control_reader.update(reference_control_writer)
429
+
430
+ # 3.1 expand the latents if we are doing classifier free guidance
431
+ latent_model_input = (
432
+ torch.cat(
433
+ [latents] * 2) if do_classifier_free_guidance else latents
434
+ )
435
+ latent_model_input = self.scheduler.scale_model_input(
436
+ latent_model_input, t
437
+ )
438
+
439
+ noise_pred = self.denoising_unet(
440
+ latent_model_input,
441
+ t,
442
+ encoder_hidden_states=image_prompt_embeds,
443
+ mask_cond_fea=mask_fea,
444
+ return_dict=False,
445
+ )[0]
446
+
447
+ # perform guidance
448
+ if do_classifier_free_guidance:
449
+ noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
450
+ noise_pred = noise_pred_uncond + guidance_scale * (
451
+ noise_pred_text - noise_pred_uncond
452
+ )
453
+
454
+ # compute the previous noisy sample x_t -> x_t-1
455
+ latents = self.scheduler.step(
456
+ noise_pred, t, latents, **extra_step_kwargs, return_dict=False
457
+ )[0]
458
+
459
+ # call the callback, if provided
460
+ if i == len(timesteps) - 1 or (
461
+ (i + 1) > num_warmup_steps and (i +
462
+ 1) % self.scheduler.order == 0
463
+ ):
464
+ progress_bar.update()
465
+ if callback is not None and i % callback_steps == 0:
466
+ step_idx = i // getattr(self.scheduler, "order", 1)
467
+ callback(step_idx, t, latents)
468
+ reference_control_reader.clear()
469
+ reference_control_writer.clear()
470
+
471
+ # Post-processing
472
+ image = self.decode_latents(latents) # (b, c, 1, h, w)
473
+
474
+ # Convert to tensor
475
+ if output_type == "tensor":
476
+ image = torch.from_numpy(image)
477
+
478
+ if not return_dict:
479
+ return image
480
+
481
+ return StaticPipelineOutput(images=image)
hallo/datasets/__init__.py ADDED
File without changes
hallo/datasets/audio_processor.py ADDED
@@ -0,0 +1,172 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # pylint: disable=C0301
2
+ '''
3
+ This module contains the AudioProcessor class and related functions for processing audio data.
4
+ It utilizes various libraries and models to perform tasks such as preprocessing, feature extraction,
5
+ and audio separation. The class is initialized with configuration parameters and can process
6
+ audio files using the provided models.
7
+ '''
8
+ import math
9
+ import os
10
+
11
+ import librosa
12
+ import numpy as np
13
+ import torch
14
+ from audio_separator.separator import Separator
15
+ from einops import rearrange
16
+ from transformers import Wav2Vec2FeatureExtractor
17
+
18
+ from hallo.models.wav2vec import Wav2VecModel
19
+ from hallo.utils.util import resample_audio
20
+
21
+
22
+ class AudioProcessor:
23
+ """
24
+ AudioProcessor is a class that handles the processing of audio files.
25
+ It takes care of preprocessing the audio files, extracting features
26
+ using wav2vec models, and separating audio signals if needed.
27
+
28
+ :param sample_rate: Sampling rate of the audio file
29
+ :param fps: Frames per second for the extracted features
30
+ :param wav2vec_model_path: Path to the wav2vec model
31
+ :param only_last_features: Whether to only use the last features
32
+ :param audio_separator_model_path: Path to the audio separator model
33
+ :param audio_separator_model_name: Name of the audio separator model
34
+ :param cache_dir: Directory to cache the intermediate results
35
+ :param device: Device to run the processing on
36
+ """
37
+ def __init__(
38
+ self,
39
+ sample_rate,
40
+ fps,
41
+ wav2vec_model_path,
42
+ only_last_features,
43
+ audio_separator_model_path:str=None,
44
+ audio_separator_model_name:str=None,
45
+ cache_dir:str='',
46
+ device="cuda:0",
47
+ ) -> None:
48
+ self.sample_rate = sample_rate
49
+ self.fps = fps
50
+ self.device = device
51
+
52
+ self.audio_encoder = Wav2VecModel.from_pretrained(wav2vec_model_path, local_files_only=True).to(device=device)
53
+ self.audio_encoder.feature_extractor._freeze_parameters()
54
+ self.only_last_features = only_last_features
55
+
56
+ if audio_separator_model_name is not None:
57
+ try:
58
+ os.makedirs(cache_dir, exist_ok=True)
59
+ except OSError as _:
60
+ print("Fail to create the output cache dir.")
61
+ self.audio_separator = Separator(
62
+ output_dir=cache_dir,
63
+ output_single_stem="vocals",
64
+ model_file_dir=audio_separator_model_path,
65
+ )
66
+ self.audio_separator.load_model(audio_separator_model_name)
67
+ assert self.audio_separator.model_instance is not None, "Fail to load audio separate model."
68
+ else:
69
+ self.audio_separator=None
70
+ print("Use audio directly without vocals seperator.")
71
+
72
+
73
+ self.wav2vec_feature_extractor = Wav2Vec2FeatureExtractor.from_pretrained(wav2vec_model_path, local_files_only=True)
74
+
75
+
76
+ def preprocess(self, wav_file: str):
77
+ """
78
+ Preprocess a WAV audio file by separating the vocals from the background and resampling it to a 16 kHz sample rate.
79
+ The separated vocal track is then converted into wav2vec2 for further processing or analysis.
80
+
81
+ Args:
82
+ wav_file (str): The path to the WAV file to be processed. This file should be accessible and in WAV format.
83
+
84
+ Raises:
85
+ RuntimeError: Raises an exception if the WAV file cannot be processed. This could be due to issues
86
+ such as file not found, unsupported file format, or errors during the audio processing steps.
87
+
88
+ Returns:
89
+ torch.tensor: Returns an audio embedding as a torch.tensor
90
+ """
91
+ if self.audio_separator is not None:
92
+ # 1. separate vocals
93
+ # TODO: process in memory
94
+ outputs = self.audio_separator.separate(wav_file)
95
+ if len(outputs) <= 0:
96
+ raise RuntimeError("Audio separate failed.")
97
+
98
+ vocal_audio_file = outputs[0]
99
+ vocal_audio_name, _ = os.path.splitext(vocal_audio_file)
100
+ vocal_audio_file = os.path.join(self.audio_separator.output_dir, vocal_audio_file)
101
+ vocal_audio_file = resample_audio(vocal_audio_file, os.path.join(self.audio_separator.output_dir, f"{vocal_audio_name}-16k.wav"), self.sample_rate)
102
+ else:
103
+ vocal_audio_file=wav_file
104
+
105
+ # 2. extract wav2vec features
106
+ speech_array, sampling_rate = librosa.load(vocal_audio_file, sr=self.sample_rate)
107
+ audio_feature = np.squeeze(self.wav2vec_feature_extractor(speech_array, sampling_rate=sampling_rate).input_values)
108
+ seq_len = math.ceil(len(audio_feature) / self.sample_rate * self.fps)
109
+
110
+ audio_feature = torch.from_numpy(audio_feature).float().to(device=self.device)
111
+ audio_feature = audio_feature.unsqueeze(0)
112
+
113
+ with torch.no_grad():
114
+ embeddings = self.audio_encoder(audio_feature, seq_len=seq_len, output_hidden_states=True)
115
+ assert len(embeddings) > 0, "Fail to extract audio embedding"
116
+ if self.only_last_features:
117
+ audio_emb = embeddings.last_hidden_state.squeeze()
118
+ else:
119
+ audio_emb = torch.stack(embeddings.hidden_states[1:], dim=1).squeeze(0)
120
+ audio_emb = rearrange(audio_emb, "b s d -> s b d")
121
+
122
+ audio_emb = audio_emb.cpu().detach()
123
+
124
+ return audio_emb
125
+
126
+ def get_embedding(self, wav_file: str):
127
+ """preprocess wav audio file convert to embeddings
128
+
129
+ Args:
130
+ wav_file (str): The path to the WAV file to be processed. This file should be accessible and in WAV format.
131
+
132
+ Returns:
133
+ torch.tensor: Returns an audio embedding as a torch.tensor
134
+ """
135
+ speech_array, sampling_rate = librosa.load(
136
+ wav_file, sr=self.sample_rate)
137
+ assert sampling_rate == 16000, "The audio sample rate must be 16000"
138
+ audio_feature = np.squeeze(self.wav2vec_feature_extractor(
139
+ speech_array, sampling_rate=sampling_rate).input_values)
140
+ seq_len = math.ceil(len(audio_feature) / self.sample_rate * self.fps)
141
+
142
+ audio_feature = torch.from_numpy(
143
+ audio_feature).float().to(device=self.device)
144
+ audio_feature = audio_feature.unsqueeze(0)
145
+
146
+ with torch.no_grad():
147
+ embeddings = self.audio_encoder(
148
+ audio_feature, seq_len=seq_len, output_hidden_states=True)
149
+ assert len(embeddings) > 0, "Fail to extract audio embedding"
150
+
151
+ if self.only_last_features:
152
+ audio_emb = embeddings.last_hidden_state.squeeze()
153
+ else:
154
+ audio_emb = torch.stack(
155
+ embeddings.hidden_states[1:], dim=1).squeeze(0)
156
+ audio_emb = rearrange(audio_emb, "b s d -> s b d")
157
+
158
+ audio_emb = audio_emb.cpu().detach()
159
+
160
+ return audio_emb
161
+
162
+ def close(self):
163
+ """
164
+ TODO: to be implemented
165
+ """
166
+ return self
167
+
168
+ def __enter__(self):
169
+ return self
170
+
171
+ def __exit__(self, _exc_type, _exc_val, _exc_tb):
172
+ self.close()
hallo/datasets/image_processor.py ADDED
@@ -0,0 +1,201 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """
2
+ This module is responsible for processing images, particularly for face-related tasks.
3
+ It uses various libraries such as OpenCV, NumPy, and InsightFace to perform tasks like
4
+ face detection, augmentation, and mask rendering. The ImageProcessor class encapsulates
5
+ the functionality for these operations.
6
+ """
7
+ import os
8
+ from typing import List
9
+
10
+ import cv2
11
+ import numpy as np
12
+ import torch
13
+ from insightface.app import FaceAnalysis
14
+ from PIL import Image
15
+ from torchvision import transforms
16
+
17
+ from ..utils.util import get_mask
18
+
19
+ MEAN = 0.5
20
+ STD = 0.5
21
+
22
+ class ImageProcessor:
23
+ """
24
+ ImageProcessor is a class responsible for processing images, particularly for face-related tasks.
25
+ It takes in an image and performs various operations such as augmentation, face detection,
26
+ face embedding extraction, and rendering a face mask. The processed images are then used for
27
+ further analysis or recognition purposes.
28
+
29
+ Attributes:
30
+ img_size (int): The size of the image to be processed.
31
+ face_analysis_model_path (str): The path to the face analysis model.
32
+
33
+ Methods:
34
+ preprocess(source_image_path, cache_dir):
35
+ Preprocesses the input image by performing augmentation, face detection,
36
+ face embedding extraction, and rendering a face mask.
37
+
38
+ close():
39
+ Closes the ImageProcessor and releases any resources being used.
40
+
41
+ _augmentation(images, transform, state=None):
42
+ Applies image augmentation to the input images using the given transform and state.
43
+
44
+ __enter__():
45
+ Enters a runtime context and returns the ImageProcessor object.
46
+
47
+ __exit__(_exc_type, _exc_val, _exc_tb):
48
+ Exits a runtime context and handles any exceptions that occurred during the processing.
49
+ """
50
+ def __init__(self, img_size, face_analysis_model_path) -> None:
51
+ self.img_size = img_size
52
+
53
+ self.pixel_transform = transforms.Compose(
54
+ [
55
+ transforms.Resize(self.img_size),
56
+ transforms.ToTensor(),
57
+ transforms.Normalize([MEAN], [STD]),
58
+ ]
59
+ )
60
+
61
+ self.cond_transform = transforms.Compose(
62
+ [
63
+ transforms.Resize(self.img_size),
64
+ transforms.ToTensor(),
65
+ ]
66
+ )
67
+
68
+ self.attn_transform_64 = transforms.Compose(
69
+ [
70
+ transforms.Resize(
71
+ (self.img_size[0] // 8, self.img_size[0] // 8)),
72
+ transforms.ToTensor(),
73
+ ]
74
+ )
75
+ self.attn_transform_32 = transforms.Compose(
76
+ [
77
+ transforms.Resize(
78
+ (self.img_size[0] // 16, self.img_size[0] // 16)),
79
+ transforms.ToTensor(),
80
+ ]
81
+ )
82
+ self.attn_transform_16 = transforms.Compose(
83
+ [
84
+ transforms.Resize(
85
+ (self.img_size[0] // 32, self.img_size[0] // 32)),
86
+ transforms.ToTensor(),
87
+ ]
88
+ )
89
+ self.attn_transform_8 = transforms.Compose(
90
+ [
91
+ transforms.Resize(
92
+ (self.img_size[0] // 64, self.img_size[0] // 64)),
93
+ transforms.ToTensor(),
94
+ ]
95
+ )
96
+
97
+ self.face_analysis = FaceAnalysis(
98
+ name="",
99
+ root=face_analysis_model_path,
100
+ providers=["CUDAExecutionProvider", "CPUExecutionProvider"],
101
+ )
102
+ self.face_analysis.prepare(ctx_id=0, det_size=(640, 640))
103
+
104
+ def preprocess(self, source_image_path: str, cache_dir: str, face_region_ratio: float):
105
+ """
106
+ Apply preprocessing to the source image to prepare for face analysis.
107
+
108
+ Parameters:
109
+ source_image_path (str): The path to the source image.
110
+ cache_dir (str): The directory to cache intermediate results.
111
+
112
+ Returns:
113
+ None
114
+ """
115
+ source_image = Image.open(source_image_path)
116
+ ref_image_pil = source_image.convert("RGB")
117
+ # 1. image augmentation
118
+ pixel_values_ref_img = self._augmentation(ref_image_pil, self.pixel_transform)
119
+
120
+
121
+ # 2.1 detect face
122
+ faces = self.face_analysis.get(cv2.cvtColor(np.array(ref_image_pil.copy()), cv2.COLOR_RGB2BGR))
123
+ # use max size face
124
+ face = sorted(faces, key=lambda x: (x["bbox"][2] - x["bbox"][0]) * (x["bbox"][3] - x["bbox"][1]))[-1]
125
+
126
+ # 2.2 face embedding
127
+ face_emb = face["embedding"]
128
+
129
+ # 2.3 render face mask
130
+ get_mask(source_image_path, cache_dir, face_region_ratio)
131
+ file_name = os.path.basename(source_image_path).split(".")[0]
132
+ face_mask_pil = Image.open(
133
+ os.path.join(cache_dir, f"{file_name}_face_mask.png")).convert("RGB")
134
+
135
+ face_mask = self._augmentation(face_mask_pil, self.cond_transform)
136
+
137
+ # 2.4 detect and expand lip, face mask
138
+ sep_background_mask = Image.open(
139
+ os.path.join(cache_dir, f"{file_name}_sep_background.png"))
140
+ sep_face_mask = Image.open(
141
+ os.path.join(cache_dir, f"{file_name}_sep_face.png"))
142
+ sep_lip_mask = Image.open(
143
+ os.path.join(cache_dir, f"{file_name}_sep_lip.png"))
144
+
145
+ pixel_values_face_mask = [
146
+ self._augmentation(sep_face_mask, self.attn_transform_64),
147
+ self._augmentation(sep_face_mask, self.attn_transform_32),
148
+ self._augmentation(sep_face_mask, self.attn_transform_16),
149
+ self._augmentation(sep_face_mask, self.attn_transform_8),
150
+ ]
151
+ pixel_values_lip_mask = [
152
+ self._augmentation(sep_lip_mask, self.attn_transform_64),
153
+ self._augmentation(sep_lip_mask, self.attn_transform_32),
154
+ self._augmentation(sep_lip_mask, self.attn_transform_16),
155
+ self._augmentation(sep_lip_mask, self.attn_transform_8),
156
+ ]
157
+ pixel_values_full_mask = [
158
+ self._augmentation(sep_background_mask, self.attn_transform_64),
159
+ self._augmentation(sep_background_mask, self.attn_transform_32),
160
+ self._augmentation(sep_background_mask, self.attn_transform_16),
161
+ self._augmentation(sep_background_mask, self.attn_transform_8),
162
+ ]
163
+
164
+ pixel_values_full_mask = [mask.view(1, -1)
165
+ for mask in pixel_values_full_mask]
166
+ pixel_values_face_mask = [mask.view(1, -1)
167
+ for mask in pixel_values_face_mask]
168
+ pixel_values_lip_mask = [mask.view(1, -1)
169
+ for mask in pixel_values_lip_mask]
170
+
171
+ return pixel_values_ref_img, face_mask, face_emb, pixel_values_full_mask, pixel_values_face_mask, pixel_values_lip_mask
172
+
173
+ def close(self):
174
+ """
175
+ Closes the ImageProcessor and releases any resources held by the FaceAnalysis instance.
176
+
177
+ Args:
178
+ self: The ImageProcessor instance.
179
+
180
+ Returns:
181
+ None.
182
+ """
183
+ for _, model in self.face_analysis.models.items():
184
+ if hasattr(model, "Dispose"):
185
+ model.Dispose()
186
+
187
+ def _augmentation(self, images, transform, state=None):
188
+ if state is not None:
189
+ torch.set_rng_state(state)
190
+ if isinstance(images, List):
191
+ transformed_images = [transform(img) for img in images]
192
+ ret_tensor = torch.stack(transformed_images, dim=0) # (f, c, h, w)
193
+ else:
194
+ ret_tensor = transform(images) # (c, h, w)
195
+ return ret_tensor
196
+
197
+ def __enter__(self):
198
+ return self
199
+
200
+ def __exit__(self, _exc_type, _exc_val, _exc_tb):
201
+ self.close()
hallo/datasets/mask_image.py ADDED
@@ -0,0 +1,154 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # pylint: disable=R0801
2
+ """
3
+ This module contains the code for a dataset class called FaceMaskDataset, which is used to process and
4
+ load image data related to face masks. The dataset class inherits from the PyTorch Dataset class and
5
+ provides methods for data augmentation, getting items from the dataset, and determining the length of the
6
+ dataset. The module also includes imports for necessary libraries such as json, random, pathlib, torch,
7
+ PIL, and transformers.
8
+ """
9
+
10
+ import json
11
+ import random
12
+ from pathlib import Path
13
+
14
+ import torch
15
+ from PIL import Image
16
+ from torch.utils.data import Dataset
17
+ from torchvision import transforms
18
+ from transformers import CLIPImageProcessor
19
+
20
+
21
+ class FaceMaskDataset(Dataset):
22
+ """
23
+ FaceMaskDataset is a custom dataset for face mask images.
24
+
25
+ Args:
26
+ img_size (int): The size of the input images.
27
+ drop_ratio (float, optional): The ratio of dropped pixels during data augmentation. Defaults to 0.1.
28
+ data_meta_paths (list, optional): The paths to the metadata files containing image paths and labels. Defaults to ["./data/HDTF_meta.json"].
29
+ sample_margin (int, optional): The margin for sampling regions in the image. Defaults to 30.
30
+
31
+ Attributes:
32
+ img_size (int): The size of the input images.
33
+ drop_ratio (float): The ratio of dropped pixels during data augmentation.
34
+ data_meta_paths (list): The paths to the metadata files containing image paths and labels.
35
+ sample_margin (int): The margin for sampling regions in the image.
36
+ processor (CLIPImageProcessor): The image processor for preprocessing images.
37
+ transform (transforms.Compose): The image augmentation transform.
38
+ """
39
+
40
+ def __init__(
41
+ self,
42
+ img_size,
43
+ drop_ratio=0.1,
44
+ data_meta_paths=None,
45
+ sample_margin=30,
46
+ ):
47
+ super().__init__()
48
+
49
+ self.img_size = img_size
50
+ self.sample_margin = sample_margin
51
+
52
+ vid_meta = []
53
+ for data_meta_path in data_meta_paths:
54
+ with open(data_meta_path, "r", encoding="utf-8") as f:
55
+ vid_meta.extend(json.load(f))
56
+ self.vid_meta = vid_meta
57
+ self.length = len(self.vid_meta)
58
+
59
+ self.clip_image_processor = CLIPImageProcessor()
60
+
61
+ self.transform = transforms.Compose(
62
+ [
63
+ transforms.Resize(self.img_size),
64
+ transforms.ToTensor(),
65
+ transforms.Normalize([0.5], [0.5]),
66
+ ]
67
+ )
68
+
69
+ self.cond_transform = transforms.Compose(
70
+ [
71
+ transforms.Resize(self.img_size),
72
+ transforms.ToTensor(),
73
+ ]
74
+ )
75
+
76
+ self.drop_ratio = drop_ratio
77
+
78
+ def augmentation(self, image, transform, state=None):
79
+ """
80
+ Apply data augmentation to the input image.
81
+
82
+ Args:
83
+ image (PIL.Image): The input image.
84
+ transform (torchvision.transforms.Compose): The data augmentation transforms.
85
+ state (dict, optional): The random state for reproducibility. Defaults to None.
86
+
87
+ Returns:
88
+ PIL.Image: The augmented image.
89
+ """
90
+ if state is not None:
91
+ torch.set_rng_state(state)
92
+ return transform(image)
93
+
94
+ def __getitem__(self, index):
95
+ video_meta = self.vid_meta[index]
96
+ video_path = video_meta["image_path"]
97
+ mask_path = video_meta["mask_path"]
98
+ face_emb_path = video_meta["face_emb"]
99
+
100
+ video_frames = sorted(Path(video_path).iterdir())
101
+ video_length = len(video_frames)
102
+
103
+ margin = min(self.sample_margin, video_length)
104
+
105
+ ref_img_idx = random.randint(0, video_length - 1)
106
+ if ref_img_idx + margin < video_length:
107
+ tgt_img_idx = random.randint(
108
+ ref_img_idx + margin, video_length - 1)
109
+ elif ref_img_idx - margin > 0:
110
+ tgt_img_idx = random.randint(0, ref_img_idx - margin)
111
+ else:
112
+ tgt_img_idx = random.randint(0, video_length - 1)
113
+
114
+ ref_img_pil = Image.open(video_frames[ref_img_idx])
115
+ tgt_img_pil = Image.open(video_frames[tgt_img_idx])
116
+
117
+ tgt_mask_pil = Image.open(mask_path)
118
+
119
+ assert ref_img_pil is not None, "Fail to load reference image."
120
+ assert tgt_img_pil is not None, "Fail to load target image."
121
+ assert tgt_mask_pil is not None, "Fail to load target mask."
122
+
123
+ state = torch.get_rng_state()
124
+ tgt_img = self.augmentation(tgt_img_pil, self.transform, state)
125
+ tgt_mask_img = self.augmentation(
126
+ tgt_mask_pil, self.cond_transform, state)
127
+ tgt_mask_img = tgt_mask_img.repeat(3, 1, 1)
128
+ ref_img_vae = self.augmentation(
129
+ ref_img_pil, self.transform, state)
130
+ face_emb = torch.load(face_emb_path)
131
+
132
+
133
+ sample = {
134
+ "video_dir": video_path,
135
+ "img": tgt_img,
136
+ "tgt_mask": tgt_mask_img,
137
+ "ref_img": ref_img_vae,
138
+ "face_emb": face_emb,
139
+ }
140
+
141
+ return sample
142
+
143
+ def __len__(self):
144
+ return len(self.vid_meta)
145
+
146
+
147
+ if __name__ == "__main__":
148
+ data = FaceMaskDataset(img_size=(512, 512))
149
+ train_dataloader = torch.utils.data.DataLoader(
150
+ data, batch_size=4, shuffle=True, num_workers=1
151
+ )
152
+ for step, batch in enumerate(train_dataloader):
153
+ print(batch["tgt_mask"].shape)
154
+ break
hallo/datasets/talk_video.py ADDED
@@ -0,0 +1,312 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # pylint: disable=R0801
2
+ """
3
+ talking_video_dataset.py
4
+
5
+ This module defines the TalkingVideoDataset class, a custom PyTorch dataset
6
+ for handling talking video data. The dataset uses video files, masks, and
7
+ embeddings to prepare data for tasks such as video generation and
8
+ speech-driven video animation.
9
+
10
+ Classes:
11
+ TalkingVideoDataset
12
+
13
+ Dependencies:
14
+ json
15
+ random
16
+ torch
17
+ decord.VideoReader, decord.cpu
18
+ PIL.Image
19
+ torch.utils.data.Dataset
20
+ torchvision.transforms
21
+
22
+ Example:
23
+ from talking_video_dataset import TalkingVideoDataset
24
+ from torch.utils.data import DataLoader
25
+
26
+ # Example configuration for the Wav2Vec model
27
+ class Wav2VecConfig:
28
+ def __init__(self, audio_type, model_scale, features):
29
+ self.audio_type = audio_type
30
+ self.model_scale = model_scale
31
+ self.features = features
32
+
33
+ wav2vec_cfg = Wav2VecConfig(audio_type="wav2vec2", model_scale="base", features="feature")
34
+
35
+ # Initialize dataset
36
+ dataset = TalkingVideoDataset(
37
+ img_size=(512, 512),
38
+ sample_rate=16000,
39
+ audio_margin=2,
40
+ n_motion_frames=0,
41
+ n_sample_frames=16,
42
+ data_meta_paths=["path/to/meta1.json", "path/to/meta2.json"],
43
+ wav2vec_cfg=wav2vec_cfg,
44
+ )
45
+
46
+ # Initialize dataloader
47
+ dataloader = DataLoader(dataset, batch_size=4, shuffle=True)
48
+
49
+ # Fetch one batch of data
50
+ batch = next(iter(dataloader))
51
+ print(batch["pixel_values_vid"].shape) # Example output: (4, 16, 3, 512, 512)
52
+
53
+ The TalkingVideoDataset class provides methods for loading video frames, masks,
54
+ audio embeddings, and other relevant data, applying transformations, and preparing
55
+ the data for training and evaluation in a deep learning pipeline.
56
+
57
+ Attributes:
58
+ img_size (tuple): The dimensions to resize the video frames to.
59
+ sample_rate (int): The audio sample rate.
60
+ audio_margin (int): The margin for audio sampling.
61
+ n_motion_frames (int): The number of motion frames.
62
+ n_sample_frames (int): The number of sample frames.
63
+ data_meta_paths (list): List of paths to the JSON metadata files.
64
+ wav2vec_cfg (object): Configuration for the Wav2Vec model.
65
+
66
+ Methods:
67
+ augmentation(images, transform, state=None): Apply transformation to input images.
68
+ __getitem__(index): Get a sample from the dataset at the specified index.
69
+ __len__(): Return the length of the dataset.
70
+ """
71
+
72
+ import json
73
+ import random
74
+ from typing import List
75
+
76
+ import torch
77
+ from decord import VideoReader, cpu
78
+ from PIL import Image
79
+ from torch.utils.data import Dataset
80
+ from torchvision import transforms
81
+
82
+
83
+ class TalkingVideoDataset(Dataset):
84
+ """
85
+ A dataset class for processing talking video data.
86
+
87
+ Args:
88
+ img_size (tuple, optional): The size of the output images. Defaults to (512, 512).
89
+ sample_rate (int, optional): The sample rate of the audio data. Defaults to 16000.
90
+ audio_margin (int, optional): The margin for the audio data. Defaults to 2.
91
+ n_motion_frames (int, optional): The number of motion frames. Defaults to 0.
92
+ n_sample_frames (int, optional): The number of sample frames. Defaults to 16.
93
+ data_meta_paths (list, optional): The paths to the data metadata. Defaults to None.
94
+ wav2vec_cfg (dict, optional): The configuration for the wav2vec model. Defaults to None.
95
+
96
+ Attributes:
97
+ img_size (tuple): The size of the output images.
98
+ sample_rate (int): The sample rate of the audio data.
99
+ audio_margin (int): The margin for the audio data.
100
+ n_motion_frames (int): The number of motion frames.
101
+ n_sample_frames (int): The number of sample frames.
102
+ data_meta_paths (list): The paths to the data metadata.
103
+ wav2vec_cfg (dict): The configuration for the wav2vec model.
104
+ """
105
+
106
+ def __init__(
107
+ self,
108
+ img_size=(512, 512),
109
+ sample_rate=16000,
110
+ audio_margin=2,
111
+ n_motion_frames=0,
112
+ n_sample_frames=16,
113
+ data_meta_paths=None,
114
+ wav2vec_cfg=None,
115
+ ):
116
+ super().__init__()
117
+ self.sample_rate = sample_rate
118
+ self.img_size = img_size
119
+ self.audio_margin = audio_margin
120
+ self.n_motion_frames = n_motion_frames
121
+ self.n_sample_frames = n_sample_frames
122
+ self.audio_type = wav2vec_cfg.audio_type
123
+ self.audio_model = wav2vec_cfg.model_scale
124
+ self.audio_features = wav2vec_cfg.features
125
+
126
+ vid_meta = []
127
+ for data_meta_path in data_meta_paths:
128
+ with open(data_meta_path, "r", encoding="utf-8") as f:
129
+ vid_meta.extend(json.load(f))
130
+ self.vid_meta = vid_meta
131
+ self.length = len(self.vid_meta)
132
+ self.pixel_transform = transforms.Compose(
133
+ [
134
+ transforms.Resize(self.img_size),
135
+ transforms.ToTensor(),
136
+ transforms.Normalize([0.5], [0.5]),
137
+ ]
138
+ )
139
+
140
+ self.cond_transform = transforms.Compose(
141
+ [
142
+ transforms.Resize(self.img_size),
143
+ transforms.ToTensor(),
144
+ ]
145
+ )
146
+ self.attn_transform_64 = transforms.Compose(
147
+ [
148
+ transforms.Resize((64,64)),
149
+ transforms.ToTensor(),
150
+ ]
151
+ )
152
+ self.attn_transform_32 = transforms.Compose(
153
+ [
154
+ transforms.Resize((32, 32)),
155
+ transforms.ToTensor(),
156
+ ]
157
+ )
158
+ self.attn_transform_16 = transforms.Compose(
159
+ [
160
+ transforms.Resize((16, 16)),
161
+ transforms.ToTensor(),
162
+ ]
163
+ )
164
+ self.attn_transform_8 = transforms.Compose(
165
+ [
166
+ transforms.Resize((8, 8)),
167
+ transforms.ToTensor(),
168
+ ]
169
+ )
170
+
171
+ def augmentation(self, images, transform, state=None):
172
+ """
173
+ Apply the given transformation to the input images.
174
+
175
+ Args:
176
+ images (List[PIL.Image] or PIL.Image): The input images to be transformed.
177
+ transform (torchvision.transforms.Compose): The transformation to be applied to the images.
178
+ state (torch.ByteTensor, optional): The state of the random number generator.
179
+ If provided, it will set the RNG state to this value before applying the transformation. Defaults to None.
180
+
181
+ Returns:
182
+ torch.Tensor: The transformed images as a tensor.
183
+ If the input was a list of images, the tensor will have shape (f, c, h, w),
184
+ where f is the number of images, c is the number of channels, h is the height, and w is the width.
185
+ If the input was a single image, the tensor will have shape (c, h, w),
186
+ where c is the number of channels, h is the height, and w is the width.
187
+ """
188
+ if state is not None:
189
+ torch.set_rng_state(state)
190
+ if isinstance(images, List):
191
+ transformed_images = [transform(img) for img in images]
192
+ ret_tensor = torch.stack(transformed_images, dim=0) # (f, c, h, w)
193
+ else:
194
+ ret_tensor = transform(images) # (c, h, w)
195
+ return ret_tensor
196
+
197
+ def __getitem__(self, index):
198
+ video_meta = self.vid_meta[index]
199
+ video_path = video_meta["video_path"]
200
+ mask_path = video_meta["mask_path"]
201
+ lip_mask_union_path = video_meta.get("sep_mask_lip", None)
202
+ face_mask_union_path = video_meta.get("sep_mask_face", None)
203
+ full_mask_union_path = video_meta.get("sep_mask_border", None)
204
+ face_emb_path = video_meta["face_emb_path"]
205
+ audio_emb_path = video_meta[
206
+ f"{self.audio_type}_emb_{self.audio_model}_{self.audio_features}"
207
+ ]
208
+ tgt_mask_pil = Image.open(mask_path)
209
+ video_frames = VideoReader(video_path, ctx=cpu(0))
210
+ assert tgt_mask_pil is not None, "Fail to load target mask."
211
+ assert (video_frames is not None and len(video_frames) > 0), "Fail to load video frames."
212
+ video_length = len(video_frames)
213
+
214
+ assert (
215
+ video_length
216
+ > self.n_sample_frames + self.n_motion_frames + 2 * self.audio_margin
217
+ )
218
+ start_idx = random.randint(
219
+ self.n_motion_frames,
220
+ video_length - self.n_sample_frames - self.audio_margin - 1,
221
+ )
222
+
223
+ videos = video_frames[start_idx : start_idx + self.n_sample_frames]
224
+
225
+ frame_list = [
226
+ Image.fromarray(video).convert("RGB") for video in videos.asnumpy()
227
+ ]
228
+
229
+ face_masks_list = [Image.open(face_mask_union_path)] * self.n_sample_frames
230
+ lip_masks_list = [Image.open(lip_mask_union_path)] * self.n_sample_frames
231
+ full_masks_list = [Image.open(full_mask_union_path)] * self.n_sample_frames
232
+ assert face_masks_list[0] is not None, "Fail to load face mask."
233
+ assert lip_masks_list[0] is not None, "Fail to load lip mask."
234
+ assert full_masks_list[0] is not None, "Fail to load full mask."
235
+
236
+
237
+ face_emb = torch.load(face_emb_path)
238
+ audio_emb = torch.load(audio_emb_path)
239
+ indices = (
240
+ torch.arange(2 * self.audio_margin + 1) - self.audio_margin
241
+ ) # Generates [-2, -1, 0, 1, 2]
242
+ center_indices = torch.arange(
243
+ start_idx,
244
+ start_idx + self.n_sample_frames,
245
+ ).unsqueeze(1) + indices.unsqueeze(0)
246
+ audio_tensor = audio_emb[center_indices]
247
+
248
+ ref_img_idx = random.randint(
249
+ self.n_motion_frames,
250
+ video_length - self.n_sample_frames - self.audio_margin - 1,
251
+ )
252
+ ref_img = video_frames[ref_img_idx].asnumpy()
253
+ ref_img = Image.fromarray(ref_img)
254
+
255
+ if self.n_motion_frames > 0:
256
+ motions = video_frames[start_idx - self.n_motion_frames : start_idx]
257
+ motion_list = [
258
+ Image.fromarray(motion).convert("RGB") for motion in motions.asnumpy()
259
+ ]
260
+
261
+ # transform
262
+ state = torch.get_rng_state()
263
+ pixel_values_vid = self.augmentation(frame_list, self.pixel_transform, state)
264
+
265
+ pixel_values_mask = self.augmentation(tgt_mask_pil, self.cond_transform, state)
266
+ pixel_values_mask = pixel_values_mask.repeat(3, 1, 1)
267
+
268
+ pixel_values_face_mask = [
269
+ self.augmentation(face_masks_list, self.attn_transform_64, state),
270
+ self.augmentation(face_masks_list, self.attn_transform_32, state),
271
+ self.augmentation(face_masks_list, self.attn_transform_16, state),
272
+ self.augmentation(face_masks_list, self.attn_transform_8, state),
273
+ ]
274
+ pixel_values_lip_mask = [
275
+ self.augmentation(lip_masks_list, self.attn_transform_64, state),
276
+ self.augmentation(lip_masks_list, self.attn_transform_32, state),
277
+ self.augmentation(lip_masks_list, self.attn_transform_16, state),
278
+ self.augmentation(lip_masks_list, self.attn_transform_8, state),
279
+ ]
280
+ pixel_values_full_mask = [
281
+ self.augmentation(full_masks_list, self.attn_transform_64, state),
282
+ self.augmentation(full_masks_list, self.attn_transform_32, state),
283
+ self.augmentation(full_masks_list, self.attn_transform_16, state),
284
+ self.augmentation(full_masks_list, self.attn_transform_8, state),
285
+ ]
286
+
287
+ pixel_values_ref_img = self.augmentation(ref_img, self.pixel_transform, state)
288
+ pixel_values_ref_img = pixel_values_ref_img.unsqueeze(0)
289
+ if self.n_motion_frames > 0:
290
+ pixel_values_motion = self.augmentation(
291
+ motion_list, self.pixel_transform, state
292
+ )
293
+ pixel_values_ref_img = torch.cat(
294
+ [pixel_values_ref_img, pixel_values_motion], dim=0
295
+ )
296
+
297
+ sample = {
298
+ "video_dir": video_path,
299
+ "pixel_values_vid": pixel_values_vid,
300
+ "pixel_values_mask": pixel_values_mask,
301
+ "pixel_values_face_mask": pixel_values_face_mask,
302
+ "pixel_values_lip_mask": pixel_values_lip_mask,
303
+ "pixel_values_full_mask": pixel_values_full_mask,
304
+ "audio_tensor": audio_tensor,
305
+ "pixel_values_ref_img": pixel_values_ref_img,
306
+ "face_emb": face_emb,
307
+ }
308
+
309
+ return sample
310
+
311
+ def __len__(self):
312
+ return len(self.vid_meta)
hallo/models/__init__.py ADDED
File without changes
hallo/models/attention.py ADDED
@@ -0,0 +1,921 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # pylint: disable=R0801
2
+ # pylint: disable=C0303
3
+
4
+ """
5
+ This module contains various transformer blocks for different applications, such as BasicTransformerBlock,
6
+ TemporalBasicTransformerBlock, and AudioTemporalBasicTransformerBlock. These blocks are used in various models,
7
+ such as GLIGEN, UNet, and others. The transformer blocks implement self-attention, cross-attention, feed-forward
8
+ networks, and other related functions.
9
+
10
+ Functions and classes included in this module are:
11
+ - BasicTransformerBlock: A basic transformer block with self-attention, cross-attention, and feed-forward layers.
12
+ - TemporalBasicTransformerBlock: A transformer block with additional temporal attention mechanisms for video data.
13
+ - AudioTemporalBasicTransformerBlock: A transformer block with additional audio-specific mechanisms for audio data.
14
+ - zero_module: A function to zero out the parameters of a given module.
15
+
16
+ For more information on each specific class and function, please refer to the respective docstrings.
17
+ """
18
+
19
+ from typing import Any, Dict, List, Optional
20
+
21
+ import torch
22
+ from diffusers.models.attention import (AdaLayerNorm, AdaLayerNormZero,
23
+ Attention, FeedForward)
24
+ from diffusers.models.embeddings import SinusoidalPositionalEmbedding
25
+ from einops import rearrange
26
+ from torch import nn
27
+
28
+
29
+ class GatedSelfAttentionDense(nn.Module):
30
+ """
31
+ A gated self-attention dense layer that combines visual features and object features.
32
+
33
+ Parameters:
34
+ query_dim (`int`): The number of channels in the query.
35
+ context_dim (`int`): The number of channels in the context.
36
+ n_heads (`int`): The number of heads to use for attention.
37
+ d_head (`int`): The number of channels in each head.
38
+ """
39
+
40
+ def __init__(self, query_dim: int, context_dim: int, n_heads: int, d_head: int):
41
+ super().__init__()
42
+
43
+ # we need a linear projection since we need cat visual feature and obj feature
44
+ self.linear = nn.Linear(context_dim, query_dim)
45
+
46
+ self.attn = Attention(query_dim=query_dim, heads=n_heads, dim_head=d_head)
47
+ self.ff = FeedForward(query_dim, activation_fn="geglu")
48
+
49
+ self.norm1 = nn.LayerNorm(query_dim)
50
+ self.norm2 = nn.LayerNorm(query_dim)
51
+
52
+ self.register_parameter("alpha_attn", nn.Parameter(torch.tensor(0.0)))
53
+ self.register_parameter("alpha_dense", nn.Parameter(torch.tensor(0.0)))
54
+
55
+ self.enabled = True
56
+
57
+ def forward(self, x: torch.Tensor, objs: torch.Tensor) -> torch.Tensor:
58
+ """
59
+ Apply the Gated Self-Attention mechanism to the input tensor `x` and object tensor `objs`.
60
+
61
+ Args:
62
+ x (torch.Tensor): The input tensor.
63
+ objs (torch.Tensor): The object tensor.
64
+
65
+ Returns:
66
+ torch.Tensor: The output tensor after applying Gated Self-Attention.
67
+ """
68
+ if not self.enabled:
69
+ return x
70
+
71
+ n_visual = x.shape[1]
72
+ objs = self.linear(objs)
73
+
74
+ x = x + self.alpha_attn.tanh() * self.attn(self.norm1(torch.cat([x, objs], dim=1)))[:, :n_visual, :]
75
+ x = x + self.alpha_dense.tanh() * self.ff(self.norm2(x))
76
+
77
+ return x
78
+
79
+ class BasicTransformerBlock(nn.Module):
80
+ r"""
81
+ A basic Transformer block.
82
+
83
+ Parameters:
84
+ dim (`int`): The number of channels in the input and output.
85
+ num_attention_heads (`int`): The number of heads to use for multi-head attention.
86
+ attention_head_dim (`int`): The number of channels in each head.
87
+ dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
88
+ cross_attention_dim (`int`, *optional*): The size of the encoder_hidden_states vector for cross attention.
89
+ activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward.
90
+ num_embeds_ada_norm (:
91
+ obj: `int`, *optional*): The number of diffusion steps used during training. See `Transformer2DModel`.
92
+ attention_bias (:
93
+ obj: `bool`, *optional*, defaults to `False`): Configure if the attentions should contain a bias parameter.
94
+ only_cross_attention (`bool`, *optional*):
95
+ Whether to use only cross-attention layers. In this case two cross attention layers are used.
96
+ double_self_attention (`bool`, *optional*):
97
+ Whether to use two self-attention layers. In this case no cross attention layers are used.
98
+ upcast_attention (`bool`, *optional*):
99
+ Whether to upcast the attention computation to float32. This is useful for mixed precision training.
100
+ norm_elementwise_affine (`bool`, *optional*, defaults to `True`):
101
+ Whether to use learnable elementwise affine parameters for normalization.
102
+ norm_type (`str`, *optional*, defaults to `"layer_norm"`):
103
+ The normalization layer to use. Can be `"layer_norm"`, `"ada_norm"` or `"ada_norm_zero"`.
104
+ final_dropout (`bool` *optional*, defaults to False):
105
+ Whether to apply a final dropout after the last feed-forward layer.
106
+ attention_type (`str`, *optional*, defaults to `"default"`):
107
+ The type of attention to use. Can be `"default"` or `"gated"` or `"gated-text-image"`.
108
+ positional_embeddings (`str`, *optional*, defaults to `None`):
109
+ The type of positional embeddings to apply to.
110
+ num_positional_embeddings (`int`, *optional*, defaults to `None`):
111
+ The maximum number of positional embeddings to apply.
112
+ """
113
+
114
+ def __init__(
115
+ self,
116
+ dim: int,
117
+ num_attention_heads: int,
118
+ attention_head_dim: int,
119
+ dropout=0.0,
120
+ cross_attention_dim: Optional[int] = None,
121
+ activation_fn: str = "geglu",
122
+ num_embeds_ada_norm: Optional[int] = None,
123
+ attention_bias: bool = False,
124
+ only_cross_attention: bool = False,
125
+ double_self_attention: bool = False,
126
+ upcast_attention: bool = False,
127
+ norm_elementwise_affine: bool = True,
128
+ # 'layer_norm', 'ada_norm', 'ada_norm_zero', 'ada_norm_single'
129
+ norm_type: str = "layer_norm",
130
+ norm_eps: float = 1e-5,
131
+ final_dropout: bool = False,
132
+ attention_type: str = "default",
133
+ positional_embeddings: Optional[str] = None,
134
+ num_positional_embeddings: Optional[int] = None,
135
+ ):
136
+ super().__init__()
137
+ self.only_cross_attention = only_cross_attention
138
+
139
+ self.use_ada_layer_norm_zero = (
140
+ num_embeds_ada_norm is not None
141
+ ) and norm_type == "ada_norm_zero"
142
+ self.use_ada_layer_norm = (
143
+ num_embeds_ada_norm is not None
144
+ ) and norm_type == "ada_norm"
145
+ self.use_ada_layer_norm_single = norm_type == "ada_norm_single"
146
+ self.use_layer_norm = norm_type == "layer_norm"
147
+
148
+ if norm_type in ("ada_norm", "ada_norm_zero") and num_embeds_ada_norm is None:
149
+ raise ValueError(
150
+ f"`norm_type` is set to {norm_type}, but `num_embeds_ada_norm` is not defined. Please make sure to"
151
+ f" define `num_embeds_ada_norm` if setting `norm_type` to {norm_type}."
152
+ )
153
+
154
+ if positional_embeddings and (num_positional_embeddings is None):
155
+ raise ValueError(
156
+ "If `positional_embedding` type is defined, `num_positition_embeddings` must also be defined."
157
+ )
158
+
159
+ if positional_embeddings == "sinusoidal":
160
+ self.pos_embed = SinusoidalPositionalEmbedding(
161
+ dim, max_seq_length=num_positional_embeddings
162
+ )
163
+ else:
164
+ self.pos_embed = None
165
+
166
+ # Define 3 blocks. Each block has its own normalization layer.
167
+ # 1. Self-Attn
168
+ if self.use_ada_layer_norm:
169
+ self.norm1 = AdaLayerNorm(dim, num_embeds_ada_norm)
170
+ elif self.use_ada_layer_norm_zero:
171
+ self.norm1 = AdaLayerNormZero(dim, num_embeds_ada_norm)
172
+ else:
173
+ self.norm1 = nn.LayerNorm(
174
+ dim, elementwise_affine=norm_elementwise_affine, eps=norm_eps
175
+ )
176
+
177
+ self.attn1 = Attention(
178
+ query_dim=dim,
179
+ heads=num_attention_heads,
180
+ dim_head=attention_head_dim,
181
+ dropout=dropout,
182
+ bias=attention_bias,
183
+ cross_attention_dim=cross_attention_dim if only_cross_attention else None,
184
+ upcast_attention=upcast_attention,
185
+ )
186
+
187
+ # 2. Cross-Attn
188
+ if cross_attention_dim is not None or double_self_attention:
189
+ # We currently only use AdaLayerNormZero for self attention where there will only be one attention block.
190
+ # I.e. the number of returned modulation chunks from AdaLayerZero would not make sense if returned during
191
+ # the second cross attention block.
192
+ self.norm2 = (
193
+ AdaLayerNorm(dim, num_embeds_ada_norm)
194
+ if self.use_ada_layer_norm
195
+ else nn.LayerNorm(
196
+ dim, elementwise_affine=norm_elementwise_affine, eps=norm_eps
197
+ )
198
+ )
199
+ self.attn2 = Attention(
200
+ query_dim=dim,
201
+ cross_attention_dim=(
202
+ cross_attention_dim if not double_self_attention else None
203
+ ),
204
+ heads=num_attention_heads,
205
+ dim_head=attention_head_dim,
206
+ dropout=dropout,
207
+ bias=attention_bias,
208
+ upcast_attention=upcast_attention,
209
+ ) # is self-attn if encoder_hidden_states is none
210
+ else:
211
+ self.norm2 = None
212
+ self.attn2 = None
213
+
214
+ # 3. Feed-forward
215
+ if not self.use_ada_layer_norm_single:
216
+ self.norm3 = nn.LayerNorm(
217
+ dim, elementwise_affine=norm_elementwise_affine, eps=norm_eps
218
+ )
219
+
220
+ self.ff = FeedForward(
221
+ dim,
222
+ dropout=dropout,
223
+ activation_fn=activation_fn,
224
+ final_dropout=final_dropout,
225
+ )
226
+
227
+ # 4. Fuser
228
+ if attention_type in {"gated", "gated-text-image"}: # Updated line
229
+ self.fuser = GatedSelfAttentionDense(
230
+ dim, cross_attention_dim, num_attention_heads, attention_head_dim
231
+ )
232
+
233
+ # 5. Scale-shift for PixArt-Alpha.
234
+ if self.use_ada_layer_norm_single:
235
+ self.scale_shift_table = nn.Parameter(
236
+ torch.randn(6, dim) / dim**0.5)
237
+
238
+ # let chunk size default to None
239
+ self._chunk_size = None
240
+ self._chunk_dim = 0
241
+
242
+ def set_chunk_feed_forward(self, chunk_size: Optional[int], dim: int = 0):
243
+ """
244
+ Sets the chunk size for feed-forward processing in the transformer block.
245
+
246
+ Args:
247
+ chunk_size (Optional[int]): The size of the chunks to process in feed-forward layers.
248
+ If None, the chunk size is set to the maximum possible value.
249
+ dim (int, optional): The dimension along which to split the input tensor into chunks. Defaults to 0.
250
+
251
+ Returns:
252
+ None.
253
+ """
254
+ self._chunk_size = chunk_size
255
+ self._chunk_dim = dim
256
+
257
+ def forward(
258
+ self,
259
+ hidden_states: torch.FloatTensor,
260
+ attention_mask: Optional[torch.FloatTensor] = None,
261
+ encoder_hidden_states: Optional[torch.FloatTensor] = None,
262
+ encoder_attention_mask: Optional[torch.FloatTensor] = None,
263
+ timestep: Optional[torch.LongTensor] = None,
264
+ cross_attention_kwargs: Dict[str, Any] = None,
265
+ class_labels: Optional[torch.LongTensor] = None,
266
+ ) -> torch.FloatTensor:
267
+ """
268
+ This function defines the forward pass of the BasicTransformerBlock.
269
+
270
+ Args:
271
+ self (BasicTransformerBlock):
272
+ An instance of the BasicTransformerBlock class.
273
+ hidden_states (torch.FloatTensor):
274
+ A tensor containing the hidden states.
275
+ attention_mask (Optional[torch.FloatTensor], optional):
276
+ A tensor containing the attention mask. Defaults to None.
277
+ encoder_hidden_states (Optional[torch.FloatTensor], optional):
278
+ A tensor containing the encoder hidden states. Defaults to None.
279
+ encoder_attention_mask (Optional[torch.FloatTensor], optional):
280
+ A tensor containing the encoder attention mask. Defaults to None.
281
+ timestep (Optional[torch.LongTensor], optional):
282
+ A tensor containing the timesteps. Defaults to None.
283
+ cross_attention_kwargs (Dict[str, Any], optional):
284
+ Additional cross-attention arguments. Defaults to None.
285
+ class_labels (Optional[torch.LongTensor], optional):
286
+ A tensor containing the class labels. Defaults to None.
287
+
288
+ Returns:
289
+ torch.FloatTensor:
290
+ A tensor containing the transformed hidden states.
291
+ """
292
+ # Notice that normalization is always applied before the real computation in the following blocks.
293
+ # 0. Self-Attention
294
+ batch_size = hidden_states.shape[0]
295
+
296
+ gate_msa = None
297
+ scale_mlp = None
298
+ shift_mlp = None
299
+ gate_mlp = None
300
+ if self.use_ada_layer_norm:
301
+ norm_hidden_states = self.norm1(hidden_states, timestep)
302
+ elif self.use_ada_layer_norm_zero:
303
+ norm_hidden_states, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.norm1(
304
+ hidden_states, timestep, class_labels, hidden_dtype=hidden_states.dtype
305
+ )
306
+ elif self.use_layer_norm:
307
+ norm_hidden_states = self.norm1(hidden_states)
308
+ elif self.use_ada_layer_norm_single:
309
+ shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = (
310
+ self.scale_shift_table[None] +
311
+ timestep.reshape(batch_size, 6, -1)
312
+ ).chunk(6, dim=1)
313
+ norm_hidden_states = self.norm1(hidden_states)
314
+ norm_hidden_states = norm_hidden_states * \
315
+ (1 + scale_msa) + shift_msa
316
+ norm_hidden_states = norm_hidden_states.squeeze(1)
317
+ else:
318
+ raise ValueError("Incorrect norm used")
319
+
320
+ if self.pos_embed is not None:
321
+ norm_hidden_states = self.pos_embed(norm_hidden_states)
322
+
323
+ # 1. Retrieve lora scale.
324
+ lora_scale = (
325
+ cross_attention_kwargs.get("scale", 1.0)
326
+ if cross_attention_kwargs is not None
327
+ else 1.0
328
+ )
329
+
330
+ # 2. Prepare GLIGEN inputs
331
+ cross_attention_kwargs = (
332
+ cross_attention_kwargs.copy() if cross_attention_kwargs is not None else {}
333
+ )
334
+ gligen_kwargs = cross_attention_kwargs.pop("gligen", None)
335
+
336
+ attn_output = self.attn1(
337
+ norm_hidden_states,
338
+ encoder_hidden_states=(
339
+ encoder_hidden_states if self.only_cross_attention else None
340
+ ),
341
+ attention_mask=attention_mask,
342
+ **cross_attention_kwargs,
343
+ )
344
+ if self.use_ada_layer_norm_zero:
345
+ attn_output = gate_msa.unsqueeze(1) * attn_output
346
+ elif self.use_ada_layer_norm_single:
347
+ attn_output = gate_msa * attn_output
348
+
349
+ hidden_states = attn_output + hidden_states
350
+ if hidden_states.ndim == 4:
351
+ hidden_states = hidden_states.squeeze(1)
352
+
353
+ # 2.5 GLIGEN Control
354
+ if gligen_kwargs is not None:
355
+ hidden_states = self.fuser(hidden_states, gligen_kwargs["objs"])
356
+
357
+ # 3. Cross-Attention
358
+ if self.attn2 is not None:
359
+ if self.use_ada_layer_norm:
360
+ norm_hidden_states = self.norm2(hidden_states, timestep)
361
+ elif self.use_ada_layer_norm_zero or self.use_layer_norm:
362
+ norm_hidden_states = self.norm2(hidden_states)
363
+ elif self.use_ada_layer_norm_single:
364
+ # For PixArt norm2 isn't applied here:
365
+ # https://github.com/PixArt-alpha/PixArt-alpha/blob/0f55e922376d8b797edd44d25d0e7464b260dcab/diffusion/model/nets/PixArtMS.py#L70C1-L76C103
366
+ norm_hidden_states = hidden_states
367
+ else:
368
+ raise ValueError("Incorrect norm")
369
+
370
+ if self.pos_embed is not None and self.use_ada_layer_norm_single is False:
371
+ norm_hidden_states = self.pos_embed(norm_hidden_states)
372
+
373
+ attn_output = self.attn2(
374
+ norm_hidden_states,
375
+ encoder_hidden_states=encoder_hidden_states,
376
+ attention_mask=encoder_attention_mask,
377
+ **cross_attention_kwargs,
378
+ )
379
+ hidden_states = attn_output + hidden_states
380
+
381
+ # 4. Feed-forward
382
+ if not self.use_ada_layer_norm_single:
383
+ norm_hidden_states = self.norm3(hidden_states)
384
+
385
+ if self.use_ada_layer_norm_zero:
386
+ norm_hidden_states = (
387
+ norm_hidden_states *
388
+ (1 + scale_mlp[:, None]) + shift_mlp[:, None]
389
+ )
390
+
391
+ if self.use_ada_layer_norm_single:
392
+ norm_hidden_states = self.norm2(hidden_states)
393
+ norm_hidden_states = norm_hidden_states * \
394
+ (1 + scale_mlp) + shift_mlp
395
+
396
+ ff_output = self.ff(norm_hidden_states, scale=lora_scale)
397
+
398
+ if self.use_ada_layer_norm_zero:
399
+ ff_output = gate_mlp.unsqueeze(1) * ff_output
400
+ elif self.use_ada_layer_norm_single:
401
+ ff_output = gate_mlp * ff_output
402
+
403
+ hidden_states = ff_output + hidden_states
404
+ if hidden_states.ndim == 4:
405
+ hidden_states = hidden_states.squeeze(1)
406
+
407
+ return hidden_states
408
+
409
+
410
+ class TemporalBasicTransformerBlock(nn.Module):
411
+ """
412
+ A PyTorch module that extends the BasicTransformerBlock to include temporal attention mechanisms.
413
+ This class is particularly useful for video-related tasks where capturing temporal information within the sequence of frames is necessary.
414
+
415
+ Attributes:
416
+ dim (int): The dimension of the input and output embeddings.
417
+ num_attention_heads (int): The number of attention heads in the multi-head self-attention mechanism.
418
+ attention_head_dim (int): The dimension of each attention head.
419
+ dropout (float): The dropout probability for the attention scores.
420
+ cross_attention_dim (Optional[int]): The dimension of the cross-attention mechanism.
421
+ activation_fn (str): The activation function used in the feed-forward layer.
422
+ num_embeds_ada_norm (Optional[int]): The number of embeddings for adaptive normalization.
423
+ attention_bias (bool): If True, uses bias in the attention mechanism.
424
+ only_cross_attention (bool): If True, only uses cross-attention.
425
+ upcast_attention (bool): If True, upcasts the attention mechanism for better performance.
426
+ unet_use_cross_frame_attention (Optional[bool]): If True, uses cross-frame attention in the UNet model.
427
+ unet_use_temporal_attention (Optional[bool]): If True, uses temporal attention in the UNet model.
428
+ """
429
+ def __init__(
430
+ self,
431
+ dim: int,
432
+ num_attention_heads: int,
433
+ attention_head_dim: int,
434
+ dropout=0.0,
435
+ cross_attention_dim: Optional[int] = None,
436
+ activation_fn: str = "geglu",
437
+ num_embeds_ada_norm: Optional[int] = None,
438
+ attention_bias: bool = False,
439
+ only_cross_attention: bool = False,
440
+ upcast_attention: bool = False,
441
+ unet_use_cross_frame_attention=None,
442
+ unet_use_temporal_attention=None,
443
+ ):
444
+ """
445
+ The TemporalBasicTransformerBlock class is a PyTorch module that extends the BasicTransformerBlock to include temporal attention mechanisms.
446
+ This is particularly useful for video-related tasks, where the model needs to capture the temporal information within the sequence of frames.
447
+ The block consists of self-attention, cross-attention, feed-forward, and temporal attention mechanisms.
448
+
449
+ dim (int): The dimension of the input and output embeddings.
450
+ num_attention_heads (int): The number of attention heads in the multi-head self-attention mechanism.
451
+ attention_head_dim (int): The dimension of each attention head.
452
+ dropout (float, optional): The dropout probability for the attention scores. Defaults to 0.0.
453
+ cross_attention_dim (int, optional): The dimension of the cross-attention mechanism. Defaults to None.
454
+ activation_fn (str, optional): The activation function used in the feed-forward layer. Defaults to "geglu".
455
+ num_embeds_ada_norm (int, optional): The number of embeddings for adaptive normalization. Defaults to None.
456
+ attention_bias (bool, optional): If True, uses bias in the attention mechanism. Defaults to False.
457
+ only_cross_attention (bool, optional): If True, only uses cross-attention. Defaults to False.
458
+ upcast_attention (bool, optional): If True, upcasts the attention mechanism for better performance. Defaults to False.
459
+ unet_use_cross_frame_attention (bool, optional): If True, uses cross-frame attention in the UNet model. Defaults to None.
460
+ unet_use_temporal_attention (bool, optional): If True, uses temporal attention in the UNet model. Defaults to None.
461
+
462
+ Forward method:
463
+ hidden_states (torch.FloatTensor): The input hidden states.
464
+ encoder_hidden_states (torch.FloatTensor, optional): The encoder hidden states. Defaults to None.
465
+ timestep (torch.LongTensor, optional): The current timestep for the transformer model. Defaults to None.
466
+ attention_mask (torch.FloatTensor, optional): The attention mask for the self-attention mechanism. Defaults to None.
467
+ video_length (int, optional): The length of the video sequence. Defaults to None.
468
+
469
+ Returns:
470
+ torch.FloatTensor: The output hidden states after passing through the TemporalBasicTransformerBlock.
471
+ """
472
+ super().__init__()
473
+ self.only_cross_attention = only_cross_attention
474
+ self.use_ada_layer_norm = num_embeds_ada_norm is not None
475
+ self.unet_use_cross_frame_attention = unet_use_cross_frame_attention
476
+ self.unet_use_temporal_attention = unet_use_temporal_attention
477
+
478
+ # SC-Attn
479
+ self.attn1 = Attention(
480
+ query_dim=dim,
481
+ heads=num_attention_heads,
482
+ dim_head=attention_head_dim,
483
+ dropout=dropout,
484
+ bias=attention_bias,
485
+ upcast_attention=upcast_attention,
486
+ )
487
+ self.norm1 = (
488
+ AdaLayerNorm(dim, num_embeds_ada_norm)
489
+ if self.use_ada_layer_norm
490
+ else nn.LayerNorm(dim)
491
+ )
492
+
493
+ # Cross-Attn
494
+ if cross_attention_dim is not None:
495
+ self.attn2 = Attention(
496
+ query_dim=dim,
497
+ cross_attention_dim=cross_attention_dim,
498
+ heads=num_attention_heads,
499
+ dim_head=attention_head_dim,
500
+ dropout=dropout,
501
+ bias=attention_bias,
502
+ upcast_attention=upcast_attention,
503
+ )
504
+ else:
505
+ self.attn2 = None
506
+
507
+ if cross_attention_dim is not None:
508
+ self.norm2 = (
509
+ AdaLayerNorm(dim, num_embeds_ada_norm)
510
+ if self.use_ada_layer_norm
511
+ else nn.LayerNorm(dim)
512
+ )
513
+ else:
514
+ self.norm2 = None
515
+
516
+ # Feed-forward
517
+ self.ff = FeedForward(dim, dropout=dropout,
518
+ activation_fn=activation_fn)
519
+ self.norm3 = nn.LayerNorm(dim)
520
+ self.use_ada_layer_norm_zero = False
521
+
522
+ # Temp-Attn
523
+ # assert unet_use_temporal_attention is not None
524
+ if unet_use_temporal_attention is None:
525
+ unet_use_temporal_attention = False
526
+ if unet_use_temporal_attention:
527
+ self.attn_temp = Attention(
528
+ query_dim=dim,
529
+ heads=num_attention_heads,
530
+ dim_head=attention_head_dim,
531
+ dropout=dropout,
532
+ bias=attention_bias,
533
+ upcast_attention=upcast_attention,
534
+ )
535
+ nn.init.zeros_(self.attn_temp.to_out[0].weight.data)
536
+ self.norm_temp = (
537
+ AdaLayerNorm(dim, num_embeds_ada_norm)
538
+ if self.use_ada_layer_norm
539
+ else nn.LayerNorm(dim)
540
+ )
541
+
542
+ def forward(
543
+ self,
544
+ hidden_states,
545
+ encoder_hidden_states=None,
546
+ timestep=None,
547
+ attention_mask=None,
548
+ video_length=None,
549
+ ):
550
+ """
551
+ Forward pass for the TemporalBasicTransformerBlock.
552
+
553
+ Args:
554
+ hidden_states (torch.FloatTensor): The input hidden states with shape (batch_size, seq_len, dim).
555
+ encoder_hidden_states (torch.FloatTensor, optional): The encoder hidden states with shape (batch_size, src_seq_len, dim).
556
+ timestep (torch.LongTensor, optional): The timestep for the transformer block.
557
+ attention_mask (torch.FloatTensor, optional): The attention mask with shape (batch_size, seq_len, seq_len).
558
+ video_length (int, optional): The length of the video sequence.
559
+
560
+ Returns:
561
+ torch.FloatTensor: The output tensor after passing through the transformer block with shape (batch_size, seq_len, dim).
562
+ """
563
+ norm_hidden_states = (
564
+ self.norm1(hidden_states, timestep)
565
+ if self.use_ada_layer_norm
566
+ else self.norm1(hidden_states)
567
+ )
568
+
569
+ if self.unet_use_cross_frame_attention:
570
+ hidden_states = (
571
+ self.attn1(
572
+ norm_hidden_states,
573
+ attention_mask=attention_mask,
574
+ video_length=video_length,
575
+ )
576
+ + hidden_states
577
+ )
578
+ else:
579
+ hidden_states = (
580
+ self.attn1(norm_hidden_states, attention_mask=attention_mask)
581
+ + hidden_states
582
+ )
583
+
584
+ if self.attn2 is not None:
585
+ # Cross-Attention
586
+ norm_hidden_states = (
587
+ self.norm2(hidden_states, timestep)
588
+ if self.use_ada_layer_norm
589
+ else self.norm2(hidden_states)
590
+ )
591
+ hidden_states = (
592
+ self.attn2(
593
+ norm_hidden_states,
594
+ encoder_hidden_states=encoder_hidden_states,
595
+ attention_mask=attention_mask,
596
+ )
597
+ + hidden_states
598
+ )
599
+
600
+ # Feed-forward
601
+ hidden_states = self.ff(self.norm3(hidden_states)) + hidden_states
602
+
603
+ # Temporal-Attention
604
+ if self.unet_use_temporal_attention:
605
+ d = hidden_states.shape[1]
606
+ hidden_states = rearrange(
607
+ hidden_states, "(b f) d c -> (b d) f c", f=video_length
608
+ )
609
+ norm_hidden_states = (
610
+ self.norm_temp(hidden_states, timestep)
611
+ if self.use_ada_layer_norm
612
+ else self.norm_temp(hidden_states)
613
+ )
614
+ hidden_states = self.attn_temp(norm_hidden_states) + hidden_states
615
+ hidden_states = rearrange(
616
+ hidden_states, "(b d) f c -> (b f) d c", d=d)
617
+
618
+ return hidden_states
619
+
620
+
621
+ class AudioTemporalBasicTransformerBlock(nn.Module):
622
+ """
623
+ A PyTorch module designed to handle audio data within a transformer framework, including temporal attention mechanisms.
624
+
625
+ Attributes:
626
+ dim (int): The dimension of the input and output embeddings.
627
+ num_attention_heads (int): The number of attention heads.
628
+ attention_head_dim (int): The dimension of each attention head.
629
+ dropout (float): The dropout probability.
630
+ cross_attention_dim (Optional[int]): The dimension of the cross-attention mechanism.
631
+ activation_fn (str): The activation function for the feed-forward network.
632
+ num_embeds_ada_norm (Optional[int]): The number of embeddings for adaptive normalization.
633
+ attention_bias (bool): If True, uses bias in the attention mechanism.
634
+ only_cross_attention (bool): If True, only uses cross-attention.
635
+ upcast_attention (bool): If True, upcasts the attention mechanism to float32.
636
+ unet_use_cross_frame_attention (Optional[bool]): If True, uses cross-frame attention in UNet.
637
+ unet_use_temporal_attention (Optional[bool]): If True, uses temporal attention in UNet.
638
+ depth (int): The depth of the transformer block.
639
+ unet_block_name (Optional[str]): The name of the UNet block.
640
+ stack_enable_blocks_name (Optional[List[str]]): The list of enabled blocks in the stack.
641
+ stack_enable_blocks_depth (Optional[List[int]]): The list of depths for the enabled blocks in the stack.
642
+ """
643
+ def __init__(
644
+ self,
645
+ dim: int,
646
+ num_attention_heads: int,
647
+ attention_head_dim: int,
648
+ dropout=0.0,
649
+ cross_attention_dim: Optional[int] = None,
650
+ activation_fn: str = "geglu",
651
+ num_embeds_ada_norm: Optional[int] = None,
652
+ attention_bias: bool = False,
653
+ only_cross_attention: bool = False,
654
+ upcast_attention: bool = False,
655
+ unet_use_cross_frame_attention=None,
656
+ unet_use_temporal_attention=None,
657
+ depth=0,
658
+ unet_block_name=None,
659
+ stack_enable_blocks_name: Optional[List[str]] = None,
660
+ stack_enable_blocks_depth: Optional[List[int]] = None,
661
+ ):
662
+ """
663
+ Initializes the AudioTemporalBasicTransformerBlock module.
664
+
665
+ Args:
666
+ dim (int): The dimension of the input and output embeddings.
667
+ num_attention_heads (int): The number of attention heads in the multi-head self-attention mechanism.
668
+ attention_head_dim (int): The dimension of each attention head.
669
+ dropout (float, optional): The dropout probability for the attention mechanism. Defaults to 0.0.
670
+ cross_attention_dim (Optional[int], optional): The dimension of the cross-attention mechanism. Defaults to None.
671
+ activation_fn (str, optional): The activation function to be used in the feed-forward network. Defaults to "geglu".
672
+ num_embeds_ada_norm (Optional[int], optional): The number of embeddings for adaptive normalization. Defaults to None.
673
+ attention_bias (bool, optional): If True, uses bias in the attention mechanism. Defaults to False.
674
+ only_cross_attention (bool, optional): If True, only uses cross-attention. Defaults to False.
675
+ upcast_attention (bool, optional): If True, upcasts the attention mechanism to float32. Defaults to False.
676
+ unet_use_cross_frame_attention (Optional[bool], optional): If True, uses cross-frame attention in UNet. Defaults to None.
677
+ unet_use_temporal_attention (Optional[bool], optional): If True, uses temporal attention in UNet. Defaults to None.
678
+ depth (int, optional): The depth of the transformer block. Defaults to 0.
679
+ unet_block_name (Optional[str], optional): The name of the UNet block. Defaults to None.
680
+ stack_enable_blocks_name (Optional[List[str]], optional): The list of enabled blocks in the stack. Defaults to None.
681
+ stack_enable_blocks_depth (Optional[List[int]], optional): The list of depths for the enabled blocks in the stack. Defaults to None.
682
+ """
683
+ super().__init__()
684
+ self.only_cross_attention = only_cross_attention
685
+ self.use_ada_layer_norm = num_embeds_ada_norm is not None
686
+ self.unet_use_cross_frame_attention = unet_use_cross_frame_attention
687
+ self.unet_use_temporal_attention = unet_use_temporal_attention
688
+ self.unet_block_name = unet_block_name
689
+ self.depth = depth
690
+
691
+ zero_conv_full = nn.Conv2d(
692
+ dim, dim, kernel_size=1)
693
+ self.zero_conv_full = zero_module(zero_conv_full)
694
+
695
+ zero_conv_face = nn.Conv2d(
696
+ dim, dim, kernel_size=1)
697
+ self.zero_conv_face = zero_module(zero_conv_face)
698
+
699
+ zero_conv_lip = nn.Conv2d(
700
+ dim, dim, kernel_size=1)
701
+ self.zero_conv_lip = zero_module(zero_conv_lip)
702
+ # SC-Attn
703
+ self.attn1 = Attention(
704
+ query_dim=dim,
705
+ heads=num_attention_heads,
706
+ dim_head=attention_head_dim,
707
+ dropout=dropout,
708
+ bias=attention_bias,
709
+ upcast_attention=upcast_attention,
710
+ )
711
+ self.norm1 = (
712
+ AdaLayerNorm(dim, num_embeds_ada_norm)
713
+ if self.use_ada_layer_norm
714
+ else nn.LayerNorm(dim)
715
+ )
716
+
717
+ # Cross-Attn
718
+ if cross_attention_dim is not None:
719
+ if (stack_enable_blocks_name is not None and
720
+ stack_enable_blocks_depth is not None and
721
+ self.unet_block_name in stack_enable_blocks_name and
722
+ self.depth in stack_enable_blocks_depth):
723
+ self.attn2_0 = Attention(
724
+ query_dim=dim,
725
+ cross_attention_dim=cross_attention_dim,
726
+ heads=num_attention_heads,
727
+ dim_head=attention_head_dim,
728
+ dropout=dropout,
729
+ bias=attention_bias,
730
+ upcast_attention=upcast_attention,
731
+ )
732
+ self.attn2_1 = Attention(
733
+ query_dim=dim,
734
+ cross_attention_dim=cross_attention_dim,
735
+ heads=num_attention_heads,
736
+ dim_head=attention_head_dim,
737
+ dropout=dropout,
738
+ bias=attention_bias,
739
+ upcast_attention=upcast_attention,
740
+ )
741
+ self.attn2_2 = Attention(
742
+ query_dim=dim,
743
+ cross_attention_dim=cross_attention_dim,
744
+ heads=num_attention_heads,
745
+ dim_head=attention_head_dim,
746
+ dropout=dropout,
747
+ bias=attention_bias,
748
+ upcast_attention=upcast_attention,
749
+ )
750
+ self.attn2 = None
751
+
752
+ else:
753
+ self.attn2 = Attention(
754
+ query_dim=dim,
755
+ cross_attention_dim=cross_attention_dim,
756
+ heads=num_attention_heads,
757
+ dim_head=attention_head_dim,
758
+ dropout=dropout,
759
+ bias=attention_bias,
760
+ upcast_attention=upcast_attention,
761
+ )
762
+ self.attn2_0=None
763
+ else:
764
+ self.attn2 = None
765
+ self.attn2_0 = None
766
+
767
+ if cross_attention_dim is not None:
768
+ self.norm2 = (
769
+ AdaLayerNorm(dim, num_embeds_ada_norm)
770
+ if self.use_ada_layer_norm
771
+ else nn.LayerNorm(dim)
772
+ )
773
+ else:
774
+ self.norm2 = None
775
+
776
+ # Feed-forward
777
+ self.ff = FeedForward(dim, dropout=dropout,
778
+ activation_fn=activation_fn)
779
+ self.norm3 = nn.LayerNorm(dim)
780
+ self.use_ada_layer_norm_zero = False
781
+
782
+
783
+
784
+ def forward(
785
+ self,
786
+ hidden_states,
787
+ encoder_hidden_states=None,
788
+ timestep=None,
789
+ attention_mask=None,
790
+ full_mask=None,
791
+ face_mask=None,
792
+ lip_mask=None,
793
+ motion_scale=None,
794
+ video_length=None,
795
+ ):
796
+ """
797
+ Forward pass for the AudioTemporalBasicTransformerBlock.
798
+
799
+ Args:
800
+ hidden_states (torch.FloatTensor): The input hidden states.
801
+ encoder_hidden_states (torch.FloatTensor, optional): The encoder hidden states. Defaults to None.
802
+ timestep (torch.LongTensor, optional): The timestep for the transformer block. Defaults to None.
803
+ attention_mask (torch.FloatTensor, optional): The attention mask. Defaults to None.
804
+ full_mask (torch.FloatTensor, optional): The full mask. Defaults to None.
805
+ face_mask (torch.FloatTensor, optional): The face mask. Defaults to None.
806
+ lip_mask (torch.FloatTensor, optional): The lip mask. Defaults to None.
807
+ video_length (int, optional): The length of the video. Defaults to None.
808
+
809
+ Returns:
810
+ torch.FloatTensor: The output tensor after passing through the AudioTemporalBasicTransformerBlock.
811
+ """
812
+ norm_hidden_states = (
813
+ self.norm1(hidden_states, timestep)
814
+ if self.use_ada_layer_norm
815
+ else self.norm1(hidden_states)
816
+ )
817
+
818
+ if self.unet_use_cross_frame_attention:
819
+ hidden_states = (
820
+ self.attn1(
821
+ norm_hidden_states,
822
+ attention_mask=attention_mask,
823
+ video_length=video_length,
824
+ )
825
+ + hidden_states
826
+ )
827
+ else:
828
+ hidden_states = (
829
+ self.attn1(norm_hidden_states, attention_mask=attention_mask)
830
+ + hidden_states
831
+ )
832
+
833
+ if self.attn2 is not None:
834
+ # Cross-Attention
835
+ norm_hidden_states = (
836
+ self.norm2(hidden_states, timestep)
837
+ if self.use_ada_layer_norm
838
+ else self.norm2(hidden_states)
839
+ )
840
+ hidden_states = self.attn2(
841
+ norm_hidden_states,
842
+ encoder_hidden_states=encoder_hidden_states,
843
+ attention_mask=attention_mask,
844
+ ) + hidden_states
845
+
846
+ elif self.attn2_0 is not None:
847
+ norm_hidden_states = (
848
+ self.norm2(hidden_states, timestep)
849
+ if self.use_ada_layer_norm
850
+ else self.norm2(hidden_states)
851
+ )
852
+
853
+ level = self.depth
854
+ full_hidden_states = (
855
+ self.attn2_0(
856
+ norm_hidden_states,
857
+ encoder_hidden_states=encoder_hidden_states,
858
+ attention_mask=attention_mask,
859
+ ) * full_mask[level][:, :, None]
860
+ )
861
+ bz, sz, c = full_hidden_states.shape
862
+ sz_sqrt = int(sz ** 0.5)
863
+ full_hidden_states = full_hidden_states.reshape(
864
+ bz, sz_sqrt, sz_sqrt, c).permute(0, 3, 1, 2)
865
+ full_hidden_states = self.zero_conv_full(full_hidden_states).permute(0, 2, 3, 1).reshape(bz, -1, c)
866
+
867
+ face_hidden_state = (
868
+ self.attn2_1(
869
+ norm_hidden_states,
870
+ encoder_hidden_states=encoder_hidden_states,
871
+ attention_mask=attention_mask,
872
+ ) * face_mask[level][:, :, None]
873
+ )
874
+ face_hidden_state = face_hidden_state.reshape(
875
+ bz, sz_sqrt, sz_sqrt, c).permute(0, 3, 1, 2)
876
+ face_hidden_state = self.zero_conv_face(
877
+ face_hidden_state).permute(0, 2, 3, 1).reshape(bz, -1, c)
878
+
879
+ lip_hidden_state = (
880
+ self.attn2_2(
881
+ norm_hidden_states,
882
+ encoder_hidden_states=encoder_hidden_states,
883
+ attention_mask=attention_mask,
884
+ ) * lip_mask[level][:, :, None]
885
+
886
+ ) # [32, 4096, 320]
887
+ lip_hidden_state = lip_hidden_state.reshape(
888
+ bz, sz_sqrt, sz_sqrt, c).permute(0, 3, 1, 2)
889
+ lip_hidden_state = self.zero_conv_lip(
890
+ lip_hidden_state).permute(0, 2, 3, 1).reshape(bz, -1, c)
891
+
892
+ if motion_scale is not None:
893
+ hidden_states = (
894
+ motion_scale[0] * full_hidden_states +
895
+ motion_scale[1] * face_hidden_state +
896
+ motion_scale[2] * lip_hidden_state + hidden_states
897
+ )
898
+ else:
899
+ hidden_states = (
900
+ full_hidden_states +
901
+ face_hidden_state +
902
+ lip_hidden_state + hidden_states
903
+ )
904
+ # Feed-forward
905
+ hidden_states = self.ff(self.norm3(hidden_states)) + hidden_states
906
+
907
+ return hidden_states
908
+
909
+ def zero_module(module):
910
+ """
911
+ Zeroes out the parameters of a given module.
912
+
913
+ Args:
914
+ module (nn.Module): The module whose parameters need to be zeroed out.
915
+
916
+ Returns:
917
+ None.
918
+ """
919
+ for p in module.parameters():
920
+ nn.init.zeros_(p)
921
+ return module
hallo/models/audio_proj.py ADDED
@@ -0,0 +1,124 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """
2
+ This module provides the implementation of an Audio Projection Model, which is designed for
3
+ audio processing tasks. The model takes audio embeddings as input and outputs context tokens
4
+ that can be used for various downstream applications, such as audio analysis or synthesis.
5
+
6
+ The AudioProjModel class is based on the ModelMixin class from the diffusers library, which
7
+ provides a foundation for building custom models. This implementation includes multiple linear
8
+ layers with ReLU activation functions and a LayerNorm for normalization.
9
+
10
+ Key Features:
11
+ - Audio embedding input with flexible sequence length and block structure.
12
+ - Multiple linear layers for feature transformation.
13
+ - ReLU activation for non-linear transformation.
14
+ - LayerNorm for stabilizing and speeding up training.
15
+ - Rearrangement of input embeddings to match the model's expected input shape.
16
+ - Customizable number of blocks, channels, and context tokens for adaptability.
17
+
18
+ The module is structured to be easily integrated into larger systems or used as a standalone
19
+ component for audio feature extraction and processing.
20
+
21
+ Classes:
22
+ - AudioProjModel: A class representing the audio projection model with configurable parameters.
23
+
24
+ Functions:
25
+ - (none)
26
+
27
+ Dependencies:
28
+ - torch: For tensor operations and neural network components.
29
+ - diffusers: For the ModelMixin base class.
30
+ - einops: For tensor rearrangement operations.
31
+
32
+ """
33
+
34
+ import torch
35
+ from diffusers import ModelMixin
36
+ from einops import rearrange
37
+ from torch import nn
38
+
39
+
40
+ class AudioProjModel(ModelMixin):
41
+ """Audio Projection Model
42
+
43
+ This class defines an audio projection model that takes audio embeddings as input
44
+ and produces context tokens as output. The model is based on the ModelMixin class
45
+ and consists of multiple linear layers and activation functions. It can be used
46
+ for various audio processing tasks.
47
+
48
+ Attributes:
49
+ seq_len (int): The length of the audio sequence.
50
+ blocks (int): The number of blocks in the audio projection model.
51
+ channels (int): The number of channels in the audio projection model.
52
+ intermediate_dim (int): The intermediate dimension of the model.
53
+ context_tokens (int): The number of context tokens in the output.
54
+ output_dim (int): The output dimension of the context tokens.
55
+
56
+ Methods:
57
+ __init__(self, seq_len=5, blocks=12, channels=768, intermediate_dim=512, context_tokens=32, output_dim=768):
58
+ Initializes the AudioProjModel with the given parameters.
59
+ forward(self, audio_embeds):
60
+ Defines the forward pass for the AudioProjModel.
61
+ Parameters:
62
+ audio_embeds (torch.Tensor): The input audio embeddings with shape (batch_size, video_length, blocks, channels).
63
+ Returns:
64
+ context_tokens (torch.Tensor): The output context tokens with shape (batch_size, video_length, context_tokens, output_dim).
65
+
66
+ """
67
+
68
+ def __init__(
69
+ self,
70
+ seq_len=5,
71
+ blocks=12, # add a new parameter blocks
72
+ channels=768, # add a new parameter channels
73
+ intermediate_dim=512,
74
+ output_dim=768,
75
+ context_tokens=32,
76
+ ):
77
+ super().__init__()
78
+
79
+ self.seq_len = seq_len
80
+ self.blocks = blocks
81
+ self.channels = channels
82
+ self.input_dim = (
83
+ seq_len * blocks * channels
84
+ ) # update input_dim to be the product of blocks and channels.
85
+ self.intermediate_dim = intermediate_dim
86
+ self.context_tokens = context_tokens
87
+ self.output_dim = output_dim
88
+
89
+ # define multiple linear layers
90
+ self.proj1 = nn.Linear(self.input_dim, intermediate_dim)
91
+ self.proj2 = nn.Linear(intermediate_dim, intermediate_dim)
92
+ self.proj3 = nn.Linear(intermediate_dim, context_tokens * output_dim)
93
+
94
+ self.norm = nn.LayerNorm(output_dim)
95
+
96
+ def forward(self, audio_embeds):
97
+ """
98
+ Defines the forward pass for the AudioProjModel.
99
+
100
+ Parameters:
101
+ audio_embeds (torch.Tensor): The input audio embeddings with shape (batch_size, video_length, blocks, channels).
102
+
103
+ Returns:
104
+ context_tokens (torch.Tensor): The output context tokens with shape (batch_size, video_length, context_tokens, output_dim).
105
+ """
106
+ # merge
107
+ video_length = audio_embeds.shape[1]
108
+ audio_embeds = rearrange(audio_embeds, "bz f w b c -> (bz f) w b c")
109
+ batch_size, window_size, blocks, channels = audio_embeds.shape
110
+ audio_embeds = audio_embeds.view(batch_size, window_size * blocks * channels)
111
+
112
+ audio_embeds = torch.relu(self.proj1(audio_embeds))
113
+ audio_embeds = torch.relu(self.proj2(audio_embeds))
114
+
115
+ context_tokens = self.proj3(audio_embeds).reshape(
116
+ batch_size, self.context_tokens, self.output_dim
117
+ )
118
+
119
+ context_tokens = self.norm(context_tokens)
120
+ context_tokens = rearrange(
121
+ context_tokens, "(bz f) m c -> bz f m c", f=video_length
122
+ )
123
+
124
+ return context_tokens
hallo/models/face_locator.py ADDED
@@ -0,0 +1,113 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """
2
+ This module implements the FaceLocator class, which is a neural network model designed to
3
+ locate and extract facial features from input images or tensors. It uses a series of
4
+ convolutional layers to progressively downsample and refine the facial feature map.
5
+
6
+ The FaceLocator class is part of a larger system that may involve facial recognition or
7
+ similar tasks where precise location and extraction of facial features are required.
8
+
9
+ Attributes:
10
+ conditioning_embedding_channels (int): The number of channels in the output embedding.
11
+ conditioning_channels (int): The number of input channels for the conditioning tensor.
12
+ block_out_channels (Tuple[int]): A tuple of integers representing the output channels
13
+ for each block in the model.
14
+
15
+ The model uses the following components:
16
+ - InflatedConv3d: A convolutional layer that inflates the input to increase the depth.
17
+ - zero_module: A utility function that may set certain parameters to zero for regularization
18
+ or other purposes.
19
+
20
+ The forward method of the FaceLocator class takes a conditioning tensor as input and
21
+ produces an embedding tensor as output, which can be used for further processing or analysis.
22
+ """
23
+
24
+ from typing import Tuple
25
+
26
+ import torch.nn.functional as F
27
+ from diffusers.models.modeling_utils import ModelMixin
28
+ from torch import nn
29
+
30
+ from .motion_module import zero_module
31
+ from .resnet import InflatedConv3d
32
+
33
+
34
+ class FaceLocator(ModelMixin):
35
+ """
36
+ The FaceLocator class is a neural network model designed to process and extract facial
37
+ features from an input tensor. It consists of a series of convolutional layers that
38
+ progressively downsample the input while increasing the depth of the feature map.
39
+
40
+ The model is built using InflatedConv3d layers, which are designed to inflate the
41
+ feature channels, allowing for more complex feature extraction. The final output is a
42
+ conditioning embedding that can be used for various tasks such as facial recognition or
43
+ feature-based image manipulation.
44
+
45
+ Parameters:
46
+ conditioning_embedding_channels (int): The number of channels in the output embedding.
47
+ conditioning_channels (int, optional): The number of input channels for the conditioning tensor. Default is 3.
48
+ block_out_channels (Tuple[int], optional): A tuple of integers representing the output channels
49
+ for each block in the model. The default is (16, 32, 64, 128), which defines the
50
+ progression of the network's depth.
51
+
52
+ Attributes:
53
+ conv_in (InflatedConv3d): The initial convolutional layer that starts the feature extraction process.
54
+ blocks (ModuleList[InflatedConv3d]): A list of convolutional layers that form the core of the model.
55
+ conv_out (InflatedConv3d): The final convolutional layer that produces the output embedding.
56
+
57
+ The forward method applies the convolutional layers to the input conditioning tensor and
58
+ returns the resulting embedding tensor.
59
+ """
60
+ def __init__(
61
+ self,
62
+ conditioning_embedding_channels: int,
63
+ conditioning_channels: int = 3,
64
+ block_out_channels: Tuple[int] = (16, 32, 64, 128),
65
+ ):
66
+ super().__init__()
67
+ self.conv_in = InflatedConv3d(
68
+ conditioning_channels, block_out_channels[0], kernel_size=3, padding=1
69
+ )
70
+
71
+ self.blocks = nn.ModuleList([])
72
+
73
+ for i in range(len(block_out_channels) - 1):
74
+ channel_in = block_out_channels[i]
75
+ channel_out = block_out_channels[i + 1]
76
+ self.blocks.append(
77
+ InflatedConv3d(channel_in, channel_in, kernel_size=3, padding=1)
78
+ )
79
+ self.blocks.append(
80
+ InflatedConv3d(
81
+ channel_in, channel_out, kernel_size=3, padding=1, stride=2
82
+ )
83
+ )
84
+
85
+ self.conv_out = zero_module(
86
+ InflatedConv3d(
87
+ block_out_channels[-1],
88
+ conditioning_embedding_channels,
89
+ kernel_size=3,
90
+ padding=1,
91
+ )
92
+ )
93
+
94
+ def forward(self, conditioning):
95
+ """
96
+ Forward pass of the FaceLocator model.
97
+
98
+ Args:
99
+ conditioning (Tensor): The input conditioning tensor.
100
+
101
+ Returns:
102
+ Tensor: The output embedding tensor.
103
+ """
104
+ embedding = self.conv_in(conditioning)
105
+ embedding = F.silu(embedding)
106
+
107
+ for block in self.blocks:
108
+ embedding = block(embedding)
109
+ embedding = F.silu(embedding)
110
+
111
+ embedding = self.conv_out(embedding)
112
+
113
+ return embedding
hallo/models/image_proj.py ADDED
@@ -0,0 +1,76 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """
2
+ image_proj_model.py
3
+
4
+ This module defines the ImageProjModel class, which is responsible for
5
+ projecting image embeddings into a different dimensional space. The model
6
+ leverages a linear transformation followed by a layer normalization to
7
+ reshape and normalize the input image embeddings for further processing in
8
+ cross-attention mechanisms or other downstream tasks.
9
+
10
+ Classes:
11
+ ImageProjModel
12
+
13
+ Dependencies:
14
+ torch
15
+ diffusers.ModelMixin
16
+
17
+ """
18
+
19
+ import torch
20
+ from diffusers import ModelMixin
21
+
22
+
23
+ class ImageProjModel(ModelMixin):
24
+ """
25
+ ImageProjModel is a class that projects image embeddings into a different
26
+ dimensional space. It inherits from ModelMixin, providing additional functionalities
27
+ specific to image projection.
28
+
29
+ Attributes:
30
+ cross_attention_dim (int): The dimension of the cross attention.
31
+ clip_embeddings_dim (int): The dimension of the CLIP embeddings.
32
+ clip_extra_context_tokens (int): The number of extra context tokens in CLIP.
33
+
34
+ Methods:
35
+ forward(image_embeds): Forward pass of the ImageProjModel, which takes in image
36
+ embeddings and returns the projected tokens.
37
+
38
+ """
39
+
40
+ def __init__(
41
+ self,
42
+ cross_attention_dim=1024,
43
+ clip_embeddings_dim=1024,
44
+ clip_extra_context_tokens=4,
45
+ ):
46
+ super().__init__()
47
+
48
+ self.generator = None
49
+ self.cross_attention_dim = cross_attention_dim
50
+ self.clip_extra_context_tokens = clip_extra_context_tokens
51
+ self.proj = torch.nn.Linear(
52
+ clip_embeddings_dim, self.clip_extra_context_tokens * cross_attention_dim
53
+ )
54
+ self.norm = torch.nn.LayerNorm(cross_attention_dim)
55
+
56
+ def forward(self, image_embeds):
57
+ """
58
+ Forward pass of the ImageProjModel, which takes in image embeddings and returns the
59
+ projected tokens after reshaping and normalization.
60
+
61
+ Args:
62
+ image_embeds (torch.Tensor): The input image embeddings, with shape
63
+ batch_size x num_image_tokens x clip_embeddings_dim.
64
+
65
+ Returns:
66
+ clip_extra_context_tokens (torch.Tensor): The projected tokens after reshaping
67
+ and normalization, with shape batch_size x (clip_extra_context_tokens *
68
+ cross_attention_dim).
69
+
70
+ """
71
+ embeds = image_embeds
72
+ clip_extra_context_tokens = self.proj(embeds).reshape(
73
+ -1, self.clip_extra_context_tokens, self.cross_attention_dim
74
+ )
75
+ clip_extra_context_tokens = self.norm(clip_extra_context_tokens)
76
+ return clip_extra_context_tokens
hallo/models/motion_module.py ADDED
@@ -0,0 +1,608 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # pylint: disable=R0801
2
+ # pylint: disable=W0613
3
+ # pylint: disable=W0221
4
+
5
+ """
6
+ temporal_transformers.py
7
+
8
+ This module provides classes and functions for implementing Temporal Transformers
9
+ in PyTorch, designed for handling video data and temporal sequences within transformer-based models.
10
+
11
+ Functions:
12
+ zero_module(module)
13
+ Zero out the parameters of a module and return it.
14
+
15
+ Classes:
16
+ TemporalTransformer3DModelOutput(BaseOutput)
17
+ Dataclass for storing the output of TemporalTransformer3DModel.
18
+
19
+ VanillaTemporalModule(nn.Module)
20
+ A Vanilla Temporal Module class for handling temporal data.
21
+
22
+ TemporalTransformer3DModel(nn.Module)
23
+ A Temporal Transformer 3D Model class for transforming temporal data.
24
+
25
+ TemporalTransformerBlock(nn.Module)
26
+ A Temporal Transformer Block class for building the transformer architecture.
27
+
28
+ PositionalEncoding(nn.Module)
29
+ A Positional Encoding module for transformers to encode positional information.
30
+
31
+ Dependencies:
32
+ math
33
+ dataclasses.dataclass
34
+ typing (Callable, Optional)
35
+ torch
36
+ diffusers (FeedForward, Attention, AttnProcessor)
37
+ diffusers.utils (BaseOutput)
38
+ diffusers.utils.import_utils (is_xformers_available)
39
+ einops (rearrange, repeat)
40
+ torch.nn
41
+ xformers
42
+ xformers.ops
43
+
44
+ Example Usage:
45
+ >>> motion_module = get_motion_module(in_channels=512, motion_module_type="Vanilla", motion_module_kwargs={})
46
+ >>> output = motion_module(input_tensor, temb, encoder_hidden_states)
47
+
48
+ This module is designed to facilitate the creation, training, and inference of transformer models
49
+ that operate on temporal data, such as videos or time-series. It includes mechanisms for applying temporal attention,
50
+ managing positional encoding, and integrating with external libraries for efficient attention operations.
51
+ """
52
+
53
+ # This code is copied from https://github.com/guoyww/AnimateDiff.
54
+
55
+ import math
56
+
57
+ import torch
58
+ import xformers
59
+ import xformers.ops
60
+ from diffusers.models.attention import FeedForward
61
+ from diffusers.models.attention_processor import Attention, AttnProcessor
62
+ from diffusers.utils import BaseOutput
63
+ from diffusers.utils.import_utils import is_xformers_available
64
+ from einops import rearrange, repeat
65
+ from torch import nn
66
+
67
+
68
+ def zero_module(module):
69
+ """
70
+ Zero out the parameters of a module and return it.
71
+
72
+ Args:
73
+ - module: A PyTorch module to zero out its parameters.
74
+
75
+ Returns:
76
+ A zeroed out PyTorch module.
77
+ """
78
+ for p in module.parameters():
79
+ p.detach().zero_()
80
+ return module
81
+
82
+
83
+ class TemporalTransformer3DModelOutput(BaseOutput):
84
+ """
85
+ Output class for the TemporalTransformer3DModel.
86
+
87
+ Attributes:
88
+ sample (torch.FloatTensor): The output sample tensor from the model.
89
+ """
90
+ sample: torch.FloatTensor
91
+
92
+ def get_sample_shape(self):
93
+ """
94
+ Returns the shape of the sample tensor.
95
+
96
+ Returns:
97
+ Tuple: The shape of the sample tensor.
98
+ """
99
+ return self.sample.shape
100
+
101
+
102
+ def get_motion_module(in_channels, motion_module_type: str, motion_module_kwargs: dict):
103
+ """
104
+ This function returns a motion module based on the given type and parameters.
105
+
106
+ Args:
107
+ - in_channels (int): The number of input channels for the motion module.
108
+ - motion_module_type (str): The type of motion module to create. Currently, only "Vanilla" is supported.
109
+ - motion_module_kwargs (dict): Additional keyword arguments to pass to the motion module constructor.
110
+
111
+ Returns:
112
+ VanillaTemporalModule: The created motion module.
113
+
114
+ Raises:
115
+ ValueError: If an unsupported motion_module_type is provided.
116
+ """
117
+ if motion_module_type == "Vanilla":
118
+ return VanillaTemporalModule(
119
+ in_channels=in_channels,
120
+ **motion_module_kwargs,
121
+ )
122
+
123
+ raise ValueError
124
+
125
+
126
+ class VanillaTemporalModule(nn.Module):
127
+ """
128
+ A Vanilla Temporal Module class.
129
+
130
+ Args:
131
+ - in_channels (int): The number of input channels for the motion module.
132
+ - num_attention_heads (int): Number of attention heads.
133
+ - num_transformer_block (int): Number of transformer blocks.
134
+ - attention_block_types (tuple): Types of attention blocks.
135
+ - cross_frame_attention_mode: Mode for cross-frame attention.
136
+ - temporal_position_encoding (bool): Flag for temporal position encoding.
137
+ - temporal_position_encoding_max_len (int): Maximum length for temporal position encoding.
138
+ - temporal_attention_dim_div (int): Divisor for temporal attention dimension.
139
+ - zero_initialize (bool): Flag for zero initialization.
140
+ """
141
+
142
+ def __init__(
143
+ self,
144
+ in_channels,
145
+ num_attention_heads=8,
146
+ num_transformer_block=2,
147
+ attention_block_types=("Temporal_Self", "Temporal_Self"),
148
+ cross_frame_attention_mode=None,
149
+ temporal_position_encoding=False,
150
+ temporal_position_encoding_max_len=24,
151
+ temporal_attention_dim_div=1,
152
+ zero_initialize=True,
153
+ ):
154
+ super().__init__()
155
+
156
+ self.temporal_transformer = TemporalTransformer3DModel(
157
+ in_channels=in_channels,
158
+ num_attention_heads=num_attention_heads,
159
+ attention_head_dim=in_channels
160
+ // num_attention_heads
161
+ // temporal_attention_dim_div,
162
+ num_layers=num_transformer_block,
163
+ attention_block_types=attention_block_types,
164
+ cross_frame_attention_mode=cross_frame_attention_mode,
165
+ temporal_position_encoding=temporal_position_encoding,
166
+ temporal_position_encoding_max_len=temporal_position_encoding_max_len,
167
+ )
168
+
169
+ if zero_initialize:
170
+ self.temporal_transformer.proj_out = zero_module(
171
+ self.temporal_transformer.proj_out
172
+ )
173
+
174
+ def forward(
175
+ self,
176
+ input_tensor,
177
+ encoder_hidden_states,
178
+ attention_mask=None,
179
+ ):
180
+ """
181
+ Forward pass of the TemporalTransformer3DModel.
182
+
183
+ Args:
184
+ hidden_states (torch.Tensor): The hidden states of the model.
185
+ encoder_hidden_states (torch.Tensor, optional): The hidden states of the encoder.
186
+ attention_mask (torch.Tensor, optional): The attention mask.
187
+
188
+ Returns:
189
+ torch.Tensor: The output tensor after the forward pass.
190
+ """
191
+ hidden_states = input_tensor
192
+ hidden_states = self.temporal_transformer(
193
+ hidden_states, encoder_hidden_states
194
+ )
195
+
196
+ output = hidden_states
197
+ return output
198
+
199
+
200
+ class TemporalTransformer3DModel(nn.Module):
201
+ """
202
+ A Temporal Transformer 3D Model class.
203
+
204
+ Args:
205
+ - in_channels (int): The number of input channels.
206
+ - num_attention_heads (int): Number of attention heads.
207
+ - attention_head_dim (int): Dimension of attention heads.
208
+ - num_layers (int): Number of transformer layers.
209
+ - attention_block_types (tuple): Types of attention blocks.
210
+ - dropout (float): Dropout rate.
211
+ - norm_num_groups (int): Number of groups for normalization.
212
+ - cross_attention_dim (int): Dimension for cross-attention.
213
+ - activation_fn (str): Activation function.
214
+ - attention_bias (bool): Flag for attention bias.
215
+ - upcast_attention (bool): Flag for upcast attention.
216
+ - cross_frame_attention_mode: Mode for cross-frame attention.
217
+ - temporal_position_encoding (bool): Flag for temporal position encoding.
218
+ - temporal_position_encoding_max_len (int): Maximum length for temporal position encoding.
219
+ """
220
+ def __init__(
221
+ self,
222
+ in_channels,
223
+ num_attention_heads,
224
+ attention_head_dim,
225
+ num_layers,
226
+ attention_block_types=(
227
+ "Temporal_Self",
228
+ "Temporal_Self",
229
+ ),
230
+ dropout=0.0,
231
+ norm_num_groups=32,
232
+ cross_attention_dim=768,
233
+ activation_fn="geglu",
234
+ attention_bias=False,
235
+ upcast_attention=False,
236
+ cross_frame_attention_mode=None,
237
+ temporal_position_encoding=False,
238
+ temporal_position_encoding_max_len=24,
239
+ ):
240
+ super().__init__()
241
+
242
+ inner_dim = num_attention_heads * attention_head_dim
243
+
244
+ self.norm = torch.nn.GroupNorm(
245
+ num_groups=norm_num_groups, num_channels=in_channels, eps=1e-6, affine=True
246
+ )
247
+ self.proj_in = nn.Linear(in_channels, inner_dim)
248
+
249
+ self.transformer_blocks = nn.ModuleList(
250
+ [
251
+ TemporalTransformerBlock(
252
+ dim=inner_dim,
253
+ num_attention_heads=num_attention_heads,
254
+ attention_head_dim=attention_head_dim,
255
+ attention_block_types=attention_block_types,
256
+ dropout=dropout,
257
+ cross_attention_dim=cross_attention_dim,
258
+ activation_fn=activation_fn,
259
+ attention_bias=attention_bias,
260
+ upcast_attention=upcast_attention,
261
+ cross_frame_attention_mode=cross_frame_attention_mode,
262
+ temporal_position_encoding=temporal_position_encoding,
263
+ temporal_position_encoding_max_len=temporal_position_encoding_max_len,
264
+ )
265
+ for d in range(num_layers)
266
+ ]
267
+ )
268
+ self.proj_out = nn.Linear(inner_dim, in_channels)
269
+
270
+ def forward(self, hidden_states, encoder_hidden_states=None):
271
+ """
272
+ Forward pass for the TemporalTransformer3DModel.
273
+
274
+ Args:
275
+ hidden_states (torch.Tensor): The input hidden states with shape (batch_size, sequence_length, in_channels).
276
+ encoder_hidden_states (torch.Tensor, optional): The encoder hidden states with shape (batch_size, encoder_sequence_length, in_channels).
277
+
278
+ Returns:
279
+ torch.Tensor: The output hidden states with shape (batch_size, sequence_length, in_channels).
280
+ """
281
+ assert (
282
+ hidden_states.dim() == 5
283
+ ), f"Expected hidden_states to have ndim=5, but got ndim={hidden_states.dim()}."
284
+ video_length = hidden_states.shape[2]
285
+ hidden_states = rearrange(hidden_states, "b c f h w -> (b f) c h w")
286
+
287
+ batch, _, height, weight = hidden_states.shape
288
+ residual = hidden_states
289
+
290
+ hidden_states = self.norm(hidden_states)
291
+ inner_dim = hidden_states.shape[1]
292
+ hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(
293
+ batch, height * weight, inner_dim
294
+ )
295
+ hidden_states = self.proj_in(hidden_states)
296
+
297
+ # Transformer Blocks
298
+ for block in self.transformer_blocks:
299
+ hidden_states = block(
300
+ hidden_states,
301
+ encoder_hidden_states=encoder_hidden_states,
302
+ video_length=video_length,
303
+ )
304
+
305
+ # output
306
+ hidden_states = self.proj_out(hidden_states)
307
+ hidden_states = (
308
+ hidden_states.reshape(batch, height, weight, inner_dim)
309
+ .permute(0, 3, 1, 2)
310
+ .contiguous()
311
+ )
312
+
313
+ output = hidden_states + residual
314
+ output = rearrange(output, "(b f) c h w -> b c f h w", f=video_length)
315
+
316
+ return output
317
+
318
+
319
+ class TemporalTransformerBlock(nn.Module):
320
+ """
321
+ A Temporal Transformer Block class.
322
+
323
+ Args:
324
+ - dim (int): Dimension of the block.
325
+ - num_attention_heads (int): Number of attention heads.
326
+ - attention_head_dim (int): Dimension of attention heads.
327
+ - attention_block_types (tuple): Types of attention blocks.
328
+ - dropout (float): Dropout rate.
329
+ - cross_attention_dim (int): Dimension for cross-attention.
330
+ - activation_fn (str): Activation function.
331
+ - attention_bias (bool): Flag for attention bias.
332
+ - upcast_attention (bool): Flag for upcast attention.
333
+ - cross_frame_attention_mode: Mode for cross-frame attention.
334
+ - temporal_position_encoding (bool): Flag for temporal position encoding.
335
+ - temporal_position_encoding_max_len (int): Maximum length for temporal position encoding.
336
+ """
337
+ def __init__(
338
+ self,
339
+ dim,
340
+ num_attention_heads,
341
+ attention_head_dim,
342
+ attention_block_types=(
343
+ "Temporal_Self",
344
+ "Temporal_Self",
345
+ ),
346
+ dropout=0.0,
347
+ cross_attention_dim=768,
348
+ activation_fn="geglu",
349
+ attention_bias=False,
350
+ upcast_attention=False,
351
+ cross_frame_attention_mode=None,
352
+ temporal_position_encoding=False,
353
+ temporal_position_encoding_max_len=24,
354
+ ):
355
+ super().__init__()
356
+
357
+ attention_blocks = []
358
+ norms = []
359
+
360
+ for block_name in attention_block_types:
361
+ attention_blocks.append(
362
+ VersatileAttention(
363
+ attention_mode=block_name.split("_", maxsplit=1)[0],
364
+ cross_attention_dim=cross_attention_dim
365
+ if block_name.endswith("_Cross")
366
+ else None,
367
+ query_dim=dim,
368
+ heads=num_attention_heads,
369
+ dim_head=attention_head_dim,
370
+ dropout=dropout,
371
+ bias=attention_bias,
372
+ upcast_attention=upcast_attention,
373
+ cross_frame_attention_mode=cross_frame_attention_mode,
374
+ temporal_position_encoding=temporal_position_encoding,
375
+ temporal_position_encoding_max_len=temporal_position_encoding_max_len,
376
+ )
377
+ )
378
+ norms.append(nn.LayerNorm(dim))
379
+
380
+ self.attention_blocks = nn.ModuleList(attention_blocks)
381
+ self.norms = nn.ModuleList(norms)
382
+
383
+ self.ff = FeedForward(dim, dropout=dropout,
384
+ activation_fn=activation_fn)
385
+ self.ff_norm = nn.LayerNorm(dim)
386
+
387
+ def forward(
388
+ self,
389
+ hidden_states,
390
+ encoder_hidden_states=None,
391
+ video_length=None,
392
+ ):
393
+ """
394
+ Forward pass for the TemporalTransformerBlock.
395
+
396
+ Args:
397
+ hidden_states (torch.Tensor): The input hidden states with shape
398
+ (batch_size, video_length, in_channels).
399
+ encoder_hidden_states (torch.Tensor, optional): The encoder hidden states
400
+ with shape (batch_size, encoder_length, in_channels).
401
+ video_length (int, optional): The length of the video.
402
+
403
+ Returns:
404
+ torch.Tensor: The output hidden states with shape
405
+ (batch_size, video_length, in_channels).
406
+ """
407
+ for attention_block, norm in zip(self.attention_blocks, self.norms):
408
+ norm_hidden_states = norm(hidden_states)
409
+ hidden_states = (
410
+ attention_block(
411
+ norm_hidden_states,
412
+ encoder_hidden_states=encoder_hidden_states
413
+ if attention_block.is_cross_attention
414
+ else None,
415
+ video_length=video_length,
416
+ )
417
+ + hidden_states
418
+ )
419
+
420
+ hidden_states = self.ff(self.ff_norm(hidden_states)) + hidden_states
421
+
422
+ output = hidden_states
423
+ return output
424
+
425
+
426
+ class PositionalEncoding(nn.Module):
427
+ """
428
+ Positional Encoding module for transformers.
429
+
430
+ Args:
431
+ - d_model (int): Model dimension.
432
+ - dropout (float): Dropout rate.
433
+ - max_len (int): Maximum length for positional encoding.
434
+ """
435
+ def __init__(self, d_model, dropout=0.0, max_len=24):
436
+ super().__init__()
437
+ self.dropout = nn.Dropout(p=dropout)
438
+ position = torch.arange(max_len).unsqueeze(1)
439
+ div_term = torch.exp(
440
+ torch.arange(0, d_model, 2) * (-math.log(10000.0) / d_model)
441
+ )
442
+ pe = torch.zeros(1, max_len, d_model)
443
+ pe[0, :, 0::2] = torch.sin(position * div_term)
444
+ pe[0, :, 1::2] = torch.cos(position * div_term)
445
+ self.register_buffer("pe", pe)
446
+
447
+ def forward(self, x):
448
+ """
449
+ Forward pass of the PositionalEncoding module.
450
+
451
+ This method takes an input tensor `x` and adds the positional encoding to it. The positional encoding is
452
+ generated based on the input tensor's shape and is added to the input tensor element-wise.
453
+
454
+ Args:
455
+ x (torch.Tensor): The input tensor to be positionally encoded.
456
+
457
+ Returns:
458
+ torch.Tensor: The positionally encoded tensor.
459
+ """
460
+ x = x + self.pe[:, : x.size(1)]
461
+ return self.dropout(x)
462
+
463
+
464
+ class VersatileAttention(Attention):
465
+ """
466
+ Versatile Attention class.
467
+
468
+ Args:
469
+ - attention_mode: Attention mode.
470
+ - temporal_position_encoding (bool): Flag for temporal position encoding.
471
+ - temporal_position_encoding_max_len (int): Maximum length for temporal position encoding.
472
+ """
473
+ def __init__(
474
+ self,
475
+ *args,
476
+ attention_mode=None,
477
+ cross_frame_attention_mode=None,
478
+ temporal_position_encoding=False,
479
+ temporal_position_encoding_max_len=24,
480
+ **kwargs,
481
+ ):
482
+ super().__init__(*args, **kwargs)
483
+ assert attention_mode == "Temporal"
484
+
485
+ self.attention_mode = attention_mode
486
+ self.is_cross_attention = kwargs.get("cross_attention_dim") is not None
487
+
488
+ self.pos_encoder = (
489
+ PositionalEncoding(
490
+ kwargs["query_dim"],
491
+ dropout=0.0,
492
+ max_len=temporal_position_encoding_max_len,
493
+ )
494
+ if (temporal_position_encoding and attention_mode == "Temporal")
495
+ else None
496
+ )
497
+
498
+ def extra_repr(self):
499
+ """
500
+ Returns a string representation of the module with information about the attention mode and whether it is cross-attention.
501
+
502
+ Returns:
503
+ str: A string representation of the module.
504
+ """
505
+ return f"(Module Info) Attention_Mode: {self.attention_mode}, Is_Cross_Attention: {self.is_cross_attention}"
506
+
507
+ def set_use_memory_efficient_attention_xformers(
508
+ self,
509
+ use_memory_efficient_attention_xformers: bool,
510
+ ):
511
+ """
512
+ Sets the use of memory-efficient attention xformers for the VersatileAttention class.
513
+
514
+ Args:
515
+ use_memory_efficient_attention_xformers (bool): A boolean flag indicating whether to use memory-efficient attention xformers or not.
516
+
517
+ Returns:
518
+ None
519
+
520
+ """
521
+ if use_memory_efficient_attention_xformers:
522
+ if not is_xformers_available():
523
+ raise ModuleNotFoundError(
524
+ (
525
+ "Refer to https://github.com/facebookresearch/xformers for more information on how to install"
526
+ " xformers"
527
+ ),
528
+ name="xformers",
529
+ )
530
+
531
+ if not torch.cuda.is_available():
532
+ raise ValueError(
533
+ "torch.cuda.is_available() should be True but is False. xformers' memory efficient attention is"
534
+ " only available for GPU "
535
+ )
536
+
537
+ try:
538
+ # Make sure we can run the memory efficient attention
539
+ _ = xformers.ops.memory_efficient_attention(
540
+ torch.randn((1, 2, 40), device="cuda"),
541
+ torch.randn((1, 2, 40), device="cuda"),
542
+ torch.randn((1, 2, 40), device="cuda"),
543
+ )
544
+ except Exception as e:
545
+ raise e
546
+ processor = AttnProcessor()
547
+ else:
548
+ processor = AttnProcessor()
549
+
550
+ self.set_processor(processor)
551
+
552
+ def forward(
553
+ self,
554
+ hidden_states,
555
+ encoder_hidden_states=None,
556
+ attention_mask=None,
557
+ video_length=None,
558
+ **cross_attention_kwargs,
559
+ ):
560
+ """
561
+ Args:
562
+ hidden_states (`torch.Tensor`):
563
+ The hidden states to be passed through the model.
564
+ encoder_hidden_states (`torch.Tensor`, optional):
565
+ The encoder hidden states to be passed through the model.
566
+ attention_mask (`torch.Tensor`, optional):
567
+ The attention mask to be used in the model.
568
+ video_length (`int`, optional):
569
+ The length of the video.
570
+ cross_attention_kwargs (`dict`, optional):
571
+ Additional keyword arguments to be used for cross-attention.
572
+
573
+ Returns:
574
+ `torch.Tensor`:
575
+ The output tensor after passing through the model.
576
+
577
+ """
578
+ if self.attention_mode == "Temporal":
579
+ d = hidden_states.shape[1] # d means HxW
580
+ hidden_states = rearrange(
581
+ hidden_states, "(b f) d c -> (b d) f c", f=video_length
582
+ )
583
+
584
+ if self.pos_encoder is not None:
585
+ hidden_states = self.pos_encoder(hidden_states)
586
+
587
+ encoder_hidden_states = (
588
+ repeat(encoder_hidden_states, "b n c -> (b d) n c", d=d)
589
+ if encoder_hidden_states is not None
590
+ else encoder_hidden_states
591
+ )
592
+
593
+ else:
594
+ raise NotImplementedError
595
+
596
+ hidden_states = self.processor(
597
+ self,
598
+ hidden_states,
599
+ encoder_hidden_states=encoder_hidden_states,
600
+ attention_mask=attention_mask,
601
+ **cross_attention_kwargs,
602
+ )
603
+
604
+ if self.attention_mode == "Temporal":
605
+ hidden_states = rearrange(
606
+ hidden_states, "(b d) f c -> (b f) d c", d=d)
607
+
608
+ return hidden_states
hallo/models/mutual_self_attention.py ADDED
@@ -0,0 +1,496 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # pylint: disable=E1120
2
+ """
3
+ This module contains the implementation of mutual self-attention,
4
+ which is a type of attention mechanism used in deep learning models.
5
+ The module includes several classes and functions related to attention mechanisms,
6
+ such as BasicTransformerBlock and TemporalBasicTransformerBlock.
7
+ The main purpose of this module is to provide a comprehensive attention mechanism for various tasks in deep learning,
8
+ such as image and video processing, natural language processing, and so on.
9
+ """
10
+
11
+ from typing import Any, Dict, Optional
12
+
13
+ import torch
14
+ from einops import rearrange
15
+
16
+ from .attention import BasicTransformerBlock, TemporalBasicTransformerBlock
17
+
18
+
19
+ def torch_dfs(model: torch.nn.Module):
20
+ """
21
+ Perform a depth-first search (DFS) traversal on a PyTorch model's neural network architecture.
22
+
23
+ This function recursively traverses all the children modules of a given PyTorch model and returns a list
24
+ containing all the modules in the model's architecture. The DFS approach starts with the input model and
25
+ explores its children modules depth-wise before backtracking and exploring other branches.
26
+
27
+ Args:
28
+ model (torch.nn.Module): The root module of the neural network to traverse.
29
+
30
+ Returns:
31
+ list: A list of all the modules in the model's architecture.
32
+ """
33
+ result = [model]
34
+ for child in model.children():
35
+ result += torch_dfs(child)
36
+ return result
37
+
38
+
39
+ class ReferenceAttentionControl:
40
+ """
41
+ This class is used to control the reference attention mechanism in a neural network model.
42
+ It is responsible for managing the guidance and fusion blocks, and modifying the self-attention
43
+ and group normalization mechanisms. The class also provides methods for registering reference hooks
44
+ and updating/clearing the internal state of the attention control object.
45
+
46
+ Attributes:
47
+ unet: The UNet model associated with this attention control object.
48
+ mode: The operating mode of the attention control object, either 'write' or 'read'.
49
+ do_classifier_free_guidance: Whether to use classifier-free guidance in the attention mechanism.
50
+ attention_auto_machine_weight: The weight assigned to the attention auto-machine.
51
+ gn_auto_machine_weight: The weight assigned to the group normalization auto-machine.
52
+ style_fidelity: The style fidelity parameter for the attention mechanism.
53
+ reference_attn: Whether to use reference attention in the model.
54
+ reference_adain: Whether to use reference AdaIN in the model.
55
+ fusion_blocks: The type of fusion blocks to use in the model ('midup', 'late', or 'nofusion').
56
+ batch_size: The batch size used for processing video frames.
57
+
58
+ Methods:
59
+ register_reference_hooks: Registers the reference hooks for the attention control object.
60
+ hacked_basic_transformer_inner_forward: The modified inner forward method for the basic transformer block.
61
+ update: Updates the internal state of the attention control object using the provided writer and dtype.
62
+ clear: Clears the internal state of the attention control object.
63
+ """
64
+ def __init__(
65
+ self,
66
+ unet,
67
+ mode="write",
68
+ do_classifier_free_guidance=False,
69
+ attention_auto_machine_weight=float("inf"),
70
+ gn_auto_machine_weight=1.0,
71
+ style_fidelity=1.0,
72
+ reference_attn=True,
73
+ reference_adain=False,
74
+ fusion_blocks="midup",
75
+ batch_size=1,
76
+ ) -> None:
77
+ """
78
+ Initializes the ReferenceAttentionControl class.
79
+
80
+ Args:
81
+ unet (torch.nn.Module): The UNet model.
82
+ mode (str, optional): The mode of operation. Defaults to "write".
83
+ do_classifier_free_guidance (bool, optional): Whether to do classifier-free guidance. Defaults to False.
84
+ attention_auto_machine_weight (float, optional): The weight for attention auto-machine. Defaults to infinity.
85
+ gn_auto_machine_weight (float, optional): The weight for group-norm auto-machine. Defaults to 1.0.
86
+ style_fidelity (float, optional): The style fidelity. Defaults to 1.0.
87
+ reference_attn (bool, optional): Whether to use reference attention. Defaults to True.
88
+ reference_adain (bool, optional): Whether to use reference AdaIN. Defaults to False.
89
+ fusion_blocks (str, optional): The fusion blocks to use. Defaults to "midup".
90
+ batch_size (int, optional): The batch size. Defaults to 1.
91
+
92
+ Raises:
93
+ ValueError: If the mode is not recognized.
94
+ ValueError: If the fusion blocks are not recognized.
95
+ """
96
+ # 10. Modify self attention and group norm
97
+ self.unet = unet
98
+ assert mode in ["read", "write"]
99
+ assert fusion_blocks in ["midup", "full"]
100
+ self.reference_attn = reference_attn
101
+ self.reference_adain = reference_adain
102
+ self.fusion_blocks = fusion_blocks
103
+ self.register_reference_hooks(
104
+ mode,
105
+ do_classifier_free_guidance,
106
+ attention_auto_machine_weight,
107
+ gn_auto_machine_weight,
108
+ style_fidelity,
109
+ reference_attn,
110
+ reference_adain,
111
+ fusion_blocks,
112
+ batch_size=batch_size,
113
+ )
114
+
115
+ def register_reference_hooks(
116
+ self,
117
+ mode,
118
+ do_classifier_free_guidance,
119
+ _attention_auto_machine_weight,
120
+ _gn_auto_machine_weight,
121
+ _style_fidelity,
122
+ _reference_attn,
123
+ _reference_adain,
124
+ _dtype=torch.float16,
125
+ batch_size=1,
126
+ num_images_per_prompt=1,
127
+ device=torch.device("cpu"),
128
+ _fusion_blocks="midup",
129
+ ):
130
+ """
131
+ Registers reference hooks for the model.
132
+
133
+ This function is responsible for registering reference hooks in the model,
134
+ which are used to modify the attention mechanism and group normalization layers.
135
+ It takes various parameters as input, such as mode,
136
+ do_classifier_free_guidance, _attention_auto_machine_weight, _gn_auto_machine_weight, _style_fidelity,
137
+ _reference_attn, _reference_adain, _dtype, batch_size, num_images_per_prompt, device, and _fusion_blocks.
138
+
139
+ Args:
140
+ self: Reference to the instance of the class.
141
+ mode: The mode of operation for the reference hooks.
142
+ do_classifier_free_guidance: A boolean flag indicating whether to use classifier-free guidance.
143
+ _attention_auto_machine_weight: The weight for the attention auto-machine.
144
+ _gn_auto_machine_weight: The weight for the group normalization auto-machine.
145
+ _style_fidelity: The style fidelity for the reference hooks.
146
+ _reference_attn: A boolean flag indicating whether to use reference attention.
147
+ _reference_adain: A boolean flag indicating whether to use reference AdaIN.
148
+ _dtype: The data type for the reference hooks.
149
+ batch_size: The batch size for the reference hooks.
150
+ num_images_per_prompt: The number of images per prompt for the reference hooks.
151
+ device: The device for the reference hooks.
152
+ _fusion_blocks: The fusion blocks for the reference hooks.
153
+
154
+ Returns:
155
+ None
156
+ """
157
+ MODE = mode
158
+ if do_classifier_free_guidance:
159
+ uc_mask = (
160
+ torch.Tensor(
161
+ [1] * batch_size * num_images_per_prompt * 16
162
+ + [0] * batch_size * num_images_per_prompt * 16
163
+ )
164
+ .to(device)
165
+ .bool()
166
+ )
167
+ else:
168
+ uc_mask = (
169
+ torch.Tensor([0] * batch_size * num_images_per_prompt * 2)
170
+ .to(device)
171
+ .bool()
172
+ )
173
+
174
+ def hacked_basic_transformer_inner_forward(
175
+ self,
176
+ hidden_states: torch.FloatTensor,
177
+ attention_mask: Optional[torch.FloatTensor] = None,
178
+ encoder_hidden_states: Optional[torch.FloatTensor] = None,
179
+ encoder_attention_mask: Optional[torch.FloatTensor] = None,
180
+ timestep: Optional[torch.LongTensor] = None,
181
+ cross_attention_kwargs: Dict[str, Any] = None,
182
+ class_labels: Optional[torch.LongTensor] = None,
183
+ video_length=None,
184
+ ):
185
+ gate_msa = None
186
+ shift_mlp = None
187
+ scale_mlp = None
188
+ gate_mlp = None
189
+
190
+ if self.use_ada_layer_norm: # False
191
+ norm_hidden_states = self.norm1(hidden_states, timestep)
192
+ elif self.use_ada_layer_norm_zero:
193
+ (
194
+ norm_hidden_states,
195
+ gate_msa,
196
+ shift_mlp,
197
+ scale_mlp,
198
+ gate_mlp,
199
+ ) = self.norm1(
200
+ hidden_states,
201
+ timestep,
202
+ class_labels,
203
+ hidden_dtype=hidden_states.dtype,
204
+ )
205
+ else:
206
+ norm_hidden_states = self.norm1(hidden_states)
207
+
208
+ # 1. Self-Attention
209
+ # self.only_cross_attention = False
210
+ cross_attention_kwargs = (
211
+ cross_attention_kwargs if cross_attention_kwargs is not None else {}
212
+ )
213
+ if self.only_cross_attention:
214
+ attn_output = self.attn1(
215
+ norm_hidden_states,
216
+ encoder_hidden_states=(
217
+ encoder_hidden_states if self.only_cross_attention else None
218
+ ),
219
+ attention_mask=attention_mask,
220
+ **cross_attention_kwargs,
221
+ )
222
+ else:
223
+ if MODE == "write":
224
+ self.bank.append(norm_hidden_states.clone())
225
+ attn_output = self.attn1(
226
+ norm_hidden_states,
227
+ encoder_hidden_states=(
228
+ encoder_hidden_states if self.only_cross_attention else None
229
+ ),
230
+ attention_mask=attention_mask,
231
+ **cross_attention_kwargs,
232
+ )
233
+ if MODE == "read":
234
+
235
+ bank_fea = [
236
+ rearrange(
237
+ rearrange(
238
+ d,
239
+ "(b s) l c -> b s l c",
240
+ b=norm_hidden_states.shape[0] // video_length,
241
+ )[:, 0, :, :]
242
+ # .unsqueeze(1)
243
+ .repeat(1, video_length, 1, 1),
244
+ "b t l c -> (b t) l c",
245
+ )
246
+ for d in self.bank
247
+ ]
248
+ motion_frames_fea = [rearrange(
249
+ d,
250
+ "(b s) l c -> b s l c",
251
+ b=norm_hidden_states.shape[0] // video_length,
252
+ )[:, 1:, :, :] for d in self.bank]
253
+ modify_norm_hidden_states = torch.cat(
254
+ [norm_hidden_states] + bank_fea, dim=1
255
+ )
256
+ hidden_states_uc = (
257
+ self.attn1(
258
+ norm_hidden_states,
259
+ encoder_hidden_states=modify_norm_hidden_states,
260
+ attention_mask=attention_mask,
261
+ )
262
+ + hidden_states
263
+ )
264
+ if do_classifier_free_guidance:
265
+ hidden_states_c = hidden_states_uc.clone()
266
+ _uc_mask = uc_mask.clone()
267
+ if hidden_states.shape[0] != _uc_mask.shape[0]:
268
+ _uc_mask = (
269
+ torch.Tensor(
270
+ [1] * (hidden_states.shape[0] // 2)
271
+ + [0] * (hidden_states.shape[0] // 2)
272
+ )
273
+ .to(device)
274
+ .bool()
275
+ )
276
+ hidden_states_c[_uc_mask] = (
277
+ self.attn1(
278
+ norm_hidden_states[_uc_mask],
279
+ encoder_hidden_states=norm_hidden_states[_uc_mask],
280
+ attention_mask=attention_mask,
281
+ )
282
+ + hidden_states[_uc_mask]
283
+ )
284
+ hidden_states = hidden_states_c.clone()
285
+ else:
286
+ hidden_states = hidden_states_uc
287
+
288
+ # self.bank.clear()
289
+ if self.attn2 is not None:
290
+ # Cross-Attention
291
+ norm_hidden_states = (
292
+ self.norm2(hidden_states, timestep)
293
+ if self.use_ada_layer_norm
294
+ else self.norm2(hidden_states)
295
+ )
296
+ hidden_states = (
297
+ self.attn2(
298
+ norm_hidden_states,
299
+ encoder_hidden_states=encoder_hidden_states,
300
+ attention_mask=attention_mask,
301
+ )
302
+ + hidden_states
303
+ )
304
+
305
+ # Feed-forward
306
+ hidden_states = self.ff(self.norm3(
307
+ hidden_states)) + hidden_states
308
+
309
+ # Temporal-Attention
310
+ if self.unet_use_temporal_attention:
311
+ d = hidden_states.shape[1]
312
+ hidden_states = rearrange(
313
+ hidden_states, "(b f) d c -> (b d) f c", f=video_length
314
+ )
315
+ norm_hidden_states = (
316
+ self.norm_temp(hidden_states, timestep)
317
+ if self.use_ada_layer_norm
318
+ else self.norm_temp(hidden_states)
319
+ )
320
+ hidden_states = (
321
+ self.attn_temp(norm_hidden_states) + hidden_states
322
+ )
323
+ hidden_states = rearrange(
324
+ hidden_states, "(b d) f c -> (b f) d c", d=d
325
+ )
326
+
327
+ return hidden_states, motion_frames_fea
328
+
329
+ if self.use_ada_layer_norm_zero:
330
+ attn_output = gate_msa.unsqueeze(1) * attn_output
331
+ hidden_states = attn_output + hidden_states
332
+
333
+ if self.attn2 is not None:
334
+ norm_hidden_states = (
335
+ self.norm2(hidden_states, timestep)
336
+ if self.use_ada_layer_norm
337
+ else self.norm2(hidden_states)
338
+ )
339
+
340
+ # 2. Cross-Attention
341
+ tmp = norm_hidden_states.shape[0] // encoder_hidden_states.shape[0]
342
+ attn_output = self.attn2(
343
+ norm_hidden_states,
344
+ # TODO: repeat这个地方需要斟酌一下
345
+ encoder_hidden_states=encoder_hidden_states.repeat(
346
+ tmp, 1, 1),
347
+ attention_mask=encoder_attention_mask,
348
+ **cross_attention_kwargs,
349
+ )
350
+ hidden_states = attn_output + hidden_states
351
+
352
+ # 3. Feed-forward
353
+ norm_hidden_states = self.norm3(hidden_states)
354
+
355
+ if self.use_ada_layer_norm_zero:
356
+ norm_hidden_states = (
357
+ norm_hidden_states *
358
+ (1 + scale_mlp[:, None]) + shift_mlp[:, None]
359
+ )
360
+
361
+ ff_output = self.ff(norm_hidden_states)
362
+
363
+ if self.use_ada_layer_norm_zero:
364
+ ff_output = gate_mlp.unsqueeze(1) * ff_output
365
+
366
+ hidden_states = ff_output + hidden_states
367
+
368
+ return hidden_states
369
+
370
+ if self.reference_attn:
371
+ if self.fusion_blocks == "midup":
372
+ attn_modules = [
373
+ module
374
+ for module in (
375
+ torch_dfs(self.unet.mid_block) +
376
+ torch_dfs(self.unet.up_blocks)
377
+ )
378
+ if isinstance(module, (BasicTransformerBlock, TemporalBasicTransformerBlock))
379
+ ]
380
+ elif self.fusion_blocks == "full":
381
+ attn_modules = [
382
+ module
383
+ for module in torch_dfs(self.unet)
384
+ if isinstance(module, (BasicTransformerBlock, TemporalBasicTransformerBlock))
385
+ ]
386
+ attn_modules = sorted(
387
+ attn_modules, key=lambda x: -x.norm1.normalized_shape[0]
388
+ )
389
+
390
+ for i, module in enumerate(attn_modules):
391
+ module._original_inner_forward = module.forward
392
+ if isinstance(module, BasicTransformerBlock):
393
+ module.forward = hacked_basic_transformer_inner_forward.__get__(
394
+ module,
395
+ BasicTransformerBlock)
396
+ if isinstance(module, TemporalBasicTransformerBlock):
397
+ module.forward = hacked_basic_transformer_inner_forward.__get__(
398
+ module,
399
+ TemporalBasicTransformerBlock)
400
+
401
+ module.bank = []
402
+ module.attn_weight = float(i) / float(len(attn_modules))
403
+
404
+ def update(self, writer, dtype=torch.float16):
405
+ """
406
+ Update the model's parameters.
407
+
408
+ Args:
409
+ writer (torch.nn.Module): The model's writer object.
410
+ dtype (torch.dtype, optional): The data type to be used for the update. Defaults to torch.float16.
411
+
412
+ Returns:
413
+ None.
414
+ """
415
+ if self.reference_attn:
416
+ if self.fusion_blocks == "midup":
417
+ reader_attn_modules = [
418
+ module
419
+ for module in (
420
+ torch_dfs(self.unet.mid_block) +
421
+ torch_dfs(self.unet.up_blocks)
422
+ )
423
+ if isinstance(module, TemporalBasicTransformerBlock)
424
+ ]
425
+ writer_attn_modules = [
426
+ module
427
+ for module in (
428
+ torch_dfs(writer.unet.mid_block)
429
+ + torch_dfs(writer.unet.up_blocks)
430
+ )
431
+ if isinstance(module, BasicTransformerBlock)
432
+ ]
433
+ elif self.fusion_blocks == "full":
434
+ reader_attn_modules = [
435
+ module
436
+ for module in torch_dfs(self.unet)
437
+ if isinstance(module, TemporalBasicTransformerBlock)
438
+ ]
439
+ writer_attn_modules = [
440
+ module
441
+ for module in torch_dfs(writer.unet)
442
+ if isinstance(module, BasicTransformerBlock)
443
+ ]
444
+
445
+ assert len(reader_attn_modules) == len(writer_attn_modules)
446
+ reader_attn_modules = sorted(
447
+ reader_attn_modules, key=lambda x: -x.norm1.normalized_shape[0]
448
+ )
449
+ writer_attn_modules = sorted(
450
+ writer_attn_modules, key=lambda x: -x.norm1.normalized_shape[0]
451
+ )
452
+ for r, w in zip(reader_attn_modules, writer_attn_modules):
453
+ r.bank = [v.clone().to(dtype) for v in w.bank]
454
+
455
+
456
+ def clear(self):
457
+ """
458
+ Clears the attention bank of all reader attention modules.
459
+
460
+ This method is used when the `reference_attn` attribute is set to `True`.
461
+ It clears the attention bank of all reader attention modules inside the UNet
462
+ model based on the selected `fusion_blocks` mode.
463
+
464
+ If `fusion_blocks` is set to "midup", it searches for reader attention modules
465
+ in both the mid block and up blocks of the UNet model. If `fusion_blocks` is set
466
+ to "full", it searches for reader attention modules in the entire UNet model.
467
+
468
+ It sorts the reader attention modules by the number of neurons in their
469
+ `norm1.normalized_shape[0]` attribute in descending order. This sorting ensures
470
+ that the modules with more neurons are cleared first.
471
+
472
+ Finally, it iterates through the sorted list of reader attention modules and
473
+ calls the `clear()` method on each module's `bank` attribute to clear the
474
+ attention bank.
475
+ """
476
+ if self.reference_attn:
477
+ if self.fusion_blocks == "midup":
478
+ reader_attn_modules = [
479
+ module
480
+ for module in (
481
+ torch_dfs(self.unet.mid_block) +
482
+ torch_dfs(self.unet.up_blocks)
483
+ )
484
+ if isinstance(module, (BasicTransformerBlock, TemporalBasicTransformerBlock))
485
+ ]
486
+ elif self.fusion_blocks == "full":
487
+ reader_attn_modules = [
488
+ module
489
+ for module in torch_dfs(self.unet)
490
+ if isinstance(module, (BasicTransformerBlock, TemporalBasicTransformerBlock))
491
+ ]
492
+ reader_attn_modules = sorted(
493
+ reader_attn_modules, key=lambda x: -x.norm1.normalized_shape[0]
494
+ )
495
+ for r in reader_attn_modules:
496
+ r.bank.clear()
hallo/models/resnet.py ADDED
@@ -0,0 +1,435 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # pylint: disable=E1120
2
+ # pylint: disable=E1102
3
+ # pylint: disable=W0237
4
+
5
+ # src/models/resnet.py
6
+
7
+ """
8
+ This module defines various components used in the ResNet model, such as InflatedConv3D, InflatedGroupNorm,
9
+ Upsample3D, Downsample3D, ResnetBlock3D, and Mish activation function. These components are used to construct
10
+ a deep neural network model for image classification or other computer vision tasks.
11
+
12
+ Classes:
13
+ - InflatedConv3d: An inflated 3D convolutional layer, inheriting from nn.Conv2d.
14
+ - InflatedGroupNorm: An inflated group normalization layer, inheriting from nn.GroupNorm.
15
+ - Upsample3D: A 3D upsampling module, used to increase the resolution of the input tensor.
16
+ - Downsample3D: A 3D downsampling module, used to decrease the resolution of the input tensor.
17
+ - ResnetBlock3D: A 3D residual block, commonly used in ResNet architectures.
18
+ - Mish: A Mish activation function, which is a smooth, non-monotonic activation function.
19
+
20
+ To use this module, simply import the classes and functions you need and follow the instructions provided in
21
+ the respective class and function docstrings.
22
+ """
23
+
24
+ import torch
25
+ import torch.nn.functional as F
26
+ from einops import rearrange
27
+ from torch import nn
28
+
29
+
30
+ class InflatedConv3d(nn.Conv2d):
31
+ """
32
+ InflatedConv3d is a class that inherits from torch.nn.Conv2d and overrides the forward method.
33
+
34
+ This class is used to perform 3D convolution on input tensor x. It is a specialized type of convolutional layer
35
+ commonly used in deep learning models for computer vision tasks. The main difference between a regular Conv2d and
36
+ InflatedConv3d is that InflatedConv3d is designed to handle 3D input tensors, which are typically the result of
37
+ inflating 2D convolutional layers to 3D for use in 3D deep learning tasks.
38
+
39
+ Attributes:
40
+ Same as torch.nn.Conv2d.
41
+
42
+ Methods:
43
+ forward(self, x):
44
+ Performs 3D convolution on the input tensor x using the InflatedConv3d layer.
45
+
46
+ Example:
47
+ conv_layer = InflatedConv3d(in_channels=3, out_channels=64, kernel_size=3, stride=1, padding=1)
48
+ output = conv_layer(input_tensor)
49
+ """
50
+ def forward(self, x):
51
+ """
52
+ Forward pass of the InflatedConv3d layer.
53
+
54
+ Args:
55
+ x (torch.Tensor): Input tensor to the layer.
56
+
57
+ Returns:
58
+ torch.Tensor: Output tensor after applying the InflatedConv3d layer.
59
+ """
60
+ video_length = x.shape[2]
61
+
62
+ x = rearrange(x, "b c f h w -> (b f) c h w")
63
+ x = super().forward(x)
64
+ x = rearrange(x, "(b f) c h w -> b c f h w", f=video_length)
65
+
66
+ return x
67
+
68
+
69
+ class InflatedGroupNorm(nn.GroupNorm):
70
+ """
71
+ InflatedGroupNorm is a custom class that inherits from torch.nn.GroupNorm.
72
+ It is used to apply group normalization to 3D tensors.
73
+
74
+ Args:
75
+ num_groups (int): The number of groups to divide the channels into.
76
+ num_channels (int): The number of channels in the input tensor.
77
+ eps (float, optional): A small constant to add to the variance to avoid division by zero. Defaults to 1e-5.
78
+ affine (bool, optional): If True, the module has learnable affine parameters. Defaults to True.
79
+
80
+ Attributes:
81
+ weight (torch.Tensor): The learnable weight tensor for scale.
82
+ bias (torch.Tensor): The learnable bias tensor for shift.
83
+
84
+ Forward method:
85
+ x (torch.Tensor): Input tensor to be normalized.
86
+ return (torch.Tensor): Normalized tensor.
87
+ """
88
+ def forward(self, x):
89
+ """
90
+ Performs a forward pass through the CustomClassName.
91
+
92
+ :param x: Input tensor of shape (batch_size, channels, video_length, height, width).
93
+ :return: Output tensor of shape (batch_size, channels, video_length, height, width).
94
+ """
95
+ video_length = x.shape[2]
96
+
97
+ x = rearrange(x, "b c f h w -> (b f) c h w")
98
+ x = super().forward(x)
99
+ x = rearrange(x, "(b f) c h w -> b c f h w", f=video_length)
100
+
101
+ return x
102
+
103
+
104
+ class Upsample3D(nn.Module):
105
+ """
106
+ Upsample3D is a PyTorch module that upsamples a 3D tensor.
107
+
108
+ Args:
109
+ channels (int): The number of channels in the input tensor.
110
+ use_conv (bool): Whether to use a convolutional layer for upsampling.
111
+ use_conv_transpose (bool): Whether to use a transposed convolutional layer for upsampling.
112
+ out_channels (int): The number of channels in the output tensor.
113
+ name (str): The name of the convolutional layer.
114
+ """
115
+ def __init__(
116
+ self,
117
+ channels,
118
+ use_conv=False,
119
+ use_conv_transpose=False,
120
+ out_channels=None,
121
+ name="conv",
122
+ ):
123
+ super().__init__()
124
+ self.channels = channels
125
+ self.out_channels = out_channels or channels
126
+ self.use_conv = use_conv
127
+ self.use_conv_transpose = use_conv_transpose
128
+ self.name = name
129
+
130
+ if use_conv_transpose:
131
+ raise NotImplementedError
132
+ if use_conv:
133
+ self.conv = InflatedConv3d(self.channels, self.out_channels, 3, padding=1)
134
+
135
+ def forward(self, hidden_states, output_size=None):
136
+ """
137
+ Forward pass of the Upsample3D class.
138
+
139
+ Args:
140
+ hidden_states (torch.Tensor): Input tensor to be upsampled.
141
+ output_size (tuple, optional): Desired output size of the upsampled tensor.
142
+
143
+ Returns:
144
+ torch.Tensor: Upsampled tensor.
145
+
146
+ Raises:
147
+ AssertionError: If the number of channels in the input tensor does not match the expected channels.
148
+ """
149
+ assert hidden_states.shape[1] == self.channels
150
+
151
+ if self.use_conv_transpose:
152
+ raise NotImplementedError
153
+
154
+ # Cast to float32 to as 'upsample_nearest2d_out_frame' op does not support bfloat16
155
+ dtype = hidden_states.dtype
156
+ if dtype == torch.bfloat16:
157
+ hidden_states = hidden_states.to(torch.float32)
158
+
159
+ # upsample_nearest_nhwc fails with large batch sizes. see https://github.com/huggingface/diffusers/issues/984
160
+ if hidden_states.shape[0] >= 64:
161
+ hidden_states = hidden_states.contiguous()
162
+
163
+ # if `output_size` is passed we force the interpolation output
164
+ # size and do not make use of `scale_factor=2`
165
+ if output_size is None:
166
+ hidden_states = F.interpolate(
167
+ hidden_states, scale_factor=[1.0, 2.0, 2.0], mode="nearest"
168
+ )
169
+ else:
170
+ hidden_states = F.interpolate(
171
+ hidden_states, size=output_size, mode="nearest"
172
+ )
173
+
174
+ # If the input is bfloat16, we cast back to bfloat16
175
+ if dtype == torch.bfloat16:
176
+ hidden_states = hidden_states.to(dtype)
177
+
178
+ # if self.use_conv:
179
+ # if self.name == "conv":
180
+ # hidden_states = self.conv(hidden_states)
181
+ # else:
182
+ # hidden_states = self.Conv2d_0(hidden_states)
183
+ hidden_states = self.conv(hidden_states)
184
+
185
+ return hidden_states
186
+
187
+
188
+ class Downsample3D(nn.Module):
189
+ """
190
+ The Downsample3D class is a PyTorch module for downsampling a 3D tensor, which is used to
191
+ reduce the spatial resolution of feature maps, commonly in the encoder part of a neural network.
192
+
193
+ Attributes:
194
+ channels (int): Number of input channels.
195
+ use_conv (bool): Flag to use a convolutional layer for downsampling.
196
+ out_channels (int, optional): Number of output channels. Defaults to input channels if None.
197
+ padding (int): Padding added to the input.
198
+ name (str): Name of the convolutional layer used for downsampling.
199
+
200
+ Methods:
201
+ forward(self, hidden_states):
202
+ Downsamples the input tensor hidden_states and returns the downsampled tensor.
203
+ """
204
+ def __init__(
205
+ self, channels, use_conv=False, out_channels=None, padding=1, name="conv"
206
+ ):
207
+ """
208
+ Downsamples the given input in the 3D space.
209
+
210
+ Args:
211
+ channels: The number of input channels.
212
+ use_conv: Whether to use a convolutional layer for downsampling.
213
+ out_channels: The number of output channels. If None, the input channels are used.
214
+ padding: The amount of padding to be added to the input.
215
+ name: The name of the convolutional layer.
216
+ """
217
+ super().__init__()
218
+ self.channels = channels
219
+ self.out_channels = out_channels or channels
220
+ self.use_conv = use_conv
221
+ self.padding = padding
222
+ stride = 2
223
+ self.name = name
224
+
225
+ if use_conv:
226
+ self.conv = InflatedConv3d(
227
+ self.channels, self.out_channels, 3, stride=stride, padding=padding
228
+ )
229
+ else:
230
+ raise NotImplementedError
231
+
232
+ def forward(self, hidden_states):
233
+ """
234
+ Forward pass for the Downsample3D class.
235
+
236
+ Args:
237
+ hidden_states (torch.Tensor): Input tensor to be downsampled.
238
+
239
+ Returns:
240
+ torch.Tensor: Downsampled tensor.
241
+
242
+ Raises:
243
+ AssertionError: If the number of channels in the input tensor does not match the expected channels.
244
+ """
245
+ assert hidden_states.shape[1] == self.channels
246
+ if self.use_conv and self.padding == 0:
247
+ raise NotImplementedError
248
+
249
+ assert hidden_states.shape[1] == self.channels
250
+ hidden_states = self.conv(hidden_states)
251
+
252
+ return hidden_states
253
+
254
+
255
+ class ResnetBlock3D(nn.Module):
256
+ """
257
+ The ResnetBlock3D class defines a 3D residual block, a common building block in ResNet
258
+ architectures for both image and video modeling tasks.
259
+
260
+ Attributes:
261
+ in_channels (int): Number of input channels.
262
+ out_channels (int, optional): Number of output channels, defaults to in_channels if None.
263
+ conv_shortcut (bool): Flag to use a convolutional shortcut.
264
+ dropout (float): Dropout rate.
265
+ temb_channels (int): Number of channels in the time embedding tensor.
266
+ groups (int): Number of groups for the group normalization layers.
267
+ eps (float): Epsilon value for group normalization.
268
+ non_linearity (str): Type of nonlinearity to apply after convolutions.
269
+ time_embedding_norm (str): Type of normalization for the time embedding.
270
+ output_scale_factor (float): Scaling factor for the output tensor.
271
+ use_in_shortcut (bool): Flag to include the input tensor in the shortcut connection.
272
+ use_inflated_groupnorm (bool): Flag to use inflated group normalization layers.
273
+
274
+ Methods:
275
+ forward(self, input_tensor, temb):
276
+ Passes the input tensor and time embedding through the residual block and
277
+ returns the output tensor.
278
+ """
279
+ def __init__(
280
+ self,
281
+ *,
282
+ in_channels,
283
+ out_channels=None,
284
+ conv_shortcut=False,
285
+ dropout=0.0,
286
+ temb_channels=512,
287
+ groups=32,
288
+ groups_out=None,
289
+ pre_norm=True,
290
+ eps=1e-6,
291
+ non_linearity="swish",
292
+ time_embedding_norm="default",
293
+ output_scale_factor=1.0,
294
+ use_in_shortcut=None,
295
+ use_inflated_groupnorm=None,
296
+ ):
297
+ super().__init__()
298
+ self.pre_norm = pre_norm
299
+ self.pre_norm = True
300
+ self.in_channels = in_channels
301
+ out_channels = in_channels if out_channels is None else out_channels
302
+ self.out_channels = out_channels
303
+ self.use_conv_shortcut = conv_shortcut
304
+ self.time_embedding_norm = time_embedding_norm
305
+ self.output_scale_factor = output_scale_factor
306
+
307
+ if groups_out is None:
308
+ groups_out = groups
309
+
310
+ assert use_inflated_groupnorm is not None
311
+ if use_inflated_groupnorm:
312
+ self.norm1 = InflatedGroupNorm(
313
+ num_groups=groups, num_channels=in_channels, eps=eps, affine=True
314
+ )
315
+ else:
316
+ self.norm1 = torch.nn.GroupNorm(
317
+ num_groups=groups, num_channels=in_channels, eps=eps, affine=True
318
+ )
319
+
320
+ self.conv1 = InflatedConv3d(
321
+ in_channels, out_channels, kernel_size=3, stride=1, padding=1
322
+ )
323
+
324
+ if temb_channels is not None:
325
+ if self.time_embedding_norm == "default":
326
+ time_emb_proj_out_channels = out_channels
327
+ elif self.time_embedding_norm == "scale_shift":
328
+ time_emb_proj_out_channels = out_channels * 2
329
+ else:
330
+ raise ValueError(
331
+ f"unknown time_embedding_norm : {self.time_embedding_norm} "
332
+ )
333
+
334
+ self.time_emb_proj = torch.nn.Linear(
335
+ temb_channels, time_emb_proj_out_channels
336
+ )
337
+ else:
338
+ self.time_emb_proj = None
339
+
340
+ if use_inflated_groupnorm:
341
+ self.norm2 = InflatedGroupNorm(
342
+ num_groups=groups_out, num_channels=out_channels, eps=eps, affine=True
343
+ )
344
+ else:
345
+ self.norm2 = torch.nn.GroupNorm(
346
+ num_groups=groups_out, num_channels=out_channels, eps=eps, affine=True
347
+ )
348
+ self.dropout = torch.nn.Dropout(dropout)
349
+ self.conv2 = InflatedConv3d(
350
+ out_channels, out_channels, kernel_size=3, stride=1, padding=1
351
+ )
352
+
353
+ if non_linearity == "swish":
354
+ self.nonlinearity = F.silu()
355
+ elif non_linearity == "mish":
356
+ self.nonlinearity = Mish()
357
+ elif non_linearity == "silu":
358
+ self.nonlinearity = nn.SiLU()
359
+
360
+ self.use_in_shortcut = (
361
+ self.in_channels != self.out_channels
362
+ if use_in_shortcut is None
363
+ else use_in_shortcut
364
+ )
365
+
366
+ self.conv_shortcut = None
367
+ if self.use_in_shortcut:
368
+ self.conv_shortcut = InflatedConv3d(
369
+ in_channels, out_channels, kernel_size=1, stride=1, padding=0
370
+ )
371
+
372
+ def forward(self, input_tensor, temb):
373
+ """
374
+ Forward pass for the ResnetBlock3D class.
375
+
376
+ Args:
377
+ input_tensor (torch.Tensor): Input tensor to the ResnetBlock3D layer.
378
+ temb (torch.Tensor): Token embedding tensor.
379
+
380
+ Returns:
381
+ torch.Tensor: Output tensor after passing through the ResnetBlock3D layer.
382
+ """
383
+ hidden_states = input_tensor
384
+
385
+ hidden_states = self.norm1(hidden_states)
386
+ hidden_states = self.nonlinearity(hidden_states)
387
+
388
+ hidden_states = self.conv1(hidden_states)
389
+
390
+ if temb is not None:
391
+ temb = self.time_emb_proj(self.nonlinearity(temb))[:, :, None, None, None]
392
+
393
+ if temb is not None and self.time_embedding_norm == "default":
394
+ hidden_states = hidden_states + temb
395
+
396
+ hidden_states = self.norm2(hidden_states)
397
+
398
+ if temb is not None and self.time_embedding_norm == "scale_shift":
399
+ scale, shift = torch.chunk(temb, 2, dim=1)
400
+ hidden_states = hidden_states * (1 + scale) + shift
401
+
402
+ hidden_states = self.nonlinearity(hidden_states)
403
+
404
+ hidden_states = self.dropout(hidden_states)
405
+ hidden_states = self.conv2(hidden_states)
406
+
407
+ if self.conv_shortcut is not None:
408
+ input_tensor = self.conv_shortcut(input_tensor)
409
+
410
+ output_tensor = (input_tensor + hidden_states) / self.output_scale_factor
411
+
412
+ return output_tensor
413
+
414
+
415
+ class Mish(torch.nn.Module):
416
+ """
417
+ The Mish class implements the Mish activation function, a smooth, non-monotonic function
418
+ that can be used in neural networks as an alternative to traditional activation functions like ReLU.
419
+
420
+ Methods:
421
+ forward(self, hidden_states):
422
+ Applies the Mish activation function to the input tensor hidden_states and
423
+ returns the resulting tensor.
424
+ """
425
+ def forward(self, hidden_states):
426
+ """
427
+ Mish activation function.
428
+
429
+ Args:
430
+ hidden_states (torch.Tensor): The input tensor to apply the Mish activation function to.
431
+
432
+ Returns:
433
+ hidden_states (torch.Tensor): The output tensor after applying the Mish activation function.
434
+ """
435
+ return hidden_states * torch.tanh(torch.nn.functional.softplus(hidden_states))
hallo/models/transformer_2d.py ADDED
@@ -0,0 +1,431 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # pylint: disable=E1101
2
+ # src/models/transformer_2d.py
3
+
4
+ """
5
+ This module defines the Transformer2DModel, a PyTorch model that extends ModelMixin and ConfigMixin. It includes
6
+ methods for gradient checkpointing, forward propagation, and various utility functions. The model is designed for
7
+ 2D image-related tasks and uses LoRa (Low-Rank All-Attention) compatible layers for efficient attention computation.
8
+
9
+ The file includes the following import statements:
10
+
11
+ - From dataclasses import dataclass
12
+ - From typing import Any, Dict, Optional
13
+ - Import torch
14
+ - From diffusers.configuration_utils import ConfigMixin, register_to_config
15
+ - From diffusers.models.lora import LoRACompatibleConv, LoRACompatibleLinear
16
+ - From diffusers.models.modeling_utils import ModelMixin
17
+ - From diffusers.models.normalization import AdaLayerNormSingle
18
+ - From diffusers.utils import (USE_PEFT_BACKEND, BaseOutput, deprecate,
19
+ is_torch_version)
20
+ - From torch import nn
21
+ - From .attention import BasicTransformerBlock
22
+
23
+ The file also includes the following classes and functions:
24
+
25
+ - Transformer2DModel: A model class that extends ModelMixin and ConfigMixin. It includes methods for gradient
26
+ checkpointing, forward propagation, and various utility functions.
27
+ - _set_gradient_checkpointing: A utility function to set gradient checkpointing for a given module.
28
+ - forward: The forward propagation method for the Transformer2DModel.
29
+
30
+ To use this module, you can import the Transformer2DModel class and create an instance of the model with the desired
31
+ configuration. Then, you can use the forward method to pass input tensors through the model and get the output tensors.
32
+ """
33
+
34
+ from dataclasses import dataclass
35
+ from typing import Any, Dict, Optional
36
+
37
+ import torch
38
+ from diffusers.configuration_utils import ConfigMixin, register_to_config
39
+ # from diffusers.models.embeddings import CaptionProjection
40
+ from diffusers.models.lora import LoRACompatibleConv, LoRACompatibleLinear
41
+ from diffusers.models.modeling_utils import ModelMixin
42
+ from diffusers.models.normalization import AdaLayerNormSingle
43
+ from diffusers.utils import (USE_PEFT_BACKEND, BaseOutput, deprecate,
44
+ is_torch_version)
45
+ from torch import nn
46
+
47
+ from .attention import BasicTransformerBlock
48
+
49
+
50
+ @dataclass
51
+ class Transformer2DModelOutput(BaseOutput):
52
+ """
53
+ The output of [`Transformer2DModel`].
54
+
55
+ Args:
56
+ sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`
57
+ or `(batch size, num_vector_embeds - 1, num_latent_pixels)` if [`Transformer2DModel`] is discrete):
58
+ The hidden states output conditioned on the `encoder_hidden_states` input. If discrete, returns probability
59
+ distributions for the unnoised latent pixels.
60
+ """
61
+
62
+ sample: torch.FloatTensor
63
+ ref_feature: torch.FloatTensor
64
+
65
+
66
+ class Transformer2DModel(ModelMixin, ConfigMixin):
67
+ """
68
+ A 2D Transformer model for image-like data.
69
+
70
+ Parameters:
71
+ num_attention_heads (`int`, *optional*, defaults to 16): The number of heads to use for multi-head attention.
72
+ attention_head_dim (`int`, *optional*, defaults to 88): The number of channels in each head.
73
+ in_channels (`int`, *optional*):
74
+ The number of channels in the input and output (specify if the input is **continuous**).
75
+ num_layers (`int`, *optional*, defaults to 1): The number of layers of Transformer blocks to use.
76
+ dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
77
+ cross_attention_dim (`int`, *optional*): The number of `encoder_hidden_states` dimensions to use.
78
+ sample_size (`int`, *optional*): The width of the latent images (specify if the input is **discrete**).
79
+ This is fixed during training since it is used to learn a number of position embeddings.
80
+ num_vector_embeds (`int`, *optional*):
81
+ The number of classes of the vector embeddings of the latent pixels (specify if the input is **discrete**).
82
+ Includes the class for the masked latent pixel.
83
+ activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to use in feed-forward.
84
+ num_embeds_ada_norm ( `int`, *optional*):
85
+ The number of diffusion steps used during training. Pass if at least one of the norm_layers is
86
+ `AdaLayerNorm`. This is fixed during training since it is used to learn a number of embeddings that are
87
+ added to the hidden states.
88
+
89
+ During inference, you can denoise for up to but not more steps than `num_embeds_ada_norm`.
90
+ attention_bias (`bool`, *optional*):
91
+ Configure if the `TransformerBlocks` attention should contain a bias parameter.
92
+ """
93
+
94
+ _supports_gradient_checkpointing = True
95
+
96
+ @register_to_config
97
+ def __init__(
98
+ self,
99
+ num_attention_heads: int = 16,
100
+ attention_head_dim: int = 88,
101
+ in_channels: Optional[int] = None,
102
+ out_channels: Optional[int] = None,
103
+ num_layers: int = 1,
104
+ dropout: float = 0.0,
105
+ norm_num_groups: int = 32,
106
+ cross_attention_dim: Optional[int] = None,
107
+ attention_bias: bool = False,
108
+ num_vector_embeds: Optional[int] = None,
109
+ patch_size: Optional[int] = None,
110
+ activation_fn: str = "geglu",
111
+ num_embeds_ada_norm: Optional[int] = None,
112
+ use_linear_projection: bool = False,
113
+ only_cross_attention: bool = False,
114
+ double_self_attention: bool = False,
115
+ upcast_attention: bool = False,
116
+ norm_type: str = "layer_norm",
117
+ norm_elementwise_affine: bool = True,
118
+ norm_eps: float = 1e-5,
119
+ attention_type: str = "default",
120
+ ):
121
+ super().__init__()
122
+ self.use_linear_projection = use_linear_projection
123
+ self.num_attention_heads = num_attention_heads
124
+ self.attention_head_dim = attention_head_dim
125
+ inner_dim = num_attention_heads * attention_head_dim
126
+
127
+ conv_cls = nn.Conv2d if USE_PEFT_BACKEND else LoRACompatibleConv
128
+ linear_cls = nn.Linear if USE_PEFT_BACKEND else LoRACompatibleLinear
129
+
130
+ # 1. Transformer2DModel can process both standard continuous images of
131
+ # shape `(batch_size, num_channels, width, height)` as well as quantized image embeddings of
132
+ # shape `(batch_size, num_image_vectors)`
133
+ # Define whether input is continuous or discrete depending on configuration
134
+ self.is_input_continuous = (in_channels is not None) and (patch_size is None)
135
+ self.is_input_vectorized = num_vector_embeds is not None
136
+ self.is_input_patches = in_channels is not None and patch_size is not None
137
+
138
+ if norm_type == "layer_norm" and num_embeds_ada_norm is not None:
139
+ deprecation_message = (
140
+ f"The configuration file of this model: {self.__class__} is outdated. `norm_type` is either not set or"
141
+ " incorrectly set to `'layer_norm'`.Make sure to set `norm_type` to `'ada_norm'` in the config."
142
+ " Please make sure to update the config accordingly as leaving `norm_type` might led to incorrect"
143
+ " results in future versions. If you have downloaded this checkpoint from the Hugging Face Hub, it"
144
+ " would be very nice if you could open a Pull request for the `transformer/config.json` file"
145
+ )
146
+ deprecate(
147
+ "norm_type!=num_embeds_ada_norm",
148
+ "1.0.0",
149
+ deprecation_message,
150
+ standard_warn=False,
151
+ )
152
+ norm_type = "ada_norm"
153
+
154
+ if self.is_input_continuous and self.is_input_vectorized:
155
+ raise ValueError(
156
+ f"Cannot define both `in_channels`: {in_channels} and `num_vector_embeds`: {num_vector_embeds}. Make"
157
+ " sure that either `in_channels` or `num_vector_embeds` is None."
158
+ )
159
+
160
+ if self.is_input_vectorized and self.is_input_patches:
161
+ raise ValueError(
162
+ f"Cannot define both `num_vector_embeds`: {num_vector_embeds} and `patch_size`: {patch_size}. Make"
163
+ " sure that either `num_vector_embeds` or `num_patches` is None."
164
+ )
165
+
166
+ if (
167
+ not self.is_input_continuous
168
+ and not self.is_input_vectorized
169
+ and not self.is_input_patches
170
+ ):
171
+ raise ValueError(
172
+ f"Has to define `in_channels`: {in_channels}, `num_vector_embeds`: {num_vector_embeds}, or patch_size:"
173
+ f" {patch_size}. Make sure that `in_channels`, `num_vector_embeds` or `num_patches` is not None."
174
+ )
175
+
176
+ # 2. Define input layers
177
+ self.in_channels = in_channels
178
+
179
+ self.norm = torch.nn.GroupNorm(
180
+ num_groups=norm_num_groups,
181
+ num_channels=in_channels,
182
+ eps=1e-6,
183
+ affine=True,
184
+ )
185
+ if use_linear_projection:
186
+ self.proj_in = linear_cls(in_channels, inner_dim)
187
+ else:
188
+ self.proj_in = conv_cls(
189
+ in_channels, inner_dim, kernel_size=1, stride=1, padding=0
190
+ )
191
+
192
+ # 3. Define transformers blocks
193
+ self.transformer_blocks = nn.ModuleList(
194
+ [
195
+ BasicTransformerBlock(
196
+ inner_dim,
197
+ num_attention_heads,
198
+ attention_head_dim,
199
+ dropout=dropout,
200
+ cross_attention_dim=cross_attention_dim,
201
+ activation_fn=activation_fn,
202
+ num_embeds_ada_norm=num_embeds_ada_norm,
203
+ attention_bias=attention_bias,
204
+ only_cross_attention=only_cross_attention,
205
+ double_self_attention=double_self_attention,
206
+ upcast_attention=upcast_attention,
207
+ norm_type=norm_type,
208
+ norm_elementwise_affine=norm_elementwise_affine,
209
+ norm_eps=norm_eps,
210
+ attention_type=attention_type,
211
+ )
212
+ for d in range(num_layers)
213
+ ]
214
+ )
215
+
216
+ # 4. Define output layers
217
+ self.out_channels = in_channels if out_channels is None else out_channels
218
+ # TODO: should use out_channels for continuous projections
219
+ if use_linear_projection:
220
+ self.proj_out = linear_cls(inner_dim, in_channels)
221
+ else:
222
+ self.proj_out = conv_cls(
223
+ inner_dim, in_channels, kernel_size=1, stride=1, padding=0
224
+ )
225
+
226
+ # 5. PixArt-Alpha blocks.
227
+ self.adaln_single = None
228
+ self.use_additional_conditions = False
229
+ if norm_type == "ada_norm_single":
230
+ self.use_additional_conditions = self.config.sample_size == 128
231
+ # TODO(Sayak, PVP) clean this, for now we use sample size to determine whether to use
232
+ # additional conditions until we find better name
233
+ self.adaln_single = AdaLayerNormSingle(
234
+ inner_dim, use_additional_conditions=self.use_additional_conditions
235
+ )
236
+
237
+ self.caption_projection = None
238
+
239
+ self.gradient_checkpointing = False
240
+
241
+ def _set_gradient_checkpointing(self, module, value=False):
242
+ if hasattr(module, "gradient_checkpointing"):
243
+ module.gradient_checkpointing = value
244
+
245
+ def forward(
246
+ self,
247
+ hidden_states: torch.Tensor,
248
+ encoder_hidden_states: Optional[torch.Tensor] = None,
249
+ timestep: Optional[torch.LongTensor] = None,
250
+ _added_cond_kwargs: Dict[str, torch.Tensor] = None,
251
+ class_labels: Optional[torch.LongTensor] = None,
252
+ cross_attention_kwargs: Dict[str, Any] = None,
253
+ attention_mask: Optional[torch.Tensor] = None,
254
+ encoder_attention_mask: Optional[torch.Tensor] = None,
255
+ return_dict: bool = True,
256
+ ):
257
+ """
258
+ The [`Transformer2DModel`] forward method.
259
+
260
+ Args:
261
+ hidden_states (`torch.LongTensor` of shape `(batch size, num latent pixels)` if discrete,
262
+ `torch.FloatTensor` of shape `(batch size, channel, height, width)` if continuous):
263
+ Input `hidden_states`.
264
+ encoder_hidden_states ( `torch.FloatTensor` of shape `(batch size, sequence len, embed dims)`, *optional*):
265
+ Conditional embeddings for cross attention layer. If not given, cross-attention defaults to
266
+ self-attention.
267
+ timestep ( `torch.LongTensor`, *optional*):
268
+ Used to indicate denoising step. Optional timestep to be applied as an embedding in `AdaLayerNorm`.
269
+ class_labels ( `torch.LongTensor` of shape `(batch size, num classes)`, *optional*):
270
+ Used to indicate class labels conditioning. Optional class labels to be applied as an embedding in
271
+ `AdaLayerZeroNorm`.
272
+ cross_attention_kwargs ( `Dict[str, Any]`, *optional*):
273
+ A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
274
+ `self.processor` in
275
+ [diffusers.models.attention_processor]
276
+ (https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
277
+ attention_mask ( `torch.Tensor`, *optional*):
278
+ An attention mask of shape `(batch, key_tokens)` is applied to `encoder_hidden_states`. If `1` the mask
279
+ is kept, otherwise if `0` it is discarded. Mask will be converted into a bias, which adds large
280
+ negative values to the attention scores corresponding to "discard" tokens.
281
+ encoder_attention_mask ( `torch.Tensor`, *optional*):
282
+ Cross-attention mask applied to `encoder_hidden_states`. Two formats supported:
283
+
284
+ * Mask `(batch, sequence_length)` True = keep, False = discard.
285
+ * Bias `(batch, 1, sequence_length)` 0 = keep, -10000 = discard.
286
+
287
+ If `ndim == 2`: will be interpreted as a mask, then converted into a bias consistent with the format
288
+ above. This bias will be added to the cross-attention scores.
289
+ return_dict (`bool`, *optional*, defaults to `True`):
290
+ Whether or not to return a [`~models.unet_2d_condition.UNet2DConditionOutput`] instead of a plain
291
+ tuple.
292
+
293
+ Returns:
294
+ If `return_dict` is True, an [`~models.transformer_2d.Transformer2DModelOutput`] is returned, otherwise a
295
+ `tuple` where the first element is the sample tensor.
296
+ """
297
+ # ensure attention_mask is a bias, and give it a singleton query_tokens dimension.
298
+ # we may have done this conversion already, e.g. if we came here via UNet2DConditionModel#forward.
299
+ # we can tell by counting dims; if ndim == 2: it's a mask rather than a bias.
300
+ # expects mask of shape:
301
+ # [batch, key_tokens]
302
+ # adds singleton query_tokens dimension:
303
+ # [batch, 1, key_tokens]
304
+ # this helps to broadcast it as a bias over attention scores, which will be in one of the following shapes:
305
+ # [batch, heads, query_tokens, key_tokens] (e.g. torch sdp attn)
306
+ # [batch * heads, query_tokens, key_tokens] (e.g. xformers or classic attn)
307
+ if attention_mask is not None and attention_mask.ndim == 2:
308
+ # assume that mask is expressed as:
309
+ # (1 = keep, 0 = discard)
310
+ # convert mask into a bias that can be added to attention scores:
311
+ # (keep = +0, discard = -10000.0)
312
+ attention_mask = (1 - attention_mask.to(hidden_states.dtype)) * -10000.0
313
+ attention_mask = attention_mask.unsqueeze(1)
314
+
315
+ # convert encoder_attention_mask to a bias the same way we do for attention_mask
316
+ if encoder_attention_mask is not None and encoder_attention_mask.ndim == 2:
317
+ encoder_attention_mask = (
318
+ 1 - encoder_attention_mask.to(hidden_states.dtype)
319
+ ) * -10000.0
320
+ encoder_attention_mask = encoder_attention_mask.unsqueeze(1)
321
+
322
+ # Retrieve lora scale.
323
+ lora_scale = (
324
+ cross_attention_kwargs.get("scale", 1.0)
325
+ if cross_attention_kwargs is not None
326
+ else 1.0
327
+ )
328
+
329
+ # 1. Input
330
+ batch, _, height, width = hidden_states.shape
331
+ residual = hidden_states
332
+
333
+ hidden_states = self.norm(hidden_states)
334
+ if not self.use_linear_projection:
335
+ hidden_states = (
336
+ self.proj_in(hidden_states, scale=lora_scale)
337
+ if not USE_PEFT_BACKEND
338
+ else self.proj_in(hidden_states)
339
+ )
340
+ inner_dim = hidden_states.shape[1]
341
+ hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(
342
+ batch, height * width, inner_dim
343
+ )
344
+ else:
345
+ inner_dim = hidden_states.shape[1]
346
+ hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(
347
+ batch, height * width, inner_dim
348
+ )
349
+ hidden_states = (
350
+ self.proj_in(hidden_states, scale=lora_scale)
351
+ if not USE_PEFT_BACKEND
352
+ else self.proj_in(hidden_states)
353
+ )
354
+
355
+ # 2. Blocks
356
+ if self.caption_projection is not None:
357
+ batch_size = hidden_states.shape[0]
358
+ encoder_hidden_states = self.caption_projection(encoder_hidden_states)
359
+ encoder_hidden_states = encoder_hidden_states.view(
360
+ batch_size, -1, hidden_states.shape[-1]
361
+ )
362
+
363
+ ref_feature = hidden_states.reshape(batch, height, width, inner_dim)
364
+ for block in self.transformer_blocks:
365
+ if self.training and self.gradient_checkpointing:
366
+
367
+ def create_custom_forward(module, return_dict=None):
368
+ def custom_forward(*inputs):
369
+ if return_dict is not None:
370
+ return module(*inputs, return_dict=return_dict)
371
+
372
+ return module(*inputs)
373
+
374
+ return custom_forward
375
+
376
+ ckpt_kwargs: Dict[str, Any] = (
377
+ {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
378
+ )
379
+ hidden_states = torch.utils.checkpoint.checkpoint(
380
+ create_custom_forward(block),
381
+ hidden_states,
382
+ attention_mask,
383
+ encoder_hidden_states,
384
+ encoder_attention_mask,
385
+ timestep,
386
+ cross_attention_kwargs,
387
+ class_labels,
388
+ **ckpt_kwargs,
389
+ )
390
+ else:
391
+ hidden_states = block(
392
+ hidden_states, # shape [5, 4096, 320]
393
+ attention_mask=attention_mask,
394
+ encoder_hidden_states=encoder_hidden_states, # shape [1,4,768]
395
+ encoder_attention_mask=encoder_attention_mask,
396
+ timestep=timestep,
397
+ cross_attention_kwargs=cross_attention_kwargs,
398
+ class_labels=class_labels,
399
+ )
400
+
401
+ # 3. Output
402
+ output = None
403
+ if self.is_input_continuous:
404
+ if not self.use_linear_projection:
405
+ hidden_states = (
406
+ hidden_states.reshape(batch, height, width, inner_dim)
407
+ .permute(0, 3, 1, 2)
408
+ .contiguous()
409
+ )
410
+ hidden_states = (
411
+ self.proj_out(hidden_states, scale=lora_scale)
412
+ if not USE_PEFT_BACKEND
413
+ else self.proj_out(hidden_states)
414
+ )
415
+ else:
416
+ hidden_states = (
417
+ self.proj_out(hidden_states, scale=lora_scale)
418
+ if not USE_PEFT_BACKEND
419
+ else self.proj_out(hidden_states)
420
+ )
421
+ hidden_states = (
422
+ hidden_states.reshape(batch, height, width, inner_dim)
423
+ .permute(0, 3, 1, 2)
424
+ .contiguous()
425
+ )
426
+
427
+ output = hidden_states + residual
428
+ if not return_dict:
429
+ return (output, ref_feature)
430
+
431
+ return Transformer2DModelOutput(sample=output, ref_feature=ref_feature)
hallo/models/transformer_3d.py ADDED
@@ -0,0 +1,257 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # pylint: disable=R0801
2
+ """
3
+ This module implements the Transformer3DModel, a PyTorch model designed for processing
4
+ 3D data such as videos. It extends ModelMixin and ConfigMixin to provide a transformer
5
+ model with support for gradient checkpointing and various types of attention mechanisms.
6
+ The model can be configured with different parameters such as the number of attention heads,
7
+ attention head dimension, and the number of layers. It also supports the use of audio modules
8
+ for enhanced feature extraction from video data.
9
+ """
10
+
11
+ from dataclasses import dataclass
12
+ from typing import Optional
13
+
14
+ import torch
15
+ from diffusers.configuration_utils import ConfigMixin, register_to_config
16
+ from diffusers.models import ModelMixin
17
+ from diffusers.utils import BaseOutput
18
+ from einops import rearrange, repeat
19
+ from torch import nn
20
+
21
+ from .attention import (AudioTemporalBasicTransformerBlock,
22
+ TemporalBasicTransformerBlock)
23
+
24
+
25
+ @dataclass
26
+ class Transformer3DModelOutput(BaseOutput):
27
+ """
28
+ The output of the [`Transformer3DModel`].
29
+
30
+ Attributes:
31
+ sample (`torch.FloatTensor`):
32
+ The output tensor from the transformer model, which is the result of processing the input
33
+ hidden states through the transformer blocks and any subsequent layers.
34
+ """
35
+ sample: torch.FloatTensor
36
+
37
+
38
+ class Transformer3DModel(ModelMixin, ConfigMixin):
39
+ """
40
+ Transformer3DModel is a PyTorch model that extends `ModelMixin` and `ConfigMixin` to create a 3D transformer model.
41
+ It implements the forward pass for processing input hidden states, encoder hidden states, and various types of attention masks.
42
+ The model supports gradient checkpointing, which can be enabled by calling the `enable_gradient_checkpointing()` method.
43
+ """
44
+ _supports_gradient_checkpointing = True
45
+
46
+ @register_to_config
47
+ def __init__(
48
+ self,
49
+ num_attention_heads: int = 16,
50
+ attention_head_dim: int = 88,
51
+ in_channels: Optional[int] = None,
52
+ num_layers: int = 1,
53
+ dropout: float = 0.0,
54
+ norm_num_groups: int = 32,
55
+ cross_attention_dim: Optional[int] = None,
56
+ attention_bias: bool = False,
57
+ activation_fn: str = "geglu",
58
+ num_embeds_ada_norm: Optional[int] = None,
59
+ use_linear_projection: bool = False,
60
+ only_cross_attention: bool = False,
61
+ upcast_attention: bool = False,
62
+ unet_use_cross_frame_attention=None,
63
+ unet_use_temporal_attention=None,
64
+ use_audio_module=False,
65
+ depth=0,
66
+ unet_block_name=None,
67
+ stack_enable_blocks_name = None,
68
+ stack_enable_blocks_depth = None,
69
+ ):
70
+ super().__init__()
71
+ self.use_linear_projection = use_linear_projection
72
+ self.num_attention_heads = num_attention_heads
73
+ self.attention_head_dim = attention_head_dim
74
+ inner_dim = num_attention_heads * attention_head_dim
75
+ self.use_audio_module = use_audio_module
76
+ # Define input layers
77
+ self.in_channels = in_channels
78
+
79
+ self.norm = torch.nn.GroupNorm(
80
+ num_groups=norm_num_groups, num_channels=in_channels, eps=1e-6, affine=True
81
+ )
82
+ if use_linear_projection:
83
+ self.proj_in = nn.Linear(in_channels, inner_dim)
84
+ else:
85
+ self.proj_in = nn.Conv2d(
86
+ in_channels, inner_dim, kernel_size=1, stride=1, padding=0
87
+ )
88
+
89
+ if use_audio_module:
90
+ self.transformer_blocks = nn.ModuleList(
91
+ [
92
+ AudioTemporalBasicTransformerBlock(
93
+ inner_dim,
94
+ num_attention_heads,
95
+ attention_head_dim,
96
+ dropout=dropout,
97
+ cross_attention_dim=cross_attention_dim,
98
+ activation_fn=activation_fn,
99
+ num_embeds_ada_norm=num_embeds_ada_norm,
100
+ attention_bias=attention_bias,
101
+ only_cross_attention=only_cross_attention,
102
+ upcast_attention=upcast_attention,
103
+ unet_use_cross_frame_attention=unet_use_cross_frame_attention,
104
+ unet_use_temporal_attention=unet_use_temporal_attention,
105
+ depth=depth,
106
+ unet_block_name=unet_block_name,
107
+ stack_enable_blocks_name=stack_enable_blocks_name,
108
+ stack_enable_blocks_depth=stack_enable_blocks_depth,
109
+ )
110
+ for d in range(num_layers)
111
+ ]
112
+ )
113
+ else:
114
+ # Define transformers blocks
115
+ self.transformer_blocks = nn.ModuleList(
116
+ [
117
+ TemporalBasicTransformerBlock(
118
+ inner_dim,
119
+ num_attention_heads,
120
+ attention_head_dim,
121
+ dropout=dropout,
122
+ cross_attention_dim=cross_attention_dim,
123
+ activation_fn=activation_fn,
124
+ num_embeds_ada_norm=num_embeds_ada_norm,
125
+ attention_bias=attention_bias,
126
+ only_cross_attention=only_cross_attention,
127
+ upcast_attention=upcast_attention,
128
+ )
129
+ for d in range(num_layers)
130
+ ]
131
+ )
132
+
133
+ # 4. Define output layers
134
+ if use_linear_projection:
135
+ self.proj_out = nn.Linear(in_channels, inner_dim)
136
+ else:
137
+ self.proj_out = nn.Conv2d(
138
+ inner_dim, in_channels, kernel_size=1, stride=1, padding=0
139
+ )
140
+
141
+ self.gradient_checkpointing = False
142
+
143
+ def _set_gradient_checkpointing(self, module, value=False):
144
+ if hasattr(module, "gradient_checkpointing"):
145
+ module.gradient_checkpointing = value
146
+
147
+ def forward(
148
+ self,
149
+ hidden_states,
150
+ encoder_hidden_states=None,
151
+ attention_mask=None,
152
+ full_mask=None,
153
+ face_mask=None,
154
+ lip_mask=None,
155
+ motion_scale=None,
156
+ timestep=None,
157
+ return_dict: bool = True,
158
+ ):
159
+ """
160
+ Forward pass for the Transformer3DModel.
161
+
162
+ Args:
163
+ hidden_states (torch.Tensor): The input hidden states.
164
+ encoder_hidden_states (torch.Tensor, optional): The input encoder hidden states.
165
+ attention_mask (torch.Tensor, optional): The attention mask.
166
+ full_mask (torch.Tensor, optional): The full mask.
167
+ face_mask (torch.Tensor, optional): The face mask.
168
+ lip_mask (torch.Tensor, optional): The lip mask.
169
+ timestep (int, optional): The current timestep.
170
+ return_dict (bool, optional): Whether to return a dictionary or a tuple.
171
+
172
+ Returns:
173
+ output (Union[Tuple, BaseOutput]): The output of the Transformer3DModel.
174
+ """
175
+ # Input
176
+ assert (
177
+ hidden_states.dim() == 5
178
+ ), f"Expected hidden_states to have ndim=5, but got ndim={hidden_states.dim()}."
179
+ video_length = hidden_states.shape[2]
180
+ hidden_states = rearrange(hidden_states, "b c f h w -> (b f) c h w")
181
+
182
+ # TODO
183
+ if self.use_audio_module:
184
+ encoder_hidden_states = rearrange(
185
+ encoder_hidden_states,
186
+ "bs f margin dim -> (bs f) margin dim",
187
+ )
188
+ else:
189
+ if encoder_hidden_states.shape[0] != hidden_states.shape[0]:
190
+ encoder_hidden_states = repeat(
191
+ encoder_hidden_states, "b n c -> (b f) n c", f=video_length
192
+ )
193
+
194
+ batch, _, height, weight = hidden_states.shape
195
+ residual = hidden_states
196
+
197
+ hidden_states = self.norm(hidden_states)
198
+ if not self.use_linear_projection:
199
+ hidden_states = self.proj_in(hidden_states)
200
+ inner_dim = hidden_states.shape[1]
201
+ hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(
202
+ batch, height * weight, inner_dim
203
+ )
204
+ else:
205
+ inner_dim = hidden_states.shape[1]
206
+ hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(
207
+ batch, height * weight, inner_dim
208
+ )
209
+ hidden_states = self.proj_in(hidden_states)
210
+
211
+ # Blocks
212
+ motion_frames = []
213
+ for _, block in enumerate(self.transformer_blocks):
214
+ if isinstance(block, TemporalBasicTransformerBlock):
215
+ hidden_states, motion_frame_fea = block(
216
+ hidden_states,
217
+ encoder_hidden_states=encoder_hidden_states,
218
+ timestep=timestep,
219
+ video_length=video_length,
220
+ )
221
+ motion_frames.append(motion_frame_fea)
222
+ else:
223
+ hidden_states = block(
224
+ hidden_states, # shape [2, 4096, 320]
225
+ encoder_hidden_states=encoder_hidden_states, # shape [2, 20, 640]
226
+ attention_mask=attention_mask,
227
+ full_mask=full_mask,
228
+ face_mask=face_mask,
229
+ lip_mask=lip_mask,
230
+ timestep=timestep,
231
+ video_length=video_length,
232
+ motion_scale=motion_scale,
233
+ )
234
+
235
+ # Output
236
+ if not self.use_linear_projection:
237
+ hidden_states = (
238
+ hidden_states.reshape(batch, height, weight, inner_dim)
239
+ .permute(0, 3, 1, 2)
240
+ .contiguous()
241
+ )
242
+ hidden_states = self.proj_out(hidden_states)
243
+ else:
244
+ hidden_states = self.proj_out(hidden_states)
245
+ hidden_states = (
246
+ hidden_states.reshape(batch, height, weight, inner_dim)
247
+ .permute(0, 3, 1, 2)
248
+ .contiguous()
249
+ )
250
+
251
+ output = hidden_states + residual
252
+
253
+ output = rearrange(output, "(b f) c h w -> b c f h w", f=video_length)
254
+ if not return_dict:
255
+ return (output, motion_frames)
256
+
257
+ return Transformer3DModelOutput(sample=output)
hallo/models/unet_2d_blocks.py ADDED
@@ -0,0 +1,1343 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # pylint: disable=R0801
2
+ # pylint: disable=W1203
3
+
4
+ """
5
+ This file defines the 2D blocks for the UNet model in a PyTorch implementation.
6
+ The UNet model is a popular architecture for image segmentation tasks,
7
+ which consists of an encoder, a decoder, and a skip connection mechanism.
8
+ The 2D blocks in this file include various types of layers, such as ResNet blocks,
9
+ Transformer blocks, and cross-attention blocks,
10
+ which are used to build the encoder and decoder parts of the UNet model.
11
+ The AutoencoderTinyBlock class is a simple autoencoder block for tiny models,
12
+ and the UNetMidBlock2D and CrossAttnDownBlock2D, DownBlock2D, CrossAttnUpBlock2D,
13
+ and UpBlock2D classes are used for the middle and decoder parts of the UNet model.
14
+ The classes and functions in this file provide a flexible and modular way
15
+ to construct the UNet model for different image segmentation tasks.
16
+ """
17
+
18
+ from typing import Any, Dict, Optional, Tuple, Union
19
+
20
+ import torch
21
+ from diffusers.models.activations import get_activation
22
+ from diffusers.models.attention_processor import Attention
23
+ from diffusers.models.resnet import Downsample2D, ResnetBlock2D, Upsample2D
24
+ from diffusers.models.transformers.dual_transformer_2d import \
25
+ DualTransformer2DModel
26
+ from diffusers.utils import is_torch_version, logging
27
+ from diffusers.utils.torch_utils import apply_freeu
28
+ from torch import nn
29
+
30
+ from .transformer_2d import Transformer2DModel
31
+
32
+ logger = logging.get_logger(__name__) # pylint: disable=invalid-name
33
+
34
+
35
+ def get_down_block(
36
+ down_block_type: str,
37
+ num_layers: int,
38
+ in_channels: int,
39
+ out_channels: int,
40
+ temb_channels: int,
41
+ add_downsample: bool,
42
+ resnet_eps: float,
43
+ resnet_act_fn: str,
44
+ transformer_layers_per_block: int = 1,
45
+ num_attention_heads: Optional[int] = None,
46
+ resnet_groups: Optional[int] = None,
47
+ cross_attention_dim: Optional[int] = None,
48
+ downsample_padding: Optional[int] = None,
49
+ dual_cross_attention: bool = False,
50
+ use_linear_projection: bool = False,
51
+ only_cross_attention: bool = False,
52
+ upcast_attention: bool = False,
53
+ resnet_time_scale_shift: str = "default",
54
+ attention_type: str = "default",
55
+ attention_head_dim: Optional[int] = None,
56
+ dropout: float = 0.0,
57
+ ):
58
+ """ This function creates and returns a UpBlock2D or CrossAttnUpBlock2D object based on the given up_block_type.
59
+
60
+ Args:
61
+ up_block_type (str): The type of up block to create. Must be either "UpBlock2D" or "CrossAttnUpBlock2D".
62
+ num_layers (int): The number of layers in the ResNet block.
63
+ in_channels (int): The number of input channels.
64
+ out_channels (int): The number of output channels.
65
+ prev_output_channel (int): The number of channels in the previous output.
66
+ temb_channels (int): The number of channels in the token embedding.
67
+ add_upsample (bool): Whether to add an upsample layer after the ResNet block. Defaults to True.
68
+ resnet_eps (float): The epsilon value for the ResNet block. Defaults to 1e-6.
69
+ resnet_act_fn (str): The activation function to use in the ResNet block. Defaults to "swish".
70
+ resnet_groups (int): The number of groups in the ResNet block. Defaults to 32.
71
+ resnet_pre_norm (bool): Whether to use pre-normalization in the ResNet block. Defaults to True.
72
+ output_scale_factor (float): The scale factor to apply to the output. Defaults to 1.0.
73
+
74
+ Returns:
75
+ nn.Module: The created UpBlock2D or CrossAttnUpBlock2D object.
76
+ """
77
+ # If attn head dim is not defined, we default it to the number of heads
78
+ if attention_head_dim is None:
79
+ logger.warning("It is recommended to provide `attention_head_dim` when calling `get_down_block`.")
80
+ logger.warning(f"Defaulting `attention_head_dim` to {num_attention_heads}.")
81
+ attention_head_dim = num_attention_heads
82
+
83
+ down_block_type = (
84
+ down_block_type[7:]
85
+ if down_block_type.startswith("UNetRes")
86
+ else down_block_type
87
+ )
88
+ if down_block_type == "DownBlock2D":
89
+ return DownBlock2D(
90
+ num_layers=num_layers,
91
+ in_channels=in_channels,
92
+ out_channels=out_channels,
93
+ temb_channels=temb_channels,
94
+ dropout=dropout,
95
+ add_downsample=add_downsample,
96
+ resnet_eps=resnet_eps,
97
+ resnet_act_fn=resnet_act_fn,
98
+ resnet_groups=resnet_groups,
99
+ downsample_padding=downsample_padding,
100
+ resnet_time_scale_shift=resnet_time_scale_shift,
101
+ )
102
+
103
+ if down_block_type == "CrossAttnDownBlock2D":
104
+ if cross_attention_dim is None:
105
+ raise ValueError(
106
+ "cross_attention_dim must be specified for CrossAttnDownBlock2D"
107
+ )
108
+ return CrossAttnDownBlock2D(
109
+ num_layers=num_layers,
110
+ transformer_layers_per_block=transformer_layers_per_block,
111
+ in_channels=in_channels,
112
+ out_channels=out_channels,
113
+ temb_channels=temb_channels,
114
+ dropout=dropout,
115
+ add_downsample=add_downsample,
116
+ resnet_eps=resnet_eps,
117
+ resnet_act_fn=resnet_act_fn,
118
+ resnet_groups=resnet_groups,
119
+ downsample_padding=downsample_padding,
120
+ cross_attention_dim=cross_attention_dim,
121
+ num_attention_heads=num_attention_heads,
122
+ dual_cross_attention=dual_cross_attention,
123
+ use_linear_projection=use_linear_projection,
124
+ only_cross_attention=only_cross_attention,
125
+ upcast_attention=upcast_attention,
126
+ resnet_time_scale_shift=resnet_time_scale_shift,
127
+ attention_type=attention_type,
128
+ )
129
+ raise ValueError(f"{down_block_type} does not exist.")
130
+
131
+
132
+ def get_up_block(
133
+ up_block_type: str,
134
+ num_layers: int,
135
+ in_channels: int,
136
+ out_channels: int,
137
+ prev_output_channel: int,
138
+ temb_channels: int,
139
+ add_upsample: bool,
140
+ resnet_eps: float,
141
+ resnet_act_fn: str,
142
+ resolution_idx: Optional[int] = None,
143
+ transformer_layers_per_block: int = 1,
144
+ num_attention_heads: Optional[int] = None,
145
+ resnet_groups: Optional[int] = None,
146
+ cross_attention_dim: Optional[int] = None,
147
+ dual_cross_attention: bool = False,
148
+ use_linear_projection: bool = False,
149
+ only_cross_attention: bool = False,
150
+ upcast_attention: bool = False,
151
+ resnet_time_scale_shift: str = "default",
152
+ attention_type: str = "default",
153
+ attention_head_dim: Optional[int] = None,
154
+ dropout: float = 0.0,
155
+ ) -> nn.Module:
156
+ """ This function ...
157
+ Args:
158
+ Returns:
159
+ """
160
+ # If attn head dim is not defined, we default it to the number of heads
161
+ if attention_head_dim is None:
162
+ logger.warning("It is recommended to provide `attention_head_dim` when calling `get_up_block`.")
163
+ logger.warning(f"Defaulting `attention_head_dim` to {num_attention_heads}.")
164
+ attention_head_dim = num_attention_heads
165
+
166
+ up_block_type = (
167
+ up_block_type[7:] if up_block_type.startswith("UNetRes") else up_block_type
168
+ )
169
+ if up_block_type == "UpBlock2D":
170
+ return UpBlock2D(
171
+ num_layers=num_layers,
172
+ in_channels=in_channels,
173
+ out_channels=out_channels,
174
+ prev_output_channel=prev_output_channel,
175
+ temb_channels=temb_channels,
176
+ resolution_idx=resolution_idx,
177
+ dropout=dropout,
178
+ add_upsample=add_upsample,
179
+ resnet_eps=resnet_eps,
180
+ resnet_act_fn=resnet_act_fn,
181
+ resnet_groups=resnet_groups,
182
+ resnet_time_scale_shift=resnet_time_scale_shift,
183
+ )
184
+ if up_block_type == "CrossAttnUpBlock2D":
185
+ if cross_attention_dim is None:
186
+ raise ValueError(
187
+ "cross_attention_dim must be specified for CrossAttnUpBlock2D"
188
+ )
189
+ return CrossAttnUpBlock2D(
190
+ num_layers=num_layers,
191
+ transformer_layers_per_block=transformer_layers_per_block,
192
+ in_channels=in_channels,
193
+ out_channels=out_channels,
194
+ prev_output_channel=prev_output_channel,
195
+ temb_channels=temb_channels,
196
+ resolution_idx=resolution_idx,
197
+ dropout=dropout,
198
+ add_upsample=add_upsample,
199
+ resnet_eps=resnet_eps,
200
+ resnet_act_fn=resnet_act_fn,
201
+ resnet_groups=resnet_groups,
202
+ cross_attention_dim=cross_attention_dim,
203
+ num_attention_heads=num_attention_heads,
204
+ dual_cross_attention=dual_cross_attention,
205
+ use_linear_projection=use_linear_projection,
206
+ only_cross_attention=only_cross_attention,
207
+ upcast_attention=upcast_attention,
208
+ resnet_time_scale_shift=resnet_time_scale_shift,
209
+ attention_type=attention_type,
210
+ )
211
+
212
+ raise ValueError(f"{up_block_type} does not exist.")
213
+
214
+
215
+ class AutoencoderTinyBlock(nn.Module):
216
+ """
217
+ Tiny Autoencoder block used in [`AutoencoderTiny`]. It is a mini residual module consisting of plain conv + ReLU
218
+ blocks.
219
+
220
+ Args:
221
+ in_channels (`int`): The number of input channels.
222
+ out_channels (`int`): The number of output channels.
223
+ act_fn (`str`):
224
+ ` The activation function to use. Supported values are `"swish"`, `"mish"`, `"gelu"`, and `"relu"`.
225
+
226
+ Returns:
227
+ `torch.FloatTensor`: A tensor with the same shape as the input tensor, but with the number of channels equal to
228
+ `out_channels`.
229
+ """
230
+
231
+ def __init__(self, in_channels: int, out_channels: int, act_fn: str):
232
+ super().__init__()
233
+ act_fn = get_activation(act_fn)
234
+ self.conv = nn.Sequential(
235
+ nn.Conv2d(in_channels, out_channels, kernel_size=3, padding=1),
236
+ act_fn,
237
+ nn.Conv2d(out_channels, out_channels, kernel_size=3, padding=1),
238
+ act_fn,
239
+ nn.Conv2d(out_channels, out_channels, kernel_size=3, padding=1),
240
+ )
241
+ self.skip = (
242
+ nn.Conv2d(in_channels, out_channels, kernel_size=1, bias=False)
243
+ if in_channels != out_channels
244
+ else nn.Identity()
245
+ )
246
+ self.fuse = nn.ReLU()
247
+
248
+ def forward(self, x: torch.FloatTensor) -> torch.FloatTensor:
249
+ """
250
+ Forward pass of the AutoencoderTinyBlock class.
251
+
252
+ Parameters:
253
+ x (torch.FloatTensor): The input tensor to the AutoencoderTinyBlock.
254
+
255
+ Returns:
256
+ torch.FloatTensor: The output tensor after passing through the AutoencoderTinyBlock.
257
+ """
258
+ return self.fuse(self.conv(x) + self.skip(x))
259
+
260
+
261
+ class UNetMidBlock2D(nn.Module):
262
+ """
263
+ A 2D UNet mid-block [`UNetMidBlock2D`] with multiple residual blocks and optional attention blocks.
264
+
265
+ Args:
266
+ in_channels (`int`): The number of input channels.
267
+ temb_channels (`int`): The number of temporal embedding channels.
268
+ dropout (`float`, *optional*, defaults to 0.0): The dropout rate.
269
+ num_layers (`int`, *optional*, defaults to 1): The number of residual blocks.
270
+ resnet_eps (`float`, *optional*, 1e-6 ): The epsilon value for the resnet blocks.
271
+ resnet_time_scale_shift (`str`, *optional*, defaults to `default`):
272
+ The type of normalization to apply to the time embeddings. This can help to improve the performance of the
273
+ model on tasks with long-range temporal dependencies.
274
+ resnet_act_fn (`str`, *optional*, defaults to `swish`): The activation function for the resnet blocks.
275
+ resnet_groups (`int`, *optional*, defaults to 32):
276
+ The number of groups to use in the group normalization layers of the resnet blocks.
277
+ attn_groups (`Optional[int]`, *optional*, defaults to None): The number of groups for the attention blocks.
278
+ resnet_pre_norm (`bool`, *optional*, defaults to `True`):
279
+ Whether to use pre-normalization for the resnet blocks.
280
+ add_attention (`bool`, *optional*, defaults to `True`): Whether to add attention blocks.
281
+ attention_head_dim (`int`, *optional*, defaults to 1):
282
+ Dimension of a single attention head. The number of attention heads is determined based on this value and
283
+ the number of input channels.
284
+ output_scale_factor (`float`, *optional*, defaults to 1.0): The output scale factor.
285
+
286
+ Returns:
287
+ `torch.FloatTensor`: The output of the last residual block, which is a tensor of shape `(batch_size,
288
+ in_channels, height, width)`.
289
+
290
+ """
291
+
292
+ def __init__(
293
+ self,
294
+ in_channels: int,
295
+ temb_channels: int,
296
+ dropout: float = 0.0,
297
+ num_layers: int = 1,
298
+ resnet_eps: float = 1e-6,
299
+ resnet_time_scale_shift: str = "default", # default, spatial
300
+ resnet_act_fn: str = "swish",
301
+ resnet_groups: int = 32,
302
+ attn_groups: Optional[int] = None,
303
+ resnet_pre_norm: bool = True,
304
+ add_attention: bool = True,
305
+ attention_head_dim: int = 1,
306
+ output_scale_factor: float = 1.0,
307
+ ):
308
+ super().__init__()
309
+ resnet_groups = (
310
+ resnet_groups if resnet_groups is not None else min(in_channels // 4, 32)
311
+ )
312
+ self.add_attention = add_attention
313
+
314
+ if attn_groups is None:
315
+ attn_groups = (
316
+ resnet_groups if resnet_time_scale_shift == "default" else None
317
+ )
318
+
319
+ # there is always at least one resnet
320
+ resnets = [
321
+ ResnetBlock2D(
322
+ in_channels=in_channels,
323
+ out_channels=in_channels,
324
+ temb_channels=temb_channels,
325
+ eps=resnet_eps,
326
+ groups=resnet_groups,
327
+ dropout=dropout,
328
+ time_embedding_norm=resnet_time_scale_shift,
329
+ non_linearity=resnet_act_fn,
330
+ output_scale_factor=output_scale_factor,
331
+ pre_norm=resnet_pre_norm,
332
+ )
333
+ ]
334
+ attentions = []
335
+
336
+ if attention_head_dim is None:
337
+ logger.warning(
338
+ f"It is not recommend to pass `attention_head_dim=None`. Defaulting `attention_head_dim` to `in_channels`: {in_channels}."
339
+ )
340
+ attention_head_dim = in_channels
341
+
342
+ for _ in range(num_layers):
343
+ if self.add_attention:
344
+ attentions.append(
345
+ Attention(
346
+ in_channels,
347
+ heads=in_channels // attention_head_dim,
348
+ dim_head=attention_head_dim,
349
+ rescale_output_factor=output_scale_factor,
350
+ eps=resnet_eps,
351
+ norm_num_groups=attn_groups,
352
+ spatial_norm_dim=(
353
+ temb_channels
354
+ if resnet_time_scale_shift == "spatial"
355
+ else None
356
+ ),
357
+ residual_connection=True,
358
+ bias=True,
359
+ upcast_softmax=True,
360
+ _from_deprecated_attn_block=True,
361
+ )
362
+ )
363
+ else:
364
+ attentions.append(None)
365
+
366
+ resnets.append(
367
+ ResnetBlock2D(
368
+ in_channels=in_channels,
369
+ out_channels=in_channels,
370
+ temb_channels=temb_channels,
371
+ eps=resnet_eps,
372
+ groups=resnet_groups,
373
+ dropout=dropout,
374
+ time_embedding_norm=resnet_time_scale_shift,
375
+ non_linearity=resnet_act_fn,
376
+ output_scale_factor=output_scale_factor,
377
+ pre_norm=resnet_pre_norm,
378
+ )
379
+ )
380
+
381
+ self.attentions = nn.ModuleList(attentions)
382
+ self.resnets = nn.ModuleList(resnets)
383
+
384
+ def forward(
385
+ self, hidden_states: torch.FloatTensor, temb: Optional[torch.FloatTensor] = None
386
+ ) -> torch.FloatTensor:
387
+ """
388
+ Forward pass of the UNetMidBlock2D class.
389
+
390
+ Args:
391
+ hidden_states (torch.FloatTensor): The input tensor to the UNetMidBlock2D.
392
+ temb (Optional[torch.FloatTensor], optional): The token embedding tensor. Defaults to None.
393
+
394
+ Returns:
395
+ torch.FloatTensor: The output tensor after passing through the UNetMidBlock2D.
396
+ """
397
+ # Your implementation here
398
+ hidden_states = self.resnets[0](hidden_states, temb)
399
+ for attn, resnet in zip(self.attentions, self.resnets[1:]):
400
+ if attn is not None:
401
+ hidden_states = attn(hidden_states, temb=temb)
402
+ hidden_states = resnet(hidden_states, temb)
403
+
404
+ return hidden_states
405
+
406
+
407
+ class UNetMidBlock2DCrossAttn(nn.Module):
408
+ """
409
+ UNetMidBlock2DCrossAttn is a class that represents a mid-block 2D UNet with cross-attention.
410
+
411
+ This block is responsible for processing the input tensor with a series of residual blocks,
412
+ and applying cross-attention mechanism to attend to the global information in the encoder.
413
+
414
+ Args:
415
+ in_channels (int): The number of input channels.
416
+ temb_channels (int): The number of channels for the token embedding.
417
+ dropout (float, optional): The dropout rate. Defaults to 0.0.
418
+ num_layers (int, optional): The number of layers in the residual blocks. Defaults to 1.
419
+ resnet_eps (float, optional): The epsilon value for the residual blocks. Defaults to 1e-6.
420
+ resnet_time_scale_shift (str, optional): The time scale shift type for the residual blocks. Defaults to "default".
421
+ resnet_act_fn (str, optional): The activation function for the residual blocks. Defaults to "swish".
422
+ resnet_groups (int, optional): The number of groups for the residual blocks. Defaults to 32.
423
+ resnet_pre_norm (bool, optional): Whether to apply pre-normalization for the residual blocks. Defaults to True.
424
+ num_attention_heads (int, optional): The number of attention heads for cross-attention. Defaults to 1.
425
+ cross_attention_dim (int, optional): The dimension of the cross-attention. Defaults to 1280.
426
+ output_scale_factor (float, optional): The scale factor for the output tensor. Defaults to 1.0.
427
+ """
428
+ def __init__(
429
+ self,
430
+ in_channels: int,
431
+ temb_channels: int,
432
+ dropout: float = 0.0,
433
+ num_layers: int = 1,
434
+ transformer_layers_per_block: Union[int, Tuple[int]] = 1,
435
+ resnet_eps: float = 1e-6,
436
+ resnet_time_scale_shift: str = "default",
437
+ resnet_act_fn: str = "swish",
438
+ resnet_groups: int = 32,
439
+ resnet_pre_norm: bool = True,
440
+ num_attention_heads: int = 1,
441
+ output_scale_factor: float = 1.0,
442
+ cross_attention_dim: int = 1280,
443
+ dual_cross_attention: bool = False,
444
+ use_linear_projection: bool = False,
445
+ upcast_attention: bool = False,
446
+ attention_type: str = "default",
447
+ ):
448
+ super().__init__()
449
+
450
+ self.has_cross_attention = True
451
+ self.num_attention_heads = num_attention_heads
452
+ resnet_groups = (
453
+ resnet_groups if resnet_groups is not None else min(in_channels // 4, 32)
454
+ )
455
+
456
+ # support for variable transformer layers per block
457
+ if isinstance(transformer_layers_per_block, int):
458
+ transformer_layers_per_block = [transformer_layers_per_block] * num_layers
459
+
460
+ # there is always at least one resnet
461
+ resnets = [
462
+ ResnetBlock2D(
463
+ in_channels=in_channels,
464
+ out_channels=in_channels,
465
+ temb_channels=temb_channels,
466
+ eps=resnet_eps,
467
+ groups=resnet_groups,
468
+ dropout=dropout,
469
+ time_embedding_norm=resnet_time_scale_shift,
470
+ non_linearity=resnet_act_fn,
471
+ output_scale_factor=output_scale_factor,
472
+ pre_norm=resnet_pre_norm,
473
+ )
474
+ ]
475
+ attentions = []
476
+
477
+ for i in range(num_layers):
478
+ if not dual_cross_attention:
479
+ attentions.append(
480
+ Transformer2DModel(
481
+ num_attention_heads,
482
+ in_channels // num_attention_heads,
483
+ in_channels=in_channels,
484
+ num_layers=transformer_layers_per_block[i],
485
+ cross_attention_dim=cross_attention_dim,
486
+ norm_num_groups=resnet_groups,
487
+ use_linear_projection=use_linear_projection,
488
+ upcast_attention=upcast_attention,
489
+ attention_type=attention_type,
490
+ )
491
+ )
492
+ else:
493
+ attentions.append(
494
+ DualTransformer2DModel(
495
+ num_attention_heads,
496
+ in_channels // num_attention_heads,
497
+ in_channels=in_channels,
498
+ num_layers=1,
499
+ cross_attention_dim=cross_attention_dim,
500
+ norm_num_groups=resnet_groups,
501
+ )
502
+ )
503
+ resnets.append(
504
+ ResnetBlock2D(
505
+ in_channels=in_channels,
506
+ out_channels=in_channels,
507
+ temb_channels=temb_channels,
508
+ eps=resnet_eps,
509
+ groups=resnet_groups,
510
+ dropout=dropout,
511
+ time_embedding_norm=resnet_time_scale_shift,
512
+ non_linearity=resnet_act_fn,
513
+ output_scale_factor=output_scale_factor,
514
+ pre_norm=resnet_pre_norm,
515
+ )
516
+ )
517
+
518
+ self.attentions = nn.ModuleList(attentions)
519
+ self.resnets = nn.ModuleList(resnets)
520
+
521
+ self.gradient_checkpointing = False
522
+
523
+ def forward(
524
+ self,
525
+ hidden_states: torch.FloatTensor,
526
+ temb: Optional[torch.FloatTensor] = None,
527
+ encoder_hidden_states: Optional[torch.FloatTensor] = None,
528
+ attention_mask: Optional[torch.FloatTensor] = None,
529
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
530
+ encoder_attention_mask: Optional[torch.FloatTensor] = None,
531
+ ) -> torch.FloatTensor:
532
+ """
533
+ Forward pass for the UNetMidBlock2DCrossAttn class.
534
+
535
+ Args:
536
+ hidden_states (torch.FloatTensor): The input hidden states tensor.
537
+ temb (Optional[torch.FloatTensor], optional): The optional tensor for time embeddings.
538
+ encoder_hidden_states (Optional[torch.FloatTensor], optional): The optional encoder hidden states tensor.
539
+ attention_mask (Optional[torch.FloatTensor], optional): The optional attention mask tensor.
540
+ cross_attention_kwargs (Optional[Dict[str, Any]], optional): The optional cross-attention kwargs tensor.
541
+ encoder_attention_mask (Optional[torch.FloatTensor], optional): The optional encoder attention mask tensor.
542
+
543
+ Returns:
544
+ torch.FloatTensor: The output tensor after passing through the UNetMidBlock2DCrossAttn layers.
545
+ """
546
+ lora_scale = (
547
+ cross_attention_kwargs.get("scale", 1.0)
548
+ if cross_attention_kwargs is not None
549
+ else 1.0
550
+ )
551
+ hidden_states = self.resnets[0](hidden_states, temb, scale=lora_scale)
552
+ for attn, resnet in zip(self.attentions, self.resnets[1:]):
553
+ if self.training and self.gradient_checkpointing:
554
+
555
+ def create_custom_forward(module, return_dict=None):
556
+ def custom_forward(*inputs):
557
+ if return_dict is not None:
558
+ return module(*inputs, return_dict=return_dict)
559
+
560
+ return module(*inputs)
561
+
562
+ return custom_forward
563
+
564
+ ckpt_kwargs: Dict[str, Any] = (
565
+ {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
566
+ )
567
+ hidden_states, _ref_feature = attn(
568
+ hidden_states,
569
+ encoder_hidden_states=encoder_hidden_states,
570
+ cross_attention_kwargs=cross_attention_kwargs,
571
+ attention_mask=attention_mask,
572
+ encoder_attention_mask=encoder_attention_mask,
573
+ return_dict=False,
574
+ )
575
+ hidden_states = torch.utils.checkpoint.checkpoint(
576
+ create_custom_forward(resnet),
577
+ hidden_states,
578
+ temb,
579
+ **ckpt_kwargs,
580
+ )
581
+ else:
582
+ hidden_states, _ref_feature = attn(
583
+ hidden_states,
584
+ encoder_hidden_states=encoder_hidden_states,
585
+ cross_attention_kwargs=cross_attention_kwargs,
586
+ attention_mask=attention_mask,
587
+ encoder_attention_mask=encoder_attention_mask,
588
+ return_dict=False,
589
+ )
590
+ hidden_states = resnet(hidden_states, temb, scale=lora_scale)
591
+
592
+ return hidden_states
593
+
594
+
595
+ class CrossAttnDownBlock2D(nn.Module):
596
+ """
597
+ CrossAttnDownBlock2D is a class that represents a 2D cross-attention downsampling block.
598
+
599
+ This block is used in the UNet model and consists of a series of ResNet blocks and Transformer layers.
600
+ It takes input hidden states, a tensor embedding, and optional encoder hidden states, attention mask,
601
+ and cross-attention kwargs. The block performs a series of operations including downsampling, cross-attention,
602
+ and residual connections.
603
+
604
+ Attributes:
605
+ in_channels (int): The number of input channels.
606
+ out_channels (int): The number of output channels.
607
+ temb_channels (int): The number of tensor embedding channels.
608
+ dropout (float): The dropout rate.
609
+ num_layers (int): The number of ResNet layers.
610
+ transformer_layers_per_block (Union[int, Tuple[int]]): The number of Transformer layers per block.
611
+ resnet_eps (float): The ResNet epsilon value.
612
+ resnet_time_scale_shift (str): The ResNet time scale shift type.
613
+ resnet_act_fn (str): The ResNet activation function.
614
+ resnet_groups (int): The ResNet group size.
615
+ resnet_pre_norm (bool): Whether to use ResNet pre-normalization.
616
+ num_attention_heads (int): The number of attention heads.
617
+ cross_attention_dim (int): The cross-attention dimension.
618
+ output_scale_factor (float): The output scale factor.
619
+ downsample_padding (int): The downsampling padding.
620
+ add_downsample (bool): Whether to add downsampling.
621
+ dual_cross_attention (bool): Whether to use dual cross-attention.
622
+ use_linear_projection (bool): Whether to use linear projection.
623
+ only_cross_attention (bool): Whether to use only cross-attention.
624
+ upcast_attention (bool): Whether to upcast attention.
625
+ attention_type (str): The attention type.
626
+ """
627
+ def __init__(
628
+ self,
629
+ in_channels: int,
630
+ out_channels: int,
631
+ temb_channels: int,
632
+ dropout: float = 0.0,
633
+ num_layers: int = 1,
634
+ transformer_layers_per_block: Union[int, Tuple[int]] = 1,
635
+ resnet_eps: float = 1e-6,
636
+ resnet_time_scale_shift: str = "default",
637
+ resnet_act_fn: str = "swish",
638
+ resnet_groups: int = 32,
639
+ resnet_pre_norm: bool = True,
640
+ num_attention_heads: int = 1,
641
+ cross_attention_dim: int = 1280,
642
+ output_scale_factor: float = 1.0,
643
+ downsample_padding: int = 1,
644
+ add_downsample: bool = True,
645
+ dual_cross_attention: bool = False,
646
+ use_linear_projection: bool = False,
647
+ only_cross_attention: bool = False,
648
+ upcast_attention: bool = False,
649
+ attention_type: str = "default",
650
+ ):
651
+ super().__init__()
652
+ resnets = []
653
+ attentions = []
654
+
655
+ self.has_cross_attention = True
656
+ self.num_attention_heads = num_attention_heads
657
+ if isinstance(transformer_layers_per_block, int):
658
+ transformer_layers_per_block = [transformer_layers_per_block] * num_layers
659
+
660
+ for i in range(num_layers):
661
+ in_channels = in_channels if i == 0 else out_channels
662
+ resnets.append(
663
+ ResnetBlock2D(
664
+ in_channels=in_channels,
665
+ out_channels=out_channels,
666
+ temb_channels=temb_channels,
667
+ eps=resnet_eps,
668
+ groups=resnet_groups,
669
+ dropout=dropout,
670
+ time_embedding_norm=resnet_time_scale_shift,
671
+ non_linearity=resnet_act_fn,
672
+ output_scale_factor=output_scale_factor,
673
+ pre_norm=resnet_pre_norm,
674
+ )
675
+ )
676
+ if not dual_cross_attention:
677
+ attentions.append(
678
+ Transformer2DModel(
679
+ num_attention_heads,
680
+ out_channels // num_attention_heads,
681
+ in_channels=out_channels,
682
+ num_layers=transformer_layers_per_block[i],
683
+ cross_attention_dim=cross_attention_dim,
684
+ norm_num_groups=resnet_groups,
685
+ use_linear_projection=use_linear_projection,
686
+ only_cross_attention=only_cross_attention,
687
+ upcast_attention=upcast_attention,
688
+ attention_type=attention_type,
689
+ )
690
+ )
691
+ else:
692
+ attentions.append(
693
+ DualTransformer2DModel(
694
+ num_attention_heads,
695
+ out_channels // num_attention_heads,
696
+ in_channels=out_channels,
697
+ num_layers=1,
698
+ cross_attention_dim=cross_attention_dim,
699
+ norm_num_groups=resnet_groups,
700
+ )
701
+ )
702
+ self.attentions = nn.ModuleList(attentions)
703
+ self.resnets = nn.ModuleList(resnets)
704
+
705
+ if add_downsample:
706
+ self.downsamplers = nn.ModuleList(
707
+ [
708
+ Downsample2D(
709
+ out_channels,
710
+ use_conv=True,
711
+ out_channels=out_channels,
712
+ padding=downsample_padding,
713
+ name="op",
714
+ )
715
+ ]
716
+ )
717
+ else:
718
+ self.downsamplers = None
719
+
720
+ self.gradient_checkpointing = False
721
+
722
+ def forward(
723
+ self,
724
+ hidden_states: torch.FloatTensor,
725
+ temb: Optional[torch.FloatTensor] = None,
726
+ encoder_hidden_states: Optional[torch.FloatTensor] = None,
727
+ attention_mask: Optional[torch.FloatTensor] = None,
728
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
729
+ encoder_attention_mask: Optional[torch.FloatTensor] = None,
730
+ additional_residuals: Optional[torch.FloatTensor] = None,
731
+ ) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...]]:
732
+ """
733
+ Forward pass for the CrossAttnDownBlock2D class.
734
+
735
+ Args:
736
+ hidden_states (torch.FloatTensor): The input hidden states.
737
+ temb (Optional[torch.FloatTensor], optional): The token embeddings. Defaults to None.
738
+ encoder_hidden_states (Optional[torch.FloatTensor], optional): The encoder hidden states. Defaults to None.
739
+ attention_mask (Optional[torch.FloatTensor], optional): The attention mask. Defaults to None.
740
+ cross_attention_kwargs (Optional[Dict[str, Any]], optional): The cross-attention kwargs. Defaults to None.
741
+ encoder_attention_mask (Optional[torch.FloatTensor], optional): The encoder attention mask. Defaults to None.
742
+ additional_residuals (Optional[torch.FloatTensor], optional): The additional residuals. Defaults to None.
743
+
744
+ Returns:
745
+ Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...]]: The output hidden states and residuals.
746
+ """
747
+ output_states = ()
748
+
749
+ lora_scale = (
750
+ cross_attention_kwargs.get("scale", 1.0)
751
+ if cross_attention_kwargs is not None
752
+ else 1.0
753
+ )
754
+
755
+ blocks = list(zip(self.resnets, self.attentions))
756
+
757
+ for i, (resnet, attn) in enumerate(blocks):
758
+ if self.training and self.gradient_checkpointing:
759
+
760
+ def create_custom_forward(module, return_dict=None):
761
+ def custom_forward(*inputs):
762
+ if return_dict is not None:
763
+ return module(*inputs, return_dict=return_dict)
764
+
765
+ return module(*inputs)
766
+
767
+ return custom_forward
768
+
769
+ ckpt_kwargs: Dict[str, Any] = (
770
+ {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
771
+ )
772
+ hidden_states = torch.utils.checkpoint.checkpoint(
773
+ create_custom_forward(resnet),
774
+ hidden_states,
775
+ temb,
776
+ **ckpt_kwargs,
777
+ )
778
+ hidden_states, _ref_feature = attn(
779
+ hidden_states,
780
+ encoder_hidden_states=encoder_hidden_states,
781
+ cross_attention_kwargs=cross_attention_kwargs,
782
+ attention_mask=attention_mask,
783
+ encoder_attention_mask=encoder_attention_mask,
784
+ return_dict=False,
785
+ )
786
+ else:
787
+ hidden_states = resnet(hidden_states, temb, scale=lora_scale)
788
+ hidden_states, _ref_feature = attn(
789
+ hidden_states,
790
+ encoder_hidden_states=encoder_hidden_states,
791
+ cross_attention_kwargs=cross_attention_kwargs,
792
+ attention_mask=attention_mask,
793
+ encoder_attention_mask=encoder_attention_mask,
794
+ return_dict=False,
795
+ )
796
+
797
+ # apply additional residuals to the output of the last pair of resnet and attention blocks
798
+ if i == len(blocks) - 1 and additional_residuals is not None:
799
+ hidden_states = hidden_states + additional_residuals
800
+
801
+ output_states = output_states + (hidden_states,)
802
+
803
+ if self.downsamplers is not None:
804
+ for downsampler in self.downsamplers:
805
+ hidden_states = downsampler(hidden_states, scale=lora_scale)
806
+
807
+ output_states = output_states + (hidden_states,)
808
+
809
+ return hidden_states, output_states
810
+
811
+
812
+ class DownBlock2D(nn.Module):
813
+ """
814
+ DownBlock2D is a class that represents a 2D downsampling block in a neural network.
815
+
816
+ It takes the following parameters:
817
+ - in_channels (int): The number of input channels in the block.
818
+ - out_channels (int): The number of output channels in the block.
819
+ - temb_channels (int): The number of channels in the token embedding.
820
+ - dropout (float): The dropout rate for the block.
821
+ - num_layers (int): The number of layers in the block.
822
+ - resnet_eps (float): The epsilon value for the ResNet layer.
823
+ - resnet_time_scale_shift (str): The type of activation function for the ResNet layer.
824
+ - resnet_act_fn (str): The activation function for the ResNet layer.
825
+ - resnet_groups (int): The number of groups in the ResNet layer.
826
+ - resnet_pre_norm (bool): Whether to apply layer normalization before the ResNet layer.
827
+ - output_scale_factor (float): The scale factor for the output.
828
+ - add_downsample (bool): Whether to add a downsampling layer.
829
+ - downsample_padding (int): The padding value for the downsampling layer.
830
+
831
+ The DownBlock2D class inherits from the nn.Module class and defines the following methods:
832
+ - __init__: Initializes the DownBlock2D class with the given parameters.
833
+ - forward: Forward pass of the DownBlock2D class.
834
+
835
+ The forward method takes the following parameters:
836
+ - hidden_states (torch.FloatTensor): The input tensor to the block.
837
+ - temb (Optional[torch.FloatTensor]): The token embedding tensor.
838
+ - scale (float): The scale factor for the input tensor.
839
+
840
+ The forward method returns a tuple containing the output tensor and a tuple of hidden states.
841
+ """
842
+ def __init__(
843
+ self,
844
+ in_channels: int,
845
+ out_channels: int,
846
+ temb_channels: int,
847
+ dropout: float = 0.0,
848
+ num_layers: int = 1,
849
+ resnet_eps: float = 1e-6,
850
+ resnet_time_scale_shift: str = "default",
851
+ resnet_act_fn: str = "swish",
852
+ resnet_groups: int = 32,
853
+ resnet_pre_norm: bool = True,
854
+ output_scale_factor: float = 1.0,
855
+ add_downsample: bool = True,
856
+ downsample_padding: int = 1,
857
+ ):
858
+ super().__init__()
859
+ resnets = []
860
+
861
+ for i in range(num_layers):
862
+ in_channels = in_channels if i == 0 else out_channels
863
+ resnets.append(
864
+ ResnetBlock2D(
865
+ in_channels=in_channels,
866
+ out_channels=out_channels,
867
+ temb_channels=temb_channels,
868
+ eps=resnet_eps,
869
+ groups=resnet_groups,
870
+ dropout=dropout,
871
+ time_embedding_norm=resnet_time_scale_shift,
872
+ non_linearity=resnet_act_fn,
873
+ output_scale_factor=output_scale_factor,
874
+ pre_norm=resnet_pre_norm,
875
+ )
876
+ )
877
+
878
+ self.resnets = nn.ModuleList(resnets)
879
+
880
+ if add_downsample:
881
+ self.downsamplers = nn.ModuleList(
882
+ [
883
+ Downsample2D(
884
+ out_channels,
885
+ use_conv=True,
886
+ out_channels=out_channels,
887
+ padding=downsample_padding,
888
+ name="op",
889
+ )
890
+ ]
891
+ )
892
+ else:
893
+ self.downsamplers = None
894
+
895
+ self.gradient_checkpointing = False
896
+
897
+ def forward(
898
+ self,
899
+ hidden_states: torch.FloatTensor,
900
+ temb: Optional[torch.FloatTensor] = None,
901
+ scale: float = 1.0,
902
+ ) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...]]:
903
+ """
904
+ Forward pass of the DownBlock2D class.
905
+
906
+ Args:
907
+ hidden_states (torch.FloatTensor): The input tensor to the DownBlock2D layer.
908
+ temb (Optional[torch.FloatTensor], optional): The token embedding tensor. Defaults to None.
909
+ scale (float, optional): The scale factor for the input tensor. Defaults to 1.0.
910
+
911
+ Returns:
912
+ Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...]]: The output tensor and any additional hidden states.
913
+ """
914
+ output_states = ()
915
+
916
+ for resnet in self.resnets:
917
+ if self.training and self.gradient_checkpointing:
918
+
919
+ def create_custom_forward(module):
920
+ def custom_forward(*inputs):
921
+ return module(*inputs)
922
+
923
+ return custom_forward
924
+
925
+ if is_torch_version(">=", "1.11.0"):
926
+ hidden_states = torch.utils.checkpoint.checkpoint(
927
+ create_custom_forward(resnet),
928
+ hidden_states,
929
+ temb,
930
+ use_reentrant=False,
931
+ )
932
+ else:
933
+ hidden_states = torch.utils.checkpoint.checkpoint(
934
+ create_custom_forward(resnet), hidden_states, temb
935
+ )
936
+ else:
937
+ hidden_states = resnet(hidden_states, temb, scale=scale)
938
+
939
+ output_states = output_states + (hidden_states,)
940
+
941
+ if self.downsamplers is not None:
942
+ for downsampler in self.downsamplers:
943
+ hidden_states = downsampler(hidden_states, scale=scale)
944
+
945
+ output_states = output_states + (hidden_states,)
946
+
947
+ return hidden_states, output_states
948
+
949
+
950
+ class CrossAttnUpBlock2D(nn.Module):
951
+ """
952
+ CrossAttnUpBlock2D is a class that represents a cross-attention UpBlock in a 2D UNet architecture.
953
+
954
+ This block is responsible for upsampling the input tensor and performing cross-attention with the encoder's hidden states.
955
+
956
+ Args:
957
+ in_channels (int): The number of input channels in the tensor.
958
+ out_channels (int): The number of output channels in the tensor.
959
+ prev_output_channel (int): The number of channels in the previous output tensor.
960
+ temb_channels (int): The number of channels in the token embedding tensor.
961
+ resolution_idx (Optional[int]): The index of the resolution in the model.
962
+ dropout (float): The dropout rate for the layer.
963
+ num_layers (int): The number of layers in the ResNet block.
964
+ transformer_layers_per_block (Union[int, Tuple[int]]): The number of transformer layers per block.
965
+ resnet_eps (float): The epsilon value for the ResNet layer.
966
+ resnet_time_scale_shift (str): The type of time scale shift to be applied in the ResNet layer.
967
+ resnet_act_fn (str): The activation function to be used in the ResNet layer.
968
+ resnet_groups (int): The number of groups in the ResNet layer.
969
+ resnet_pre_norm (bool): Whether to use pre-normalization in the ResNet layer.
970
+ num_attention_heads (int): The number of attention heads in the cross-attention layer.
971
+ cross_attention_dim (int): The dimension of the cross-attention layer.
972
+ output_scale_factor (float): The scale factor for the output tensor.
973
+ add_upsample (bool): Whether to add upsampling to the block.
974
+ dual_cross_attention (bool): Whether to use dual cross-attention.
975
+ use_linear_projection (bool): Whether to use linear projection in the cross-attention layer.
976
+ only_cross_attention (bool): Whether to only use cross-attention and no self-attention.
977
+ upcast_attention (bool): Whether to upcast the attention weights.
978
+ attention_type (str): The type of attention to be used in the cross-attention layer.
979
+
980
+ Attributes:
981
+ up_block (nn.Module): The UpBlock module responsible for upsampling the input tensor.
982
+ cross_attn (nn.Module): The cross-attention module that performs attention between
983
+ the decoder's hidden states and the encoder's hidden states.
984
+ resnet_blocks (nn.ModuleList): A list of ResNet blocks that make up the ResNet portion of the block.
985
+ """
986
+
987
+ def __init__(
988
+ self,
989
+ in_channels: int,
990
+ out_channels: int,
991
+ prev_output_channel: int,
992
+ temb_channels: int,
993
+ resolution_idx: Optional[int] = None,
994
+ dropout: float = 0.0,
995
+ num_layers: int = 1,
996
+ transformer_layers_per_block: Union[int, Tuple[int]] = 1,
997
+ resnet_eps: float = 1e-6,
998
+ resnet_time_scale_shift: str = "default",
999
+ resnet_act_fn: str = "swish",
1000
+ resnet_groups: int = 32,
1001
+ resnet_pre_norm: bool = True,
1002
+ num_attention_heads: int = 1,
1003
+ cross_attention_dim: int = 1280,
1004
+ output_scale_factor: float = 1.0,
1005
+ add_upsample: bool = True,
1006
+ dual_cross_attention: bool = False,
1007
+ use_linear_projection: bool = False,
1008
+ only_cross_attention: bool = False,
1009
+ upcast_attention: bool = False,
1010
+ attention_type: str = "default",
1011
+ ):
1012
+ super().__init__()
1013
+ resnets = []
1014
+ attentions = []
1015
+
1016
+ self.has_cross_attention = True
1017
+ self.num_attention_heads = num_attention_heads
1018
+
1019
+ if isinstance(transformer_layers_per_block, int):
1020
+ transformer_layers_per_block = [transformer_layers_per_block] * num_layers
1021
+
1022
+ for i in range(num_layers):
1023
+ res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
1024
+ resnet_in_channels = prev_output_channel if i == 0 else out_channels
1025
+
1026
+ resnets.append(
1027
+ ResnetBlock2D(
1028
+ in_channels=resnet_in_channels + res_skip_channels,
1029
+ out_channels=out_channels,
1030
+ temb_channels=temb_channels,
1031
+ eps=resnet_eps,
1032
+ groups=resnet_groups,
1033
+ dropout=dropout,
1034
+ time_embedding_norm=resnet_time_scale_shift,
1035
+ non_linearity=resnet_act_fn,
1036
+ output_scale_factor=output_scale_factor,
1037
+ pre_norm=resnet_pre_norm,
1038
+ )
1039
+ )
1040
+ if not dual_cross_attention:
1041
+ attentions.append(
1042
+ Transformer2DModel(
1043
+ num_attention_heads,
1044
+ out_channels // num_attention_heads,
1045
+ in_channels=out_channels,
1046
+ num_layers=transformer_layers_per_block[i],
1047
+ cross_attention_dim=cross_attention_dim,
1048
+ norm_num_groups=resnet_groups,
1049
+ use_linear_projection=use_linear_projection,
1050
+ only_cross_attention=only_cross_attention,
1051
+ upcast_attention=upcast_attention,
1052
+ attention_type=attention_type,
1053
+ )
1054
+ )
1055
+ else:
1056
+ attentions.append(
1057
+ DualTransformer2DModel(
1058
+ num_attention_heads,
1059
+ out_channels // num_attention_heads,
1060
+ in_channels=out_channels,
1061
+ num_layers=1,
1062
+ cross_attention_dim=cross_attention_dim,
1063
+ norm_num_groups=resnet_groups,
1064
+ )
1065
+ )
1066
+ self.attentions = nn.ModuleList(attentions)
1067
+ self.resnets = nn.ModuleList(resnets)
1068
+
1069
+ if add_upsample:
1070
+ self.upsamplers = nn.ModuleList(
1071
+ [Upsample2D(out_channels, use_conv=True, out_channels=out_channels)]
1072
+ )
1073
+ else:
1074
+ self.upsamplers = None
1075
+
1076
+ self.gradient_checkpointing = False
1077
+ self.resolution_idx = resolution_idx
1078
+
1079
+ def forward(
1080
+ self,
1081
+ hidden_states: torch.FloatTensor,
1082
+ res_hidden_states_tuple: Tuple[torch.FloatTensor, ...],
1083
+ temb: Optional[torch.FloatTensor] = None,
1084
+ encoder_hidden_states: Optional[torch.FloatTensor] = None,
1085
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
1086
+ upsample_size: Optional[int] = None,
1087
+ attention_mask: Optional[torch.FloatTensor] = None,
1088
+ encoder_attention_mask: Optional[torch.FloatTensor] = None,
1089
+ ) -> torch.FloatTensor:
1090
+ """
1091
+ Forward pass for the CrossAttnUpBlock2D class.
1092
+
1093
+ Args:
1094
+ self (CrossAttnUpBlock2D): An instance of the CrossAttnUpBlock2D class.
1095
+ hidden_states (torch.FloatTensor): The input hidden states tensor.
1096
+ res_hidden_states_tuple (Tuple[torch.FloatTensor, ...]): A tuple of residual hidden states tensors.
1097
+ temb (Optional[torch.FloatTensor], optional): The token embeddings tensor. Defaults to None.
1098
+ encoder_hidden_states (Optional[torch.FloatTensor], optional): The encoder hidden states tensor. Defaults to None.
1099
+ cross_attention_kwargs (Optional[Dict[str, Any]], optional): Additional keyword arguments for cross attention. Defaults to None.
1100
+ upsample_size (Optional[int], optional): The upsample size. Defaults to None.
1101
+ attention_mask (Optional[torch.FloatTensor], optional): The attention mask tensor. Defaults to None.
1102
+ encoder_attention_mask (Optional[torch.FloatTensor], optional): The encoder attention mask tensor. Defaults to None.
1103
+
1104
+ Returns:
1105
+ torch.FloatTensor: The output tensor after passing through the block.
1106
+ """
1107
+ lora_scale = (
1108
+ cross_attention_kwargs.get("scale", 1.0)
1109
+ if cross_attention_kwargs is not None
1110
+ else 1.0
1111
+ )
1112
+ is_freeu_enabled = (
1113
+ getattr(self, "s1", None)
1114
+ and getattr(self, "s2", None)
1115
+ and getattr(self, "b1", None)
1116
+ and getattr(self, "b2", None)
1117
+ )
1118
+
1119
+ for resnet, attn in zip(self.resnets, self.attentions):
1120
+ # pop res hidden states
1121
+ res_hidden_states = res_hidden_states_tuple[-1]
1122
+ res_hidden_states_tuple = res_hidden_states_tuple[:-1]
1123
+
1124
+ # FreeU: Only operate on the first two stages
1125
+ if is_freeu_enabled:
1126
+ hidden_states, res_hidden_states = apply_freeu(
1127
+ self.resolution_idx,
1128
+ hidden_states,
1129
+ res_hidden_states,
1130
+ s1=self.s1,
1131
+ s2=self.s2,
1132
+ b1=self.b1,
1133
+ b2=self.b2,
1134
+ )
1135
+
1136
+ hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
1137
+
1138
+ if self.training and self.gradient_checkpointing:
1139
+
1140
+ def create_custom_forward(module, return_dict=None):
1141
+ def custom_forward(*inputs):
1142
+ if return_dict is not None:
1143
+ return module(*inputs, return_dict=return_dict)
1144
+
1145
+ return module(*inputs)
1146
+
1147
+ return custom_forward
1148
+
1149
+ ckpt_kwargs: Dict[str, Any] = (
1150
+ {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
1151
+ )
1152
+ hidden_states = torch.utils.checkpoint.checkpoint(
1153
+ create_custom_forward(resnet),
1154
+ hidden_states,
1155
+ temb,
1156
+ **ckpt_kwargs,
1157
+ )
1158
+ hidden_states, _ref_feature = attn(
1159
+ hidden_states,
1160
+ encoder_hidden_states=encoder_hidden_states,
1161
+ cross_attention_kwargs=cross_attention_kwargs,
1162
+ attention_mask=attention_mask,
1163
+ encoder_attention_mask=encoder_attention_mask,
1164
+ return_dict=False,
1165
+ )
1166
+ else:
1167
+ hidden_states = resnet(hidden_states, temb, scale=lora_scale)
1168
+ hidden_states, _ref_feature = attn(
1169
+ hidden_states,
1170
+ encoder_hidden_states=encoder_hidden_states,
1171
+ cross_attention_kwargs=cross_attention_kwargs,
1172
+ attention_mask=attention_mask,
1173
+ encoder_attention_mask=encoder_attention_mask,
1174
+ return_dict=False,
1175
+ )
1176
+
1177
+ if self.upsamplers is not None:
1178
+ for upsampler in self.upsamplers:
1179
+ hidden_states = upsampler(
1180
+ hidden_states, upsample_size, scale=lora_scale
1181
+ )
1182
+
1183
+ return hidden_states
1184
+
1185
+
1186
+ class UpBlock2D(nn.Module):
1187
+ """
1188
+ UpBlock2D is a class that represents a 2D upsampling block in a neural network.
1189
+
1190
+ This block is used for upsampling the input tensor by a factor of 2 in both dimensions.
1191
+ It takes the previous output channel, input channels, and output channels as input
1192
+ and applies a series of convolutional layers, batch normalization, and activation
1193
+ functions to produce the upsampled tensor.
1194
+
1195
+ Args:
1196
+ in_channels (int): The number of input channels in the tensor.
1197
+ prev_output_channel (int): The number of channels in the previous output tensor.
1198
+ out_channels (int): The number of output channels in the tensor.
1199
+ temb_channels (int): The number of channels in the time embedding tensor.
1200
+ resolution_idx (Optional[int], optional): The index of the resolution in the sequence of resolutions. Defaults to None.
1201
+ dropout (float, optional): The dropout rate to be applied to the convolutional layers. Defaults to 0.0.
1202
+ num_layers (int, optional): The number of convolutional layers in the block. Defaults to 1.
1203
+ resnet_eps (float, optional): The epsilon value used in the batch normalization layer. Defaults to 1e-6.
1204
+ resnet_time_scale_shift (str, optional): The type of activation function to be applied after the convolutional layers. Defaults to "default".
1205
+ resnet_act_fn (str, optional): The activation function to be applied after the batch normalization layer. Defaults to "swish".
1206
+ resnet_groups (int, optional): The number of groups in the group normalization layer. Defaults to 32.
1207
+ resnet_pre_norm (bool, optional): A flag indicating whether to apply layer normalization before the activation function. Defaults to True.
1208
+ output_scale_factor (float, optional): The scale factor to be applied to the output tensor. Defaults to 1.0.
1209
+ add_upsample (bool, optional): A flag indicating whether to add an upsampling layer to the block. Defaults to True.
1210
+
1211
+ Attributes:
1212
+ layers (nn.ModuleList): A list of nn.Module objects representing the convolutional layers in the block.
1213
+ upsample (nn.Module): The upsampling layer in the block, if add_upsample is True.
1214
+
1215
+ """
1216
+
1217
+ def __init__(
1218
+ self,
1219
+ in_channels: int,
1220
+ prev_output_channel: int,
1221
+ out_channels: int,
1222
+ temb_channels: int,
1223
+ resolution_idx: Optional[int] = None,
1224
+ dropout: float = 0.0,
1225
+ num_layers: int = 1,
1226
+ resnet_eps: float = 1e-6,
1227
+ resnet_time_scale_shift: str = "default",
1228
+ resnet_act_fn: str = "swish",
1229
+ resnet_groups: int = 32,
1230
+ resnet_pre_norm: bool = True,
1231
+ output_scale_factor: float = 1.0,
1232
+ add_upsample: bool = True,
1233
+ ):
1234
+ super().__init__()
1235
+ resnets = []
1236
+
1237
+ for i in range(num_layers):
1238
+ res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
1239
+ resnet_in_channels = prev_output_channel if i == 0 else out_channels
1240
+
1241
+ resnets.append(
1242
+ ResnetBlock2D(
1243
+ in_channels=resnet_in_channels + res_skip_channels,
1244
+ out_channels=out_channels,
1245
+ temb_channels=temb_channels,
1246
+ eps=resnet_eps,
1247
+ groups=resnet_groups,
1248
+ dropout=dropout,
1249
+ time_embedding_norm=resnet_time_scale_shift,
1250
+ non_linearity=resnet_act_fn,
1251
+ output_scale_factor=output_scale_factor,
1252
+ pre_norm=resnet_pre_norm,
1253
+ )
1254
+ )
1255
+
1256
+ self.resnets = nn.ModuleList(resnets)
1257
+
1258
+ if add_upsample:
1259
+ self.upsamplers = nn.ModuleList(
1260
+ [Upsample2D(out_channels, use_conv=True, out_channels=out_channels)]
1261
+ )
1262
+ else:
1263
+ self.upsamplers = None
1264
+
1265
+ self.gradient_checkpointing = False
1266
+ self.resolution_idx = resolution_idx
1267
+
1268
+ def forward(
1269
+ self,
1270
+ hidden_states: torch.FloatTensor,
1271
+ res_hidden_states_tuple: Tuple[torch.FloatTensor, ...],
1272
+ temb: Optional[torch.FloatTensor] = None,
1273
+ upsample_size: Optional[int] = None,
1274
+ scale: float = 1.0,
1275
+ ) -> torch.FloatTensor:
1276
+
1277
+ """
1278
+ Forward pass for the UpBlock2D class.
1279
+
1280
+ Args:
1281
+ self (UpBlock2D): An instance of the UpBlock2D class.
1282
+ hidden_states (torch.FloatTensor): The input tensor to the block.
1283
+ res_hidden_states_tuple (Tuple[torch.FloatTensor, ...]): A tuple of residual hidden states.
1284
+ temb (Optional[torch.FloatTensor], optional): The token embeddings. Defaults to None.
1285
+ upsample_size (Optional[int], optional): The size to upsample the input tensor to. Defaults to None.
1286
+ scale (float, optional): The scale factor to apply to the input tensor. Defaults to 1.0.
1287
+
1288
+ Returns:
1289
+ torch.FloatTensor: The output tensor after passing through the block.
1290
+ """
1291
+ is_freeu_enabled = (
1292
+ getattr(self, "s1", None)
1293
+ and getattr(self, "s2", None)
1294
+ and getattr(self, "b1", None)
1295
+ and getattr(self, "b2", None)
1296
+ )
1297
+
1298
+ for resnet in self.resnets:
1299
+ # pop res hidden states
1300
+ res_hidden_states = res_hidden_states_tuple[-1]
1301
+ res_hidden_states_tuple = res_hidden_states_tuple[:-1]
1302
+
1303
+ # FreeU: Only operate on the first two stages
1304
+ if is_freeu_enabled:
1305
+ hidden_states, res_hidden_states = apply_freeu(
1306
+ self.resolution_idx,
1307
+ hidden_states,
1308
+ res_hidden_states,
1309
+ s1=self.s1,
1310
+ s2=self.s2,
1311
+ b1=self.b1,
1312
+ b2=self.b2,
1313
+ )
1314
+
1315
+ hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
1316
+
1317
+ if self.training and self.gradient_checkpointing:
1318
+
1319
+ def create_custom_forward(module):
1320
+ def custom_forward(*inputs):
1321
+ return module(*inputs)
1322
+
1323
+ return custom_forward
1324
+
1325
+ if is_torch_version(">=", "1.11.0"):
1326
+ hidden_states = torch.utils.checkpoint.checkpoint(
1327
+ create_custom_forward(resnet),
1328
+ hidden_states,
1329
+ temb,
1330
+ use_reentrant=False,
1331
+ )
1332
+ else:
1333
+ hidden_states = torch.utils.checkpoint.checkpoint(
1334
+ create_custom_forward(resnet), hidden_states, temb
1335
+ )
1336
+ else:
1337
+ hidden_states = resnet(hidden_states, temb, scale=scale)
1338
+
1339
+ if self.upsamplers is not None:
1340
+ for upsampler in self.upsamplers:
1341
+ hidden_states = upsampler(hidden_states, upsample_size, scale=scale)
1342
+
1343
+ return hidden_states
hallo/models/unet_2d_condition.py ADDED
@@ -0,0 +1,1432 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # pylint: disable=R0801
2
+ # pylint: disable=E1101
3
+ # pylint: disable=W1203
4
+
5
+ """
6
+ This module implements the `UNet2DConditionModel`,
7
+ a variant of the 2D U-Net architecture designed for conditional image generation tasks.
8
+ The model is capable of taking a noisy input sample and conditioning it based on additional information such as class labels,
9
+ time steps, and encoder hidden states to produce a denoised output.
10
+
11
+ The `UNet2DConditionModel` leverages various components such as time embeddings,
12
+ class embeddings, and cross-attention mechanisms to integrate the conditioning information effectively.
13
+ It is built upon several sub-blocks including down-blocks, a middle block, and up-blocks,
14
+ each responsible for different stages of the U-Net's downsampling and upsampling process.
15
+
16
+ Key Features:
17
+ - Support for multiple types of down and up blocks, including those with cross-attention capabilities.
18
+ - Flexible configuration of the model's layers, including the number of layers per block and the output channels for each block.
19
+ - Integration of time embeddings and class embeddings to condition the model's output on additional information.
20
+ - Implementation of cross-attention to leverage encoder hidden states for conditional generation.
21
+ - The model supports gradient checkpointing to reduce memory usage during training.
22
+
23
+ The module also includes utility functions and classes such as `UNet2DConditionOutput` for structured output
24
+ and `load_change_cross_attention_dim` for loading and modifying pre-trained models.
25
+
26
+ Example Usage:
27
+ >>> import torch
28
+ >>> from unet_2d_condition_model import UNet2DConditionModel
29
+ >>> model = UNet2DConditionModel(
30
+ ... sample_size=(64, 64),
31
+ ... in_channels=3,
32
+ ... out_channels=3,
33
+ ... encoder_hid_dim=512,
34
+ ... cross_attention_dim=1024,
35
+ ... )
36
+ >>> # Prepare input tensors
37
+ >>> sample = torch.randn(1, 3, 64, 64)
38
+ >>> timestep = 0
39
+ >>> encoder_hidden_states = torch.randn(1, 14, 512)
40
+ >>> # Forward pass through the model
41
+ >>> output = model(sample, timestep, encoder_hidden_states)
42
+
43
+ This module is part of a larger ecosystem of diffusion models and can be used for various conditional image generation tasks.
44
+ """
45
+
46
+ from dataclasses import dataclass
47
+ from os import PathLike
48
+ from pathlib import Path
49
+ from typing import Any, Dict, List, Optional, Tuple, Union
50
+
51
+ import torch
52
+ import torch.utils.checkpoint
53
+ from diffusers.configuration_utils import ConfigMixin, register_to_config
54
+ from diffusers.loaders import UNet2DConditionLoadersMixin
55
+ from diffusers.models.activations import get_activation
56
+ from diffusers.models.attention_processor import (
57
+ ADDED_KV_ATTENTION_PROCESSORS, CROSS_ATTENTION_PROCESSORS,
58
+ AttentionProcessor, AttnAddedKVProcessor, AttnProcessor)
59
+ from diffusers.models.embeddings import (GaussianFourierProjection,
60
+ GLIGENTextBoundingboxProjection,
61
+ ImageHintTimeEmbedding,
62
+ ImageProjection, ImageTimeEmbedding,
63
+ TextImageProjection,
64
+ TextImageTimeEmbedding,
65
+ TextTimeEmbedding, TimestepEmbedding,
66
+ Timesteps)
67
+ from diffusers.models.modeling_utils import ModelMixin
68
+ from diffusers.utils import (SAFETENSORS_WEIGHTS_NAME, USE_PEFT_BACKEND,
69
+ WEIGHTS_NAME, BaseOutput, deprecate, logging,
70
+ scale_lora_layers, unscale_lora_layers)
71
+ from safetensors.torch import load_file
72
+ from torch import nn
73
+
74
+ from .unet_2d_blocks import (UNetMidBlock2D, UNetMidBlock2DCrossAttn,
75
+ get_down_block, get_up_block)
76
+
77
+ logger = logging.get_logger(__name__) # pylint: disable=invalid-name
78
+
79
+ @dataclass
80
+ class UNet2DConditionOutput(BaseOutput):
81
+ """
82
+ The output of [`UNet2DConditionModel`].
83
+
84
+ Args:
85
+ sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
86
+ The hidden states output conditioned on `encoder_hidden_states` input. Output of last layer of model.
87
+ """
88
+
89
+ sample: torch.FloatTensor = None
90
+ ref_features: Tuple[torch.FloatTensor] = None
91
+
92
+
93
+ class UNet2DConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin):
94
+ r"""
95
+ A conditional 2D UNet model that takes a noisy sample, conditional state, and a timestep and returns a sample
96
+ shaped output.
97
+
98
+ This model inherits from [`ModelMixin`]. Check the superclass documentation for it's generic methods implemented
99
+ for all models (such as downloading or saving).
100
+
101
+ Parameters:
102
+ sample_size (`int` or `Tuple[int, int]`, *optional*, defaults to `None`):
103
+ Height and width of input/output sample.
104
+ in_channels (`int`, *optional*, defaults to 4): Number of channels in the input sample.
105
+ out_channels (`int`, *optional*, defaults to 4): Number of channels in the output.
106
+ center_input_sample (`bool`, *optional*, defaults to `False`): Whether to center the input sample.
107
+ flip_sin_to_cos (`bool`, *optional*, defaults to `False`):
108
+ Whether to flip the sin to cos in the time embedding.
109
+ freq_shift (`int`, *optional*, defaults to 0): The frequency shift to apply to the time embedding.
110
+ down_block_types (`Tuple[str]`, *optional*, defaults to
111
+ `("CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "DownBlock2D")`):
112
+ The tuple of downsample blocks to use.
113
+ mid_block_type (`str`, *optional*, defaults to `"UNetMidBlock2DCrossAttn"`):
114
+ Block type for middle of UNet, it can be one of `UNetMidBlock2DCrossAttn`, `UNetMidBlock2D`, or
115
+ `UNetMidBlock2DSimpleCrossAttn`. If `None`, the mid block layer is skipped.
116
+ up_block_types (`Tuple[str]`, *optional*, defaults to `("UpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D")`):
117
+ The tuple of upsample blocks to use.
118
+ only_cross_attention(`bool` or `Tuple[bool]`, *optional*, default to `False`):
119
+ Whether to include self-attention in the basic transformer blocks, see
120
+ [`~models.attention.BasicTransformerBlock`].
121
+ block_out_channels (`Tuple[int]`, *optional*, defaults to `(320, 640, 1280, 1280)`):
122
+ The tuple of output channels for each block.
123
+ layers_per_block (`int`, *optional*, defaults to 2): The number of layers per block.
124
+ downsample_padding (`int`, *optional*, defaults to 1): The padding to use for the downsampling convolution.
125
+ mid_block_scale_factor (`float`, *optional*, defaults to 1.0): The scale factor to use for the mid block.
126
+ dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
127
+ act_fn (`str`, *optional*, defaults to `"silu"`): The activation function to use.
128
+ norm_num_groups (`int`, *optional*, defaults to 32): The number of groups to use for the normalization.
129
+ If `None`, normalization and activation layers is skipped in post-processing.
130
+ norm_eps (`float`, *optional*, defaults to 1e-5): The epsilon to use for the normalization.
131
+ cross_attention_dim (`int` or `Tuple[int]`, *optional*, defaults to 1280):
132
+ The dimension of the cross attention features.
133
+ transformer_layers_per_block (`int`, `Tuple[int]`, or `Tuple[Tuple]` , *optional*, defaults to 1):
134
+ The number of transformer blocks of type [`~models.attention.BasicTransformerBlock`]. Only relevant for
135
+ [`~models.unet_2d_blocks.CrossAttnDownBlock2D`], [`~models.unet_2d_blocks.CrossAttnUpBlock2D`],
136
+ [`~models.unet_2d_blocks.UNetMidBlock2DCrossAttn`].
137
+ reverse_transformer_layers_per_block : (`Tuple[Tuple]`, *optional*, defaults to None):
138
+ The number of transformer blocks of type [`~models.attention.BasicTransformerBlock`], in the upsampling
139
+ blocks of the U-Net. Only relevant if `transformer_layers_per_block` is of type `Tuple[Tuple]` and for
140
+ [`~models.unet_2d_blocks.CrossAttnDownBlock2D`], [`~models.unet_2d_blocks.CrossAttnUpBlock2D`],
141
+ [`~models.unet_2d_blocks.UNetMidBlock2DCrossAttn`].
142
+ encoder_hid_dim (`int`, *optional*, defaults to None):
143
+ If `encoder_hid_dim_type` is defined, `encoder_hidden_states` will be projected from `encoder_hid_dim`
144
+ dimension to `cross_attention_dim`.
145
+ encoder_hid_dim_type (`str`, *optional*, defaults to `None`):
146
+ If given, the `encoder_hidden_states` and potentially other embeddings are down-projected to text
147
+ embeddings of dimension `cross_attention` according to `encoder_hid_dim_type`.
148
+ attention_head_dim (`int`, *optional*, defaults to 8): The dimension of the attention heads.
149
+ num_attention_heads (`int`, *optional*):
150
+ The number of attention heads. If not defined, defaults to `attention_head_dim`
151
+ resnet_time_scale_shift (`str`, *optional*, defaults to `"default"`): Time scale shift config
152
+ for ResNet blocks (see [`~models.resnet.ResnetBlock2D`]). Choose from `default` or `scale_shift`.
153
+ class_embed_type (`str`, *optional*, defaults to `None`):
154
+ The type of class embedding to use which is ultimately summed with the time embeddings. Choose from `None`,
155
+ `"timestep"`, `"identity"`, `"projection"`, or `"simple_projection"`.
156
+ addition_embed_type (`str`, *optional*, defaults to `None`):
157
+ Configures an optional embedding which will be summed with the time embeddings. Choose from `None` or
158
+ "text". "text" will use the `TextTimeEmbedding` layer.
159
+ addition_time_embed_dim: (`int`, *optional*, defaults to `None`):
160
+ Dimension for the timestep embeddings.
161
+ num_class_embeds (`int`, *optional*, defaults to `None`):
162
+ Input dimension of the learnable embedding matrix to be projected to `time_embed_dim`, when performing
163
+ class conditioning with `class_embed_type` equal to `None`.
164
+ time_embedding_type (`str`, *optional*, defaults to `positional`):
165
+ The type of position embedding to use for timesteps. Choose from `positional` or `fourier`.
166
+ time_embedding_dim (`int`, *optional*, defaults to `None`):
167
+ An optional override for the dimension of the projected time embedding.
168
+ time_embedding_act_fn (`str`, *optional*, defaults to `None`):
169
+ Optional activation function to use only once on the time embeddings before they are passed to the rest of
170
+ the UNet. Choose from `silu`, `mish`, `gelu`, and `swish`.
171
+ timestep_post_act (`str`, *optional*, defaults to `None`):
172
+ The second activation function to use in timestep embedding. Choose from `silu`, `mish` and `gelu`.
173
+ time_cond_proj_dim (`int`, *optional*, defaults to `None`):
174
+ The dimension of `cond_proj` layer in the timestep embedding.
175
+ conv_in_kernel (`int`, *optional*, default to `3`): The kernel size of `conv_in` layer. conv_out_kernel (`int`,
176
+ *optional*, default to `3`): The kernel size of `conv_out` layer. projection_class_embeddings_input_dim (`int`,
177
+ *optional*): The dimension of the `class_labels` input when
178
+ `class_embed_type="projection"`. Required when `class_embed_type="projection"`.
179
+ class_embeddings_concat (`bool`, *optional*, defaults to `False`): Whether to concatenate the time
180
+ embeddings with the class embeddings.
181
+ mid_block_only_cross_attention (`bool`, *optional*, defaults to `None`):
182
+ Whether to use cross attention with the mid block when using the `UNetMidBlock2DSimpleCrossAttn`. If
183
+ `only_cross_attention` is given as a single boolean and `mid_block_only_cross_attention` is `None`, the
184
+ `only_cross_attention` value is used as the value for `mid_block_only_cross_attention`. Default to `False`
185
+ otherwise.
186
+ """
187
+
188
+ _supports_gradient_checkpointing = True
189
+
190
+ @register_to_config
191
+ def __init__(
192
+ self,
193
+ sample_size: Optional[int] = None,
194
+ in_channels: int = 4,
195
+ _out_channels: int = 4,
196
+ _center_input_sample: bool = False,
197
+ flip_sin_to_cos: bool = True,
198
+ freq_shift: int = 0,
199
+ down_block_types: Tuple[str] = (
200
+ "CrossAttnDownBlock2D",
201
+ "CrossAttnDownBlock2D",
202
+ "CrossAttnDownBlock2D",
203
+ "DownBlock2D",
204
+ ),
205
+ mid_block_type: Optional[str] = "UNetMidBlock2DCrossAttn",
206
+ up_block_types: Tuple[str] = (
207
+ "UpBlock2D",
208
+ "CrossAttnUpBlock2D",
209
+ "CrossAttnUpBlock2D",
210
+ "CrossAttnUpBlock2D",
211
+ ),
212
+ only_cross_attention: Union[bool, Tuple[bool]] = False,
213
+ block_out_channels: Tuple[int] = (320, 640, 1280, 1280),
214
+ layers_per_block: Union[int, Tuple[int]] = 2,
215
+ downsample_padding: int = 1,
216
+ mid_block_scale_factor: float = 1,
217
+ dropout: float = 0.0,
218
+ act_fn: str = "silu",
219
+ norm_num_groups: Optional[int] = 32,
220
+ norm_eps: float = 1e-5,
221
+ cross_attention_dim: Union[int, Tuple[int]] = 1280,
222
+ transformer_layers_per_block: Union[int, Tuple[int], Tuple[Tuple]] = 1,
223
+ reverse_transformer_layers_per_block: Optional[Tuple[Tuple[int]]] = None,
224
+ encoder_hid_dim: Optional[int] = None,
225
+ encoder_hid_dim_type: Optional[str] = None,
226
+ attention_head_dim: Union[int, Tuple[int]] = 8,
227
+ num_attention_heads: Optional[Union[int, Tuple[int]]] = None,
228
+ dual_cross_attention: bool = False,
229
+ use_linear_projection: bool = False,
230
+ class_embed_type: Optional[str] = None,
231
+ addition_embed_type: Optional[str] = None,
232
+ addition_time_embed_dim: Optional[int] = None,
233
+ num_class_embeds: Optional[int] = None,
234
+ upcast_attention: bool = False,
235
+ resnet_time_scale_shift: str = "default",
236
+ time_embedding_type: str = "positional",
237
+ time_embedding_dim: Optional[int] = None,
238
+ time_embedding_act_fn: Optional[str] = None,
239
+ timestep_post_act: Optional[str] = None,
240
+ time_cond_proj_dim: Optional[int] = None,
241
+ conv_in_kernel: int = 3,
242
+ projection_class_embeddings_input_dim: Optional[int] = None,
243
+ attention_type: str = "default",
244
+ class_embeddings_concat: bool = False,
245
+ mid_block_only_cross_attention: Optional[bool] = None,
246
+ addition_embed_type_num_heads=64,
247
+ _landmark_net=False,
248
+ ):
249
+ super().__init__()
250
+
251
+ self.sample_size = sample_size
252
+
253
+ if num_attention_heads is not None:
254
+ raise ValueError(
255
+ "At the moment it is not possible to define the number of attention heads via `num_attention_heads`"
256
+ "because of a naming issue as described in https://github.com/huggingface/diffusers/issues/2011#issuecomment-1547958131."
257
+ "Passing `num_attention_heads` will only be supported in diffusers v0.19."
258
+ )
259
+
260
+ # If `num_attention_heads` is not defined (which is the case for most models)
261
+ # it will default to `attention_head_dim`. This looks weird upon first reading it and it is.
262
+ # The reason for this behavior is to correct for incorrectly named variables that were introduced
263
+ # when this library was created. The incorrect naming was only discovered much later in
264
+ # https://github.com/huggingface/diffusers/issues/2011#issuecomment-1547958131
265
+ # Changing `attention_head_dim` to `num_attention_heads` for 40,000+ configurations is too backwards breaking
266
+ # which is why we correct for the naming here.
267
+ num_attention_heads = num_attention_heads or attention_head_dim
268
+
269
+ # Check inputs
270
+ if len(down_block_types) != len(up_block_types):
271
+ raise ValueError(
272
+ "Must provide the same number of `down_block_types` as `up_block_types`."
273
+ f"`down_block_types`: {down_block_types}. `up_block_types`: {up_block_types}."
274
+ )
275
+
276
+ if len(block_out_channels) != len(down_block_types):
277
+ raise ValueError(
278
+ "Must provide the same number of `block_out_channels` as `down_block_types`."
279
+ f"`block_out_channels`: {block_out_channels}. `down_block_types`: {down_block_types}."
280
+ )
281
+
282
+ if not isinstance(only_cross_attention, bool) and len(
283
+ only_cross_attention
284
+ ) != len(down_block_types):
285
+ raise ValueError(
286
+ "Must provide the same number of `only_cross_attention` as `down_block_types`."
287
+ f"`only_cross_attention`: {only_cross_attention}. `down_block_types`: {down_block_types}."
288
+ )
289
+
290
+ if not isinstance(num_attention_heads, int) and len(num_attention_heads) != len(
291
+ down_block_types
292
+ ):
293
+ raise ValueError(
294
+ "Must provide the same number of `num_attention_heads` as `down_block_types`."
295
+ f"`num_attention_heads`: {num_attention_heads}. `down_block_types`: {down_block_types}."
296
+ )
297
+
298
+ if not isinstance(attention_head_dim, int) and len(attention_head_dim) != len(
299
+ down_block_types
300
+ ):
301
+ raise ValueError(
302
+ "Must provide the same number of `attention_head_dim` as `down_block_types`."
303
+ f"`attention_head_dim`: {attention_head_dim}. `down_block_types`: {down_block_types}."
304
+ )
305
+
306
+ if isinstance(cross_attention_dim, list) and len(cross_attention_dim) != len(
307
+ down_block_types
308
+ ):
309
+ raise ValueError(
310
+ "Must provide the same number of `cross_attention_dim` as `down_block_types`."
311
+ f"`cross_attention_dim`: {cross_attention_dim}. `down_block_types`: {down_block_types}."
312
+ )
313
+
314
+ if not isinstance(layers_per_block, int) and len(layers_per_block) != len(
315
+ down_block_types
316
+ ):
317
+ raise ValueError(
318
+ "Must provide the same number of `layers_per_block` as `down_block_types`."
319
+ f"`layers_per_block`: {layers_per_block}. `down_block_types`: {down_block_types}."
320
+ )
321
+ if (
322
+ isinstance(transformer_layers_per_block, list)
323
+ and reverse_transformer_layers_per_block is None
324
+ ):
325
+ for layer_number_per_block in transformer_layers_per_block:
326
+ if isinstance(layer_number_per_block, list):
327
+ raise ValueError(
328
+ "Must provide 'reverse_transformer_layers_per_block` if using asymmetrical UNet."
329
+ )
330
+
331
+ # input
332
+ conv_in_padding = (conv_in_kernel - 1) // 2
333
+ self.conv_in = nn.Conv2d(
334
+ in_channels,
335
+ block_out_channels[0],
336
+ kernel_size=conv_in_kernel,
337
+ padding=conv_in_padding,
338
+ )
339
+
340
+ # time
341
+ if time_embedding_type == "fourier":
342
+ time_embed_dim = time_embedding_dim or block_out_channels[0] * 2
343
+ if time_embed_dim % 2 != 0:
344
+ raise ValueError(
345
+ f"`time_embed_dim` should be divisible by 2, but is {time_embed_dim}."
346
+ )
347
+ self.time_proj = GaussianFourierProjection(
348
+ time_embed_dim // 2,
349
+ set_W_to_weight=False,
350
+ log=False,
351
+ flip_sin_to_cos=flip_sin_to_cos,
352
+ )
353
+ timestep_input_dim = time_embed_dim
354
+ elif time_embedding_type == "positional":
355
+ time_embed_dim = time_embedding_dim or block_out_channels[0] * 4
356
+
357
+ self.time_proj = Timesteps(
358
+ block_out_channels[0], flip_sin_to_cos, freq_shift
359
+ )
360
+ timestep_input_dim = block_out_channels[0]
361
+ else:
362
+ raise ValueError(
363
+ f"{time_embedding_type} does not exist. Please make sure to use one of `fourier` or `positional`."
364
+ )
365
+
366
+ self.time_embedding = TimestepEmbedding(
367
+ timestep_input_dim,
368
+ time_embed_dim,
369
+ act_fn=act_fn,
370
+ post_act_fn=timestep_post_act,
371
+ cond_proj_dim=time_cond_proj_dim,
372
+ )
373
+
374
+ if encoder_hid_dim_type is None and encoder_hid_dim is not None:
375
+ encoder_hid_dim_type = "text_proj"
376
+ self.register_to_config(encoder_hid_dim_type=encoder_hid_dim_type)
377
+ logger.info(
378
+ "encoder_hid_dim_type defaults to 'text_proj' as `encoder_hid_dim` is defined."
379
+ )
380
+
381
+ if encoder_hid_dim is None and encoder_hid_dim_type is not None:
382
+ raise ValueError(
383
+ f"`encoder_hid_dim` has to be defined when `encoder_hid_dim_type` is set to {encoder_hid_dim_type}."
384
+ )
385
+
386
+ if encoder_hid_dim_type == "text_proj":
387
+ self.encoder_hid_proj = nn.Linear(
388
+ encoder_hid_dim, cross_attention_dim)
389
+ elif encoder_hid_dim_type == "text_image_proj":
390
+ # image_embed_dim DOESN'T have to be `cross_attention_dim`. To not clutter the __init__ too much
391
+ # they are set to `cross_attention_dim` here as this is exactly the required dimension for the currently only use
392
+ # case when `addition_embed_type == "text_image_proj"` (Kadinsky 2.1)`
393
+ self.encoder_hid_proj = TextImageProjection(
394
+ text_embed_dim=encoder_hid_dim,
395
+ image_embed_dim=cross_attention_dim,
396
+ cross_attention_dim=cross_attention_dim,
397
+ )
398
+ elif encoder_hid_dim_type == "image_proj":
399
+ # Kandinsky 2.2
400
+ self.encoder_hid_proj = ImageProjection(
401
+ image_embed_dim=encoder_hid_dim,
402
+ cross_attention_dim=cross_attention_dim,
403
+ )
404
+ elif encoder_hid_dim_type is not None:
405
+ raise ValueError(
406
+ f"encoder_hid_dim_type: {encoder_hid_dim_type} must be None, 'text_proj' or 'text_image_proj'."
407
+ )
408
+ else:
409
+ self.encoder_hid_proj = None
410
+
411
+ # class embedding
412
+ if class_embed_type is None and num_class_embeds is not None:
413
+ self.class_embedding = nn.Embedding(
414
+ num_class_embeds, time_embed_dim)
415
+ elif class_embed_type == "timestep":
416
+ self.class_embedding = TimestepEmbedding(
417
+ timestep_input_dim, time_embed_dim, act_fn=act_fn
418
+ )
419
+ elif class_embed_type == "identity":
420
+ self.class_embedding = nn.Identity(time_embed_dim, time_embed_dim)
421
+ elif class_embed_type == "projection":
422
+ if projection_class_embeddings_input_dim is None:
423
+ raise ValueError(
424
+ "`class_embed_type`: 'projection' requires `projection_class_embeddings_input_dim` be set"
425
+ )
426
+ # The projection `class_embed_type` is the same as the timestep `class_embed_type` except
427
+ # 1. the `class_labels` inputs are not first converted to sinusoidal embeddings
428
+ # 2. it projects from an arbitrary input dimension.
429
+ #
430
+ # Note that `TimestepEmbedding` is quite general, being mainly linear layers and activations.
431
+ # When used for embedding actual timesteps, the timesteps are first converted to sinusoidal embeddings.
432
+ # As a result, `TimestepEmbedding` can be passed arbitrary vectors.
433
+ self.class_embedding = TimestepEmbedding(
434
+ projection_class_embeddings_input_dim, time_embed_dim
435
+ )
436
+ elif class_embed_type == "simple_projection":
437
+ if projection_class_embeddings_input_dim is None:
438
+ raise ValueError(
439
+ "`class_embed_type`: 'simple_projection' requires `projection_class_embeddings_input_dim` be set"
440
+ )
441
+ self.class_embedding = nn.Linear(
442
+ projection_class_embeddings_input_dim, time_embed_dim
443
+ )
444
+ else:
445
+ self.class_embedding = None
446
+
447
+ if addition_embed_type == "text":
448
+ if encoder_hid_dim is not None:
449
+ text_time_embedding_from_dim = encoder_hid_dim
450
+ else:
451
+ text_time_embedding_from_dim = cross_attention_dim
452
+
453
+ self.add_embedding = TextTimeEmbedding(
454
+ text_time_embedding_from_dim,
455
+ time_embed_dim,
456
+ num_heads=addition_embed_type_num_heads,
457
+ )
458
+ elif addition_embed_type == "text_image":
459
+ # text_embed_dim and image_embed_dim DON'T have to be `cross_attention_dim`. To not clutter the __init__ too much
460
+ # they are set to `cross_attention_dim` here as this is exactly the required dimension for the currently only use
461
+ # case when `addition_embed_type == "text_image"` (Kadinsky 2.1)`
462
+ self.add_embedding = TextImageTimeEmbedding(
463
+ text_embed_dim=cross_attention_dim,
464
+ image_embed_dim=cross_attention_dim,
465
+ time_embed_dim=time_embed_dim,
466
+ )
467
+ elif addition_embed_type == "text_time":
468
+ self.add_time_proj = Timesteps(
469
+ addition_time_embed_dim, flip_sin_to_cos, freq_shift
470
+ )
471
+ self.add_embedding = TimestepEmbedding(
472
+ projection_class_embeddings_input_dim, time_embed_dim
473
+ )
474
+ elif addition_embed_type == "image":
475
+ # Kandinsky 2.2
476
+ self.add_embedding = ImageTimeEmbedding(
477
+ image_embed_dim=encoder_hid_dim, time_embed_dim=time_embed_dim
478
+ )
479
+ elif addition_embed_type == "image_hint":
480
+ # Kandinsky 2.2 ControlNet
481
+ self.add_embedding = ImageHintTimeEmbedding(
482
+ image_embed_dim=encoder_hid_dim, time_embed_dim=time_embed_dim
483
+ )
484
+ elif addition_embed_type is not None:
485
+ raise ValueError(
486
+ f"addition_embed_type: {addition_embed_type} must be None, 'text' or 'text_image'."
487
+ )
488
+
489
+ if time_embedding_act_fn is None:
490
+ self.time_embed_act = None
491
+ else:
492
+ self.time_embed_act = get_activation(time_embedding_act_fn)
493
+
494
+ self.down_blocks = nn.ModuleList([])
495
+ self.up_blocks = nn.ModuleList([])
496
+
497
+ if isinstance(only_cross_attention, bool):
498
+ if mid_block_only_cross_attention is None:
499
+ mid_block_only_cross_attention = only_cross_attention
500
+
501
+ only_cross_attention = [
502
+ only_cross_attention] * len(down_block_types)
503
+
504
+ if mid_block_only_cross_attention is None:
505
+ mid_block_only_cross_attention = False
506
+
507
+ if isinstance(num_attention_heads, int):
508
+ num_attention_heads = (num_attention_heads,) * \
509
+ len(down_block_types)
510
+
511
+ if isinstance(attention_head_dim, int):
512
+ attention_head_dim = (attention_head_dim,) * len(down_block_types)
513
+
514
+ if isinstance(cross_attention_dim, int):
515
+ cross_attention_dim = (cross_attention_dim,) * \
516
+ len(down_block_types)
517
+
518
+ if isinstance(layers_per_block, int):
519
+ layers_per_block = [layers_per_block] * len(down_block_types)
520
+
521
+ if isinstance(transformer_layers_per_block, int):
522
+ transformer_layers_per_block = [transformer_layers_per_block] * len(
523
+ down_block_types
524
+ )
525
+
526
+ if class_embeddings_concat:
527
+ # The time embeddings are concatenated with the class embeddings. The dimension of the
528
+ # time embeddings passed to the down, middle, and up blocks is twice the dimension of the
529
+ # regular time embeddings
530
+ blocks_time_embed_dim = time_embed_dim * 2
531
+ else:
532
+ blocks_time_embed_dim = time_embed_dim
533
+
534
+ # down
535
+ output_channel = block_out_channels[0]
536
+ for i, down_block_type in enumerate(down_block_types):
537
+ input_channel = output_channel
538
+ output_channel = block_out_channels[i]
539
+ is_final_block = i == len(block_out_channels) - 1
540
+
541
+ down_block = get_down_block(
542
+ down_block_type,
543
+ num_layers=layers_per_block[i],
544
+ transformer_layers_per_block=transformer_layers_per_block[i],
545
+ in_channels=input_channel,
546
+ out_channels=output_channel,
547
+ temb_channels=blocks_time_embed_dim,
548
+ add_downsample=not is_final_block,
549
+ resnet_eps=norm_eps,
550
+ resnet_act_fn=act_fn,
551
+ resnet_groups=norm_num_groups,
552
+ cross_attention_dim=cross_attention_dim[i],
553
+ num_attention_heads=num_attention_heads[i],
554
+ downsample_padding=downsample_padding,
555
+ dual_cross_attention=dual_cross_attention,
556
+ use_linear_projection=use_linear_projection,
557
+ only_cross_attention=only_cross_attention[i],
558
+ upcast_attention=upcast_attention,
559
+ resnet_time_scale_shift=resnet_time_scale_shift,
560
+ attention_type=attention_type,
561
+ attention_head_dim=(
562
+ attention_head_dim[i]
563
+ if attention_head_dim[i] is not None
564
+ else output_channel
565
+ ),
566
+ dropout=dropout,
567
+ )
568
+ self.down_blocks.append(down_block)
569
+
570
+ # mid
571
+ if mid_block_type == "UNetMidBlock2DCrossAttn":
572
+ self.mid_block = UNetMidBlock2DCrossAttn(
573
+ transformer_layers_per_block=transformer_layers_per_block[-1],
574
+ in_channels=block_out_channels[-1],
575
+ temb_channels=blocks_time_embed_dim,
576
+ dropout=dropout,
577
+ resnet_eps=norm_eps,
578
+ resnet_act_fn=act_fn,
579
+ output_scale_factor=mid_block_scale_factor,
580
+ resnet_time_scale_shift=resnet_time_scale_shift,
581
+ cross_attention_dim=cross_attention_dim[-1],
582
+ num_attention_heads=num_attention_heads[-1],
583
+ resnet_groups=norm_num_groups,
584
+ dual_cross_attention=dual_cross_attention,
585
+ use_linear_projection=use_linear_projection,
586
+ upcast_attention=upcast_attention,
587
+ attention_type=attention_type,
588
+ )
589
+ elif mid_block_type == "UNetMidBlock2DSimpleCrossAttn":
590
+ raise NotImplementedError(
591
+ f"Unsupport mid_block_type: {mid_block_type}")
592
+ elif mid_block_type == "UNetMidBlock2D":
593
+ self.mid_block = UNetMidBlock2D(
594
+ in_channels=block_out_channels[-1],
595
+ temb_channels=blocks_time_embed_dim,
596
+ dropout=dropout,
597
+ num_layers=0,
598
+ resnet_eps=norm_eps,
599
+ resnet_act_fn=act_fn,
600
+ output_scale_factor=mid_block_scale_factor,
601
+ resnet_groups=norm_num_groups,
602
+ resnet_time_scale_shift=resnet_time_scale_shift,
603
+ add_attention=False,
604
+ )
605
+ elif mid_block_type is None:
606
+ self.mid_block = None
607
+ else:
608
+ raise ValueError(f"unknown mid_block_type : {mid_block_type}")
609
+
610
+ # count how many layers upsample the images
611
+ self.num_upsamplers = 0
612
+
613
+ # up
614
+ reversed_block_out_channels = list(reversed(block_out_channels))
615
+ reversed_num_attention_heads = list(reversed(num_attention_heads))
616
+ reversed_layers_per_block = list(reversed(layers_per_block))
617
+ reversed_cross_attention_dim = list(reversed(cross_attention_dim))
618
+ reversed_transformer_layers_per_block = (
619
+ list(reversed(transformer_layers_per_block))
620
+ if reverse_transformer_layers_per_block is None
621
+ else reverse_transformer_layers_per_block
622
+ )
623
+ only_cross_attention = list(reversed(only_cross_attention))
624
+
625
+ output_channel = reversed_block_out_channels[0]
626
+ for i, up_block_type in enumerate(up_block_types):
627
+ is_final_block = i == len(block_out_channels) - 1
628
+
629
+ prev_output_channel = output_channel
630
+ output_channel = reversed_block_out_channels[i]
631
+ input_channel = reversed_block_out_channels[
632
+ min(i + 1, len(block_out_channels) - 1)
633
+ ]
634
+
635
+ # add upsample block for all BUT final layer
636
+ if not is_final_block:
637
+ add_upsample = True
638
+ self.num_upsamplers += 1
639
+ else:
640
+ add_upsample = False
641
+
642
+ up_block = get_up_block(
643
+ up_block_type,
644
+ num_layers=reversed_layers_per_block[i] + 1,
645
+ transformer_layers_per_block=reversed_transformer_layers_per_block[i],
646
+ in_channels=input_channel,
647
+ out_channels=output_channel,
648
+ prev_output_channel=prev_output_channel,
649
+ temb_channels=blocks_time_embed_dim,
650
+ add_upsample=add_upsample,
651
+ resnet_eps=norm_eps,
652
+ resnet_act_fn=act_fn,
653
+ resolution_idx=i,
654
+ resnet_groups=norm_num_groups,
655
+ cross_attention_dim=reversed_cross_attention_dim[i],
656
+ num_attention_heads=reversed_num_attention_heads[i],
657
+ dual_cross_attention=dual_cross_attention,
658
+ use_linear_projection=use_linear_projection,
659
+ only_cross_attention=only_cross_attention[i],
660
+ upcast_attention=upcast_attention,
661
+ resnet_time_scale_shift=resnet_time_scale_shift,
662
+ attention_type=attention_type,
663
+ attention_head_dim=(
664
+ attention_head_dim[i]
665
+ if attention_head_dim[i] is not None
666
+ else output_channel
667
+ ),
668
+ dropout=dropout,
669
+ )
670
+ self.up_blocks.append(up_block)
671
+ prev_output_channel = output_channel
672
+
673
+ # out
674
+ if norm_num_groups is not None:
675
+ self.conv_norm_out = nn.GroupNorm(
676
+ num_channels=block_out_channels[0],
677
+ num_groups=norm_num_groups,
678
+ eps=norm_eps,
679
+ )
680
+
681
+ self.conv_act = get_activation(act_fn)
682
+
683
+ else:
684
+ self.conv_norm_out = None
685
+ self.conv_act = None
686
+ self.conv_norm_out = None
687
+
688
+ if attention_type in ["gated", "gated-text-image"]:
689
+ positive_len = 768
690
+ if isinstance(cross_attention_dim, int):
691
+ positive_len = cross_attention_dim
692
+ elif isinstance(cross_attention_dim, (tuple, list)):
693
+ positive_len = cross_attention_dim[0]
694
+
695
+ feature_type = "text-only" if attention_type == "gated" else "text-image"
696
+ self.position_net = GLIGENTextBoundingboxProjection(
697
+ positive_len=positive_len,
698
+ out_dim=cross_attention_dim,
699
+ feature_type=feature_type,
700
+ )
701
+
702
+ @property
703
+ def attn_processors(self) -> Dict[str, AttentionProcessor]:
704
+ r"""
705
+ Returns:
706
+ `dict` of attention processors: A dictionary containing all attention processors used in the model with
707
+ indexed by its weight name.
708
+ """
709
+ # set recursively
710
+ processors = {}
711
+
712
+ def fn_recursive_add_processors(
713
+ name: str,
714
+ module: torch.nn.Module,
715
+ processors: Dict[str, AttentionProcessor],
716
+ ):
717
+ if hasattr(module, "get_processor"):
718
+ processors[f"{name}.processor"] = module.get_processor(
719
+ return_deprecated_lora=True
720
+ )
721
+
722
+ for sub_name, child in module.named_children():
723
+ fn_recursive_add_processors(
724
+ f"{name}.{sub_name}", child, processors)
725
+
726
+ return processors
727
+
728
+ for name, module in self.named_children():
729
+ fn_recursive_add_processors(name, module, processors)
730
+
731
+ return processors
732
+
733
+ def set_attn_processor(
734
+ self,
735
+ processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]],
736
+ _remove_lora=False,
737
+ ):
738
+ r"""
739
+ Sets the attention processor to use to compute attention.
740
+
741
+ Parameters:
742
+ processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
743
+ The instantiated processor class or a dictionary of processor classes that will be set as the processor
744
+ for **all** `Attention` layers.
745
+
746
+ If `processor` is a dict, the key needs to define the path to the corresponding cross attention
747
+ processor. This is strongly recommended when setting trainable attention processors.
748
+
749
+ """
750
+ count = len(self.attn_processors.keys())
751
+
752
+ if isinstance(processor, dict) and len(processor) != count:
753
+ raise ValueError(
754
+ f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
755
+ f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
756
+ )
757
+
758
+ def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
759
+ if hasattr(module, "set_processor"):
760
+ if not isinstance(processor, dict):
761
+ module.set_processor(processor, _remove_lora=_remove_lora)
762
+ else:
763
+ module.set_processor(
764
+ processor.pop(f"{name}.processor"), _remove_lora=_remove_lora
765
+ )
766
+
767
+ for sub_name, child in module.named_children():
768
+ fn_recursive_attn_processor(
769
+ f"{name}.{sub_name}", child, processor)
770
+
771
+ for name, module in self.named_children():
772
+ fn_recursive_attn_processor(name, module, processor)
773
+
774
+ def set_default_attn_processor(self):
775
+ """
776
+ Disables custom attention processors and sets the default attention implementation.
777
+ """
778
+ if all(
779
+ proc.__class__ in ADDED_KV_ATTENTION_PROCESSORS
780
+ for proc in self.attn_processors.values()
781
+ ):
782
+ processor = AttnAddedKVProcessor()
783
+ elif all(
784
+ proc.__class__ in CROSS_ATTENTION_PROCESSORS
785
+ for proc in self.attn_processors.values()
786
+ ):
787
+ processor = AttnProcessor()
788
+ else:
789
+ raise ValueError(
790
+ f"Cannot call `set_default_attn_processor` when attention processors are of type {next(iter(self.attn_processors.values()))}"
791
+ )
792
+
793
+ self.set_attn_processor(processor, _remove_lora=True)
794
+
795
+ def set_attention_slice(self, slice_size):
796
+ r"""
797
+ Enable sliced attention computation.
798
+
799
+ When this option is enabled, the attention module splits the input tensor in slices to compute attention in
800
+ several steps. This is useful for saving some memory in exchange for a small decrease in speed.
801
+
802
+ Args:
803
+ slice_size (`str` or `int` or `list(int)`, *optional*, defaults to `"auto"`):
804
+ When `"auto"`, input to the attention heads is halved, so attention is computed in two steps. If
805
+ `"max"`, maximum amount of memory is saved by running only one slice at a time. If a number is
806
+ provided, uses as many slices as `attention_head_dim // slice_size`. In this case, `attention_head_dim`
807
+ must be a multiple of `slice_size`.
808
+ """
809
+ sliceable_head_dims = []
810
+
811
+ def fn_recursive_retrieve_sliceable_dims(module: torch.nn.Module):
812
+ if hasattr(module, "set_attention_slice"):
813
+ sliceable_head_dims.append(module.sliceable_head_dim)
814
+
815
+ for child in module.children():
816
+ fn_recursive_retrieve_sliceable_dims(child)
817
+
818
+ # retrieve number of attention layers
819
+ for module in self.children():
820
+ fn_recursive_retrieve_sliceable_dims(module)
821
+
822
+ num_sliceable_layers = len(sliceable_head_dims)
823
+
824
+ if slice_size == "auto":
825
+ # half the attention head size is usually a good trade-off between
826
+ # speed and memory
827
+ slice_size = [dim // 2 for dim in sliceable_head_dims]
828
+ elif slice_size == "max":
829
+ # make smallest slice possible
830
+ slice_size = num_sliceable_layers * [1]
831
+
832
+ slice_size = (
833
+ num_sliceable_layers * [slice_size]
834
+ if not isinstance(slice_size, list)
835
+ else slice_size
836
+ )
837
+
838
+ if len(slice_size) != len(sliceable_head_dims):
839
+ raise ValueError(
840
+ f"You have provided {len(slice_size)}, but {self.config} has {len(sliceable_head_dims)} different"
841
+ f" attention layers. Make sure to match `len(slice_size)` to be {len(sliceable_head_dims)}."
842
+ )
843
+
844
+ for i, size in enumerate(slice_size):
845
+ dim = sliceable_head_dims[i]
846
+ if size is not None and size > dim:
847
+ raise ValueError(
848
+ f"size {size} has to be smaller or equal to {dim}.")
849
+
850
+ # Recursively walk through all the children.
851
+ # Any children which exposes the set_attention_slice method
852
+ # gets the message
853
+ def fn_recursive_set_attention_slice(
854
+ module: torch.nn.Module, slice_size: List[int]
855
+ ):
856
+ if hasattr(module, "set_attention_slice"):
857
+ module.set_attention_slice(slice_size.pop())
858
+
859
+ for child in module.children():
860
+ fn_recursive_set_attention_slice(child, slice_size)
861
+
862
+ reversed_slice_size = list(reversed(slice_size))
863
+ for module in self.children():
864
+ fn_recursive_set_attention_slice(module, reversed_slice_size)
865
+
866
+ def _set_gradient_checkpointing(self, module, value=False):
867
+ if hasattr(module, "gradient_checkpointing"):
868
+ module.gradient_checkpointing = value
869
+
870
+ def enable_freeu(self, s1, s2, b1, b2):
871
+ r"""Enables the FreeU mechanism from https://arxiv.org/abs/2309.11497.
872
+
873
+ The suffixes after the scaling factors represent the stage blocks where they are being applied.
874
+
875
+ Please refer to the [official repository](https://github.com/ChenyangSi/FreeU) for combinations of values that
876
+ are known to work well for different pipelines such as Stable Diffusion v1, v2, and Stable Diffusion XL.
877
+
878
+ Args:
879
+ s1 (`float`):
880
+ Scaling factor for stage 1 to attenuate the contributions of the skip features. This is done to
881
+ mitigate the "oversmoothing effect" in the enhanced denoising process.
882
+ s2 (`float`):
883
+ Scaling factor for stage 2 to attenuate the contributions of the skip features. This is done to
884
+ mitigate the "oversmoothing effect" in the enhanced denoising process.
885
+ b1 (`float`): Scaling factor for stage 1 to amplify the contributions of backbone features.
886
+ b2 (`float`): Scaling factor for stage 2 to amplify the contributions of backbone features.
887
+ """
888
+ for _, upsample_block in enumerate(self.up_blocks):
889
+ setattr(upsample_block, "s1", s1)
890
+ setattr(upsample_block, "s2", s2)
891
+ setattr(upsample_block, "b1", b1)
892
+ setattr(upsample_block, "b2", b2)
893
+
894
+ def disable_freeu(self):
895
+ """Disables the FreeU mechanism."""
896
+ freeu_keys = {"s1", "s2", "b1", "b2"}
897
+ for _, upsample_block in enumerate(self.up_blocks):
898
+ for k in freeu_keys:
899
+ if (
900
+ hasattr(upsample_block, k)
901
+ or getattr(upsample_block, k, None) is not None
902
+ ):
903
+ setattr(upsample_block, k, None)
904
+
905
+ def forward(
906
+ self,
907
+ sample: torch.FloatTensor,
908
+ timestep: Union[torch.Tensor, float, int],
909
+ encoder_hidden_states: torch.Tensor,
910
+ cond_tensor: torch.FloatTensor=None,
911
+ class_labels: Optional[torch.Tensor] = None,
912
+ timestep_cond: Optional[torch.Tensor] = None,
913
+ attention_mask: Optional[torch.Tensor] = None,
914
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
915
+ added_cond_kwargs: Optional[Dict[str, torch.Tensor]] = None,
916
+ down_block_additional_residuals: Optional[Tuple[torch.Tensor]] = None,
917
+ mid_block_additional_residual: Optional[torch.Tensor] = None,
918
+ down_intrablock_additional_residuals: Optional[Tuple[torch.Tensor]] = None,
919
+ encoder_attention_mask: Optional[torch.Tensor] = None,
920
+ return_dict: bool = True,
921
+ post_process: bool = False,
922
+ ) -> Union[UNet2DConditionOutput, Tuple]:
923
+ r"""
924
+ The [`UNet2DConditionModel`] forward method.
925
+
926
+ Args:
927
+ sample (`torch.FloatTensor`):
928
+ The noisy input tensor with the following shape `(batch, channel, height, width)`.
929
+ timestep (`torch.FloatTensor` or `float` or `int`): The number of timesteps to denoise an input.
930
+ encoder_hidden_states (`torch.FloatTensor`):
931
+ The encoder hidden states with shape `(batch, sequence_length, feature_dim)`.
932
+ class_labels (`torch.Tensor`, *optional*, defaults to `None`):
933
+ Optional class labels for conditioning. Their embeddings will be summed with the timestep embeddings.
934
+ timestep_cond: (`torch.Tensor`, *optional*, defaults to `None`):
935
+ Conditional embeddings for timestep. If provided, the embeddings will be summed with the samples passed
936
+ through the `self.time_embedding` layer to obtain the timestep embeddings.
937
+ attention_mask (`torch.Tensor`, *optional*, defaults to `None`):
938
+ An attention mask of shape `(batch, key_tokens)` is applied to `encoder_hidden_states`. If `1` the mask
939
+ is kept, otherwise if `0` it is discarded. Mask will be converted into a bias, which adds large
940
+ negative values to the attention scores corresponding to "discard" tokens.
941
+ cross_attention_kwargs (`dict`, *optional*):
942
+ A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
943
+ `self.processor` in
944
+ [diffusers.models.attention_processor]
945
+ (https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
946
+ added_cond_kwargs: (`dict`, *optional*):
947
+ A kwargs dictionary containing additional embeddings that if specified are added to the embeddings that
948
+ are passed along to the UNet blocks.
949
+ down_block_additional_residuals: (`tuple` of `torch.Tensor`, *optional*):
950
+ A tuple of tensors that if specified are added to the residuals of down unet blocks.
951
+ mid_block_additional_residual: (`torch.Tensor`, *optional*):
952
+ A tensor that if specified is added to the residual of the middle unet block.
953
+ encoder_attention_mask (`torch.Tensor`):
954
+ A cross-attention mask of shape `(batch, sequence_length)` is applied to `encoder_hidden_states`. If
955
+ `True` the mask is kept, otherwise if `False` it is discarded. Mask will be converted into a bias,
956
+ which adds large negative values to the attention scores corresponding to "discard" tokens.
957
+ return_dict (`bool`, *optional*, defaults to `True`):
958
+ Whether or not to return a [`~models.unet_2d_condition.UNet2DConditionOutput`] instead of a plain
959
+ tuple.
960
+ cross_attention_kwargs (`dict`, *optional*):
961
+ A kwargs dictionary that if specified is passed along to the [`AttnProcessor`].
962
+ added_cond_kwargs: (`dict`, *optional*):
963
+ A kwargs dictionary containin additional embeddings that if specified are added to the embeddings that
964
+ are passed along to the UNet blocks.
965
+ down_block_additional_residuals (`tuple` of `torch.Tensor`, *optional*):
966
+ additional residuals to be added to UNet long skip connections from down blocks to up blocks for
967
+ example from ControlNet side model(s)
968
+ mid_block_additional_residual (`torch.Tensor`, *optional*):
969
+ additional residual to be added to UNet mid block output, for example from ControlNet side model
970
+ down_intrablock_additional_residuals (`tuple` of `torch.Tensor`, *optional*):
971
+ additional residuals to be added within UNet down blocks, for example from T2I-Adapter side model(s)
972
+
973
+ Returns:
974
+ [`~models.unet_2d_condition.UNet2DConditionOutput`] or `tuple`:
975
+ If `return_dict` is True, an [`~models.unet_2d_condition.UNet2DConditionOutput`] is returned, otherwise
976
+ a `tuple` is returned where the first element is the sample tensor.
977
+ """
978
+ # By default samples have to be AT least a multiple of the overall upsampling factor.
979
+ # The overall upsampling factor is equal to 2 ** (# num of upsampling layers).
980
+ # However, the upsampling interpolation output size can be forced to fit any upsampling size
981
+ # on the fly if necessary.
982
+ default_overall_up_factor = 2**self.num_upsamplers
983
+
984
+ # upsample size should be forwarded when sample is not a multiple of `default_overall_up_factor`
985
+ forward_upsample_size = False
986
+ upsample_size = None
987
+
988
+ for dim in sample.shape[-2:]:
989
+ if dim % default_overall_up_factor != 0:
990
+ # Forward upsample size to force interpolation output size.
991
+ forward_upsample_size = True
992
+ break
993
+
994
+ # ensure attention_mask is a bias, and give it a singleton query_tokens dimension
995
+ # expects mask of shape:
996
+ # [batch, key_tokens]
997
+ # adds singleton query_tokens dimension:
998
+ # [batch, 1, key_tokens]
999
+ # this helps to broadcast it as a bias over attention scores, which will be in one of the following shapes:
1000
+ # [batch, heads, query_tokens, key_tokens] (e.g. torch sdp attn)
1001
+ # [batch * heads, query_tokens, key_tokens] (e.g. xformers or classic attn)
1002
+ if attention_mask is not None:
1003
+ # assume that mask is expressed as:
1004
+ # (1 = keep, 0 = discard)
1005
+ # convert mask into a bias that can be added to attention scores:
1006
+ # (keep = +0, discard = -10000.0)
1007
+ attention_mask = (1 - attention_mask.to(sample.dtype)) * -10000.0
1008
+ attention_mask = attention_mask.unsqueeze(1)
1009
+
1010
+ # convert encoder_attention_mask to a bias the same way we do for attention_mask
1011
+ if encoder_attention_mask is not None:
1012
+ encoder_attention_mask = (
1013
+ 1 - encoder_attention_mask.to(sample.dtype)
1014
+ ) * -10000.0
1015
+ encoder_attention_mask = encoder_attention_mask.unsqueeze(1)
1016
+
1017
+ # 0. center input if necessary
1018
+ if self.config.center_input_sample:
1019
+ sample = 2 * sample - 1.0
1020
+
1021
+ # 1. time
1022
+ timesteps = timestep
1023
+ if not torch.is_tensor(timesteps):
1024
+ # TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can
1025
+ # This would be a good case for the `match` statement (Python 3.10+)
1026
+ is_mps = sample.device.type == "mps"
1027
+ if isinstance(timestep, float):
1028
+ dtype = torch.float32 if is_mps else torch.float64
1029
+ else:
1030
+ dtype = torch.int32 if is_mps else torch.int64
1031
+ timesteps = torch.tensor(
1032
+ [timesteps], dtype=dtype, device=sample.device)
1033
+ elif len(timesteps.shape) == 0:
1034
+ timesteps = timesteps[None].to(sample.device)
1035
+
1036
+ # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
1037
+ timesteps = timesteps.expand(sample.shape[0])
1038
+
1039
+ t_emb = self.time_proj(timesteps)
1040
+
1041
+ # `Timesteps` does not contain any weights and will always return f32 tensors
1042
+ # but time_embedding might actually be running in fp16. so we need to cast here.
1043
+ # there might be better ways to encapsulate this.
1044
+ t_emb = t_emb.to(dtype=sample.dtype)
1045
+
1046
+ emb = self.time_embedding(t_emb, timestep_cond)
1047
+ aug_emb = None
1048
+
1049
+ if self.class_embedding is not None:
1050
+ if class_labels is None:
1051
+ raise ValueError(
1052
+ "class_labels should be provided when num_class_embeds > 0"
1053
+ )
1054
+
1055
+ if self.config.class_embed_type == "timestep":
1056
+ class_labels = self.time_proj(class_labels)
1057
+
1058
+ # `Timesteps` does not contain any weights and will always return f32 tensors
1059
+ # there might be better ways to encapsulate this.
1060
+ class_labels = class_labels.to(dtype=sample.dtype)
1061
+
1062
+ class_emb = self.class_embedding(
1063
+ class_labels).to(dtype=sample.dtype)
1064
+
1065
+ if self.config.class_embeddings_concat:
1066
+ emb = torch.cat([emb, class_emb], dim=-1)
1067
+ else:
1068
+ emb = emb + class_emb
1069
+
1070
+ if self.config.addition_embed_type == "text":
1071
+ aug_emb = self.add_embedding(encoder_hidden_states)
1072
+ elif self.config.addition_embed_type == "text_image":
1073
+ # Kandinsky 2.1 - style
1074
+ if "image_embeds" not in added_cond_kwargs:
1075
+ raise ValueError(
1076
+ f"{self.__class__} has the config param `addition_embed_type` set to 'text_image'"
1077
+ "which requires the keyword argument `image_embeds` to be passed in `added_cond_kwargs`"
1078
+ )
1079
+
1080
+ image_embs = added_cond_kwargs.get("image_embeds")
1081
+ text_embs = added_cond_kwargs.get(
1082
+ "text_embeds", encoder_hidden_states)
1083
+ aug_emb = self.add_embedding(text_embs, image_embs)
1084
+ elif self.config.addition_embed_type == "text_time":
1085
+ # SDXL - style
1086
+ if "text_embeds" not in added_cond_kwargs:
1087
+ raise ValueError(
1088
+ f"{self.__class__} has the config param `addition_embed_type` set to 'text_time'"
1089
+ "which requires the keyword argument `text_embeds` to be passed in `added_cond_kwargs`"
1090
+ )
1091
+ text_embeds = added_cond_kwargs.get("text_embeds")
1092
+ if "time_ids" not in added_cond_kwargs:
1093
+ raise ValueError(
1094
+ f"{self.__class__} has the config param `addition_embed_type` set to 'text_time'"
1095
+ "which requires the keyword argument `time_ids` to be passed in `added_cond_kwargs`"
1096
+ )
1097
+ time_ids = added_cond_kwargs.get("time_ids")
1098
+ time_embeds = self.add_time_proj(time_ids.flatten())
1099
+ time_embeds = time_embeds.reshape((text_embeds.shape[0], -1))
1100
+ add_embeds = torch.concat([text_embeds, time_embeds], dim=-1)
1101
+ add_embeds = add_embeds.to(emb.dtype)
1102
+ aug_emb = self.add_embedding(add_embeds)
1103
+ elif self.config.addition_embed_type == "image":
1104
+ # Kandinsky 2.2 - style
1105
+ if "image_embeds" not in added_cond_kwargs:
1106
+ raise ValueError(
1107
+ f"{self.__class__} has the config param `addition_embed_type` set to 'image'"
1108
+ "which requires the keyword argument `image_embeds` to be passed in `added_cond_kwargs`"
1109
+ )
1110
+ image_embs = added_cond_kwargs.get("image_embeds")
1111
+ aug_emb = self.add_embedding(image_embs)
1112
+ elif self.config.addition_embed_type == "image_hint":
1113
+ # Kandinsky 2.2 - style
1114
+ if (
1115
+ "image_embeds" not in added_cond_kwargs
1116
+ or "hint" not in added_cond_kwargs
1117
+ ):
1118
+ raise ValueError(
1119
+ f"{self.__class__} has the config param `addition_embed_type` set to 'image_hint'"
1120
+ "which requires the keyword arguments `image_embeds` and `hint` to be passed in `added_cond_kwargs`"
1121
+ )
1122
+ image_embs = added_cond_kwargs.get("image_embeds")
1123
+ hint = added_cond_kwargs.get("hint")
1124
+ aug_emb, hint = self.add_embedding(image_embs, hint)
1125
+ sample = torch.cat([sample, hint], dim=1)
1126
+
1127
+ emb = emb + aug_emb if aug_emb is not None else emb
1128
+
1129
+ if self.time_embed_act is not None:
1130
+ emb = self.time_embed_act(emb)
1131
+
1132
+ if (
1133
+ self.encoder_hid_proj is not None
1134
+ and self.config.encoder_hid_dim_type == "text_proj"
1135
+ ):
1136
+ encoder_hidden_states = self.encoder_hid_proj(
1137
+ encoder_hidden_states)
1138
+ elif (
1139
+ self.encoder_hid_proj is not None
1140
+ and self.config.encoder_hid_dim_type == "text_image_proj"
1141
+ ):
1142
+ # Kadinsky 2.1 - style
1143
+ if "image_embeds" not in added_cond_kwargs:
1144
+ raise ValueError(
1145
+ f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'text_image_proj'"
1146
+ "which requires the keyword argument `image_embeds` to be passed in `added_conditions`"
1147
+ )
1148
+
1149
+ image_embeds = added_cond_kwargs.get("image_embeds")
1150
+ encoder_hidden_states = self.encoder_hid_proj(
1151
+ encoder_hidden_states, image_embeds
1152
+ )
1153
+ elif (
1154
+ self.encoder_hid_proj is not None
1155
+ and self.config.encoder_hid_dim_type == "image_proj"
1156
+ ):
1157
+ # Kandinsky 2.2 - style
1158
+ if "image_embeds" not in added_cond_kwargs:
1159
+ raise ValueError(
1160
+ f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'image_proj'"
1161
+ "which requires the keyword argument `image_embeds` to be passed in `added_conditions`"
1162
+ )
1163
+ image_embeds = added_cond_kwargs.get("image_embeds")
1164
+ encoder_hidden_states = self.encoder_hid_proj(image_embeds)
1165
+ elif (
1166
+ self.encoder_hid_proj is not None
1167
+ and self.config.encoder_hid_dim_type == "ip_image_proj"
1168
+ ):
1169
+ if "image_embeds" not in added_cond_kwargs:
1170
+ raise ValueError(
1171
+ f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'ip_image_proj'"
1172
+ "which requires the keyword argument `image_embeds` to be passed in `added_conditions`"
1173
+ )
1174
+ image_embeds = added_cond_kwargs.get("image_embeds")
1175
+ image_embeds = self.encoder_hid_proj(image_embeds).to(
1176
+ encoder_hidden_states.dtype
1177
+ )
1178
+ encoder_hidden_states = torch.cat(
1179
+ [encoder_hidden_states, image_embeds], dim=1
1180
+ )
1181
+
1182
+ # 2. pre-process
1183
+ sample = self.conv_in(sample)
1184
+ if cond_tensor is not None:
1185
+ sample = sample + cond_tensor
1186
+
1187
+ # 2.5 GLIGEN position net
1188
+ if (
1189
+ cross_attention_kwargs is not None
1190
+ and cross_attention_kwargs.get("gligen", None) is not None
1191
+ ):
1192
+ cross_attention_kwargs = cross_attention_kwargs.copy()
1193
+ gligen_args = cross_attention_kwargs.pop("gligen")
1194
+ cross_attention_kwargs["gligen"] = {
1195
+ "objs": self.position_net(**gligen_args)
1196
+ }
1197
+
1198
+ # 3. down
1199
+ lora_scale = (
1200
+ cross_attention_kwargs.get("scale", 1.0)
1201
+ if cross_attention_kwargs is not None
1202
+ else 1.0
1203
+ )
1204
+ if USE_PEFT_BACKEND:
1205
+ # weight the lora layers by setting `lora_scale` for each PEFT layer
1206
+ scale_lora_layers(self, lora_scale)
1207
+
1208
+ is_controlnet = (
1209
+ mid_block_additional_residual is not None
1210
+ and down_block_additional_residuals is not None
1211
+ )
1212
+ # using new arg down_intrablock_additional_residuals for T2I-Adapters, to distinguish from controlnets
1213
+ is_adapter = down_intrablock_additional_residuals is not None
1214
+ # maintain backward compatibility for legacy usage, where
1215
+ # T2I-Adapter and ControlNet both use down_block_additional_residuals arg
1216
+ # but can only use one or the other
1217
+ if (
1218
+ not is_adapter
1219
+ and mid_block_additional_residual is None
1220
+ and down_block_additional_residuals is not None
1221
+ ):
1222
+ deprecate(
1223
+ "T2I should not use down_block_additional_residuals",
1224
+ "1.3.0",
1225
+ "Passing intrablock residual connections with `down_block_additional_residuals` is deprecated \
1226
+ and will be removed in diffusers 1.3.0. `down_block_additional_residuals` should only be used \
1227
+ for ControlNet. Please make sure use `down_intrablock_additional_residuals` instead. ",
1228
+ standard_warn=False,
1229
+ )
1230
+ down_intrablock_additional_residuals = down_block_additional_residuals
1231
+ is_adapter = True
1232
+
1233
+ down_block_res_samples = (sample,)
1234
+ for downsample_block in self.down_blocks:
1235
+ if (
1236
+ hasattr(downsample_block, "has_cross_attention")
1237
+ and downsample_block.has_cross_attention
1238
+ ):
1239
+ # For t2i-adapter CrossAttnDownBlock2D
1240
+ additional_residuals = {}
1241
+ if is_adapter and len(down_intrablock_additional_residuals) > 0:
1242
+ additional_residuals["additional_residuals"] = (
1243
+ down_intrablock_additional_residuals.pop(0)
1244
+ )
1245
+
1246
+ sample, res_samples = downsample_block(
1247
+ hidden_states=sample,
1248
+ temb=emb,
1249
+ encoder_hidden_states=encoder_hidden_states,
1250
+ attention_mask=attention_mask,
1251
+ cross_attention_kwargs=cross_attention_kwargs,
1252
+ encoder_attention_mask=encoder_attention_mask,
1253
+ **additional_residuals,
1254
+ )
1255
+ else:
1256
+ sample, res_samples = downsample_block(
1257
+ hidden_states=sample, temb=emb, scale=lora_scale
1258
+ )
1259
+ if is_adapter and len(down_intrablock_additional_residuals) > 0:
1260
+ sample += down_intrablock_additional_residuals.pop(0)
1261
+
1262
+ down_block_res_samples += res_samples
1263
+
1264
+ if is_controlnet:
1265
+ new_down_block_res_samples = ()
1266
+
1267
+ for down_block_res_sample, down_block_additional_residual in zip(
1268
+ down_block_res_samples, down_block_additional_residuals
1269
+ ):
1270
+ down_block_res_sample = (
1271
+ down_block_res_sample + down_block_additional_residual
1272
+ )
1273
+ new_down_block_res_samples = new_down_block_res_samples + (
1274
+ down_block_res_sample,
1275
+ )
1276
+
1277
+ down_block_res_samples = new_down_block_res_samples
1278
+
1279
+ # 4. mid
1280
+ if self.mid_block is not None:
1281
+ if (
1282
+ hasattr(self.mid_block, "has_cross_attention")
1283
+ and self.mid_block.has_cross_attention
1284
+ ):
1285
+ sample = self.mid_block(
1286
+ sample,
1287
+ emb,
1288
+ encoder_hidden_states=encoder_hidden_states,
1289
+ attention_mask=attention_mask,
1290
+ cross_attention_kwargs=cross_attention_kwargs,
1291
+ encoder_attention_mask=encoder_attention_mask,
1292
+ )
1293
+ else:
1294
+ sample = self.mid_block(sample, emb)
1295
+
1296
+ # To support T2I-Adapter-XL
1297
+ if (
1298
+ is_adapter
1299
+ and len(down_intrablock_additional_residuals) > 0
1300
+ and sample.shape == down_intrablock_additional_residuals[0].shape
1301
+ ):
1302
+ sample += down_intrablock_additional_residuals.pop(0)
1303
+
1304
+ if is_controlnet:
1305
+ sample = sample + mid_block_additional_residual
1306
+
1307
+ # 5. up
1308
+ for i, upsample_block in enumerate(self.up_blocks):
1309
+ is_final_block = i == len(self.up_blocks) - 1
1310
+
1311
+ res_samples = down_block_res_samples[-len(upsample_block.resnets):]
1312
+ down_block_res_samples = down_block_res_samples[
1313
+ : -len(upsample_block.resnets)
1314
+ ]
1315
+
1316
+ # if we have not reached the final block and need to forward the
1317
+ # upsample size, we do it here
1318
+ if not is_final_block and forward_upsample_size:
1319
+ upsample_size = down_block_res_samples[-1].shape[2:]
1320
+
1321
+ if (
1322
+ hasattr(upsample_block, "has_cross_attention")
1323
+ and upsample_block.has_cross_attention
1324
+ ):
1325
+ sample = upsample_block(
1326
+ hidden_states=sample,
1327
+ temb=emb,
1328
+ res_hidden_states_tuple=res_samples,
1329
+ encoder_hidden_states=encoder_hidden_states,
1330
+ cross_attention_kwargs=cross_attention_kwargs,
1331
+ upsample_size=upsample_size,
1332
+ attention_mask=attention_mask,
1333
+ encoder_attention_mask=encoder_attention_mask,
1334
+ )
1335
+ else:
1336
+ sample = upsample_block(
1337
+ hidden_states=sample,
1338
+ temb=emb,
1339
+ res_hidden_states_tuple=res_samples,
1340
+ upsample_size=upsample_size,
1341
+ scale=lora_scale,
1342
+ )
1343
+
1344
+ # 6. post-process
1345
+ if post_process:
1346
+ if self.conv_norm_out:
1347
+ sample = self.conv_norm_out(sample)
1348
+ sample = self.conv_act(sample)
1349
+ sample = self.conv_out(sample)
1350
+
1351
+ if USE_PEFT_BACKEND:
1352
+ # remove `lora_scale` from each PEFT layer
1353
+ unscale_lora_layers(self, lora_scale)
1354
+
1355
+ if not return_dict:
1356
+ return (sample,)
1357
+
1358
+ return UNet2DConditionOutput(sample=sample)
1359
+
1360
+ @classmethod
1361
+ def load_change_cross_attention_dim(
1362
+ cls,
1363
+ pretrained_model_path: PathLike,
1364
+ subfolder=None,
1365
+ # unet_additional_kwargs=None,
1366
+ ):
1367
+ """
1368
+ Load or change the cross-attention dimension of a pre-trained model.
1369
+
1370
+ Parameters:
1371
+ pretrained_model_name_or_path (:class:`~typing.Union[str, :class:`~pathlib.Path`]`):
1372
+ The identifier of the pre-trained model or the path to the local folder containing the model.
1373
+ force_download (:class:`~bool`):
1374
+ If True, re-download the model even if it is already cached.
1375
+ resume_download (:class:`~bool`):
1376
+ If True, resume the download of the model if partially downloaded.
1377
+ proxies (:class:`~dict`):
1378
+ A dictionary of proxy servers to use for downloading the model.
1379
+ cache_dir (:class:`~Optional[str]`):
1380
+ The path to the cache directory for storing downloaded models.
1381
+ use_auth_token (:class:`~bool`):
1382
+ If True, use the authentication token for private models.
1383
+ revision (:class:`~str`):
1384
+ The specific model version to use.
1385
+ use_safetensors (:class:`~bool`):
1386
+ If True, use the SafeTensors format for loading the model weights.
1387
+ **kwargs (:class:`~dict`):
1388
+ Additional keyword arguments passed to the model.
1389
+
1390
+ """
1391
+ pretrained_model_path = Path(pretrained_model_path)
1392
+ if subfolder is not None:
1393
+ pretrained_model_path = pretrained_model_path.joinpath(subfolder)
1394
+ config_file = pretrained_model_path / "config.json"
1395
+ if not (config_file.exists() and config_file.is_file()):
1396
+ raise RuntimeError(
1397
+ f"{config_file} does not exist or is not a file")
1398
+
1399
+ unet_config = cls.load_config(config_file)
1400
+ unet_config["cross_attention_dim"] = 1024
1401
+
1402
+ model = cls.from_config(unet_config)
1403
+ # load the vanilla weights
1404
+ if pretrained_model_path.joinpath(SAFETENSORS_WEIGHTS_NAME).exists():
1405
+ logger.debug(
1406
+ f"loading safeTensors weights from {pretrained_model_path} ..."
1407
+ )
1408
+ state_dict = load_file(
1409
+ pretrained_model_path.joinpath(SAFETENSORS_WEIGHTS_NAME), device="cpu"
1410
+ )
1411
+
1412
+ elif pretrained_model_path.joinpath(WEIGHTS_NAME).exists():
1413
+ logger.debug(f"loading weights from {pretrained_model_path} ...")
1414
+ state_dict = torch.load(
1415
+ pretrained_model_path.joinpath(WEIGHTS_NAME),
1416
+ map_location="cpu",
1417
+ weights_only=True,
1418
+ )
1419
+ else:
1420
+ raise FileNotFoundError(
1421
+ f"no weights file found in {pretrained_model_path}")
1422
+
1423
+ model_state_dict = model.state_dict()
1424
+ for k in state_dict:
1425
+ if k in model_state_dict:
1426
+ if state_dict[k].shape != model_state_dict[k].shape:
1427
+ state_dict[k] = model_state_dict[k]
1428
+ # load the weights into the model
1429
+ m, u = model.load_state_dict(state_dict, strict=False)
1430
+ print(m, u)
1431
+
1432
+ return model
hallo/models/unet_3d.py ADDED
@@ -0,0 +1,839 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # pylint: disable=R0801
2
+ # pylint: disable=E1101
3
+ # pylint: disable=R0402
4
+ # pylint: disable=W1203
5
+
6
+ """
7
+ This is the main file for the UNet3DConditionModel, which defines the UNet3D model architecture.
8
+
9
+ The UNet3D model is a 3D convolutional neural network designed for image segmentation and
10
+ other computer vision tasks. It consists of an encoder, a decoder, and skip connections between
11
+ the corresponding layers of the encoder and decoder. The model can handle 3D data and
12
+ performs well on tasks such as image segmentation, object detection, and video analysis.
13
+
14
+ This file contains the necessary imports, the main UNet3DConditionModel class, and its
15
+ methods for setting attention slice, setting gradient checkpointing, setting attention
16
+ processor, and the forward method for model inference.
17
+
18
+ The module provides a comprehensive solution for 3D image segmentation tasks and can be
19
+ easily extended for other computer vision tasks as well.
20
+ """
21
+
22
+ from collections import OrderedDict
23
+ from dataclasses import dataclass
24
+ from os import PathLike
25
+ from pathlib import Path
26
+ from typing import Dict, List, Optional, Tuple, Union
27
+
28
+ import torch
29
+ import torch.nn as nn
30
+ import torch.utils.checkpoint
31
+ from diffusers.configuration_utils import ConfigMixin, register_to_config
32
+ from diffusers.models.attention_processor import AttentionProcessor
33
+ from diffusers.models.embeddings import TimestepEmbedding, Timesteps
34
+ from diffusers.models.modeling_utils import ModelMixin
35
+ from diffusers.utils import (SAFETENSORS_WEIGHTS_NAME, WEIGHTS_NAME,
36
+ BaseOutput, logging)
37
+ from safetensors.torch import load_file
38
+
39
+ from .resnet import InflatedConv3d, InflatedGroupNorm
40
+ from .unet_3d_blocks import (UNetMidBlock3DCrossAttn, get_down_block,
41
+ get_up_block)
42
+
43
+ logger = logging.get_logger(__name__) # pylint: disable=invalid-name
44
+
45
+
46
+ @dataclass
47
+ class UNet3DConditionOutput(BaseOutput):
48
+ """
49
+ Data class that serves as the output of the UNet3DConditionModel.
50
+
51
+ Attributes:
52
+ sample (`torch.FloatTensor`):
53
+ A tensor representing the processed sample. The shape and nature of this tensor will depend on the
54
+ specific configuration of the model and the input data.
55
+ """
56
+ sample: torch.FloatTensor
57
+
58
+
59
+ class UNet3DConditionModel(ModelMixin, ConfigMixin):
60
+ """
61
+ A 3D UNet model designed to handle conditional image and video generation tasks. This model is particularly
62
+ suited for tasks that require the generation of 3D data, such as volumetric medical imaging or 3D video
63
+ generation, while incorporating additional conditioning information.
64
+
65
+ The model consists of an encoder-decoder structure with skip connections. It utilizes a series of downsampling
66
+ and upsampling blocks, with a middle block for further processing. Each block can be customized with different
67
+ types of layers and attention mechanisms.
68
+
69
+ Parameters:
70
+ sample_size (`int`, optional): The size of the input sample.
71
+ in_channels (`int`, defaults to 8): The number of input channels.
72
+ out_channels (`int`, defaults to 8): The number of output channels.
73
+ center_input_sample (`bool`, defaults to False): Whether to center the input sample.
74
+ flip_sin_to_cos (`bool`, defaults to True): Whether to flip the sine to cosine in the time embedding.
75
+ freq_shift (`int`, defaults to 0): The frequency shift for the time embedding.
76
+ down_block_types (`Tuple[str]`): A tuple of strings specifying the types of downsampling blocks.
77
+ mid_block_type (`str`): The type of middle block.
78
+ up_block_types (`Tuple[str]`): A tuple of strings specifying the types of upsampling blocks.
79
+ only_cross_attention (`Union[bool, Tuple[bool]]`): Whether to use only cross-attention.
80
+ block_out_channels (`Tuple[int]`): A tuple of integers specifying the output channels for each block.
81
+ layers_per_block (`int`, defaults to 2): The number of layers per block.
82
+ downsample_padding (`int`, defaults to 1): The padding used in downsampling.
83
+ mid_block_scale_factor (`float`, defaults to 1): The scale factor for the middle block.
84
+ act_fn (`str`, defaults to 'silu'): The activation function to be used.
85
+ norm_num_groups (`int`, defaults to 32): The number of groups for normalization.
86
+ norm_eps (`float`, defaults to 1e-5): The epsilon for normalization.
87
+ cross_attention_dim (`int`, defaults to 1280): The dimension for cross-attention.
88
+ attention_head_dim (`Union[int, Tuple[int]]`): The dimension for attention heads.
89
+ dual_cross_attention (`bool`, defaults to False): Whether to use dual cross-attention.
90
+ use_linear_projection (`bool`, defaults to False): Whether to use linear projection.
91
+ class_embed_type (`str`, optional): The type of class embedding.
92
+ num_class_embeds (`int`, optional): The number of class embeddings.
93
+ upcast_attention (`bool`, defaults to False): Whether to upcast attention.
94
+ resnet_time_scale_shift (`str`, defaults to 'default'): The time scale shift for the ResNet.
95
+ use_inflated_groupnorm (`bool`, defaults to False): Whether to use inflated group normalization.
96
+ use_motion_module (`bool`, defaults to False): Whether to use a motion module.
97
+ motion_module_resolutions (`Tuple[int]`): A tuple of resolutions for the motion module.
98
+ motion_module_mid_block (`bool`, defaults to False): Whether to use a motion module in the middle block.
99
+ motion_module_decoder_only (`bool`, defaults to False): Whether to use the motion module only in the decoder.
100
+ motion_module_type (`str`, optional): The type of motion module.
101
+ motion_module_kwargs (`dict`): Keyword arguments for the motion module.
102
+ unet_use_cross_frame_attention (`bool`, optional): Whether to use cross-frame attention in the UNet.
103
+ unet_use_temporal_attention (`bool`, optional): Whether to use temporal attention in the UNet.
104
+ use_audio_module (`bool`, defaults to False): Whether to use an audio module.
105
+ audio_attention_dim (`int`, defaults to 768): The dimension for audio attention.
106
+
107
+ The model supports various features such as gradient checkpointing, attention processors, and sliced attention
108
+ computation, making it flexible and efficient for different computational requirements and use cases.
109
+
110
+ The forward method of the model accepts a sample, timestep, and encoder hidden states as input, and it returns
111
+ the processed sample as output. The method also supports additional conditioning information such as class
112
+ labels, audio embeddings, and masks for specialized tasks.
113
+
114
+ The from_pretrained_2d class method allows loading a pre-trained 2D UNet model and adapting it for 3D tasks by
115
+ incorporating motion modules and other 3D specific features.
116
+ """
117
+
118
+ _supports_gradient_checkpointing = True
119
+
120
+ @register_to_config
121
+ def __init__(
122
+ self,
123
+ sample_size: Optional[int] = None,
124
+ in_channels: int = 8,
125
+ out_channels: int = 8,
126
+ flip_sin_to_cos: bool = True,
127
+ freq_shift: int = 0,
128
+ down_block_types: Tuple[str] = (
129
+ "CrossAttnDownBlock3D",
130
+ "CrossAttnDownBlock3D",
131
+ "CrossAttnDownBlock3D",
132
+ "DownBlock3D",
133
+ ),
134
+ mid_block_type: str = "UNetMidBlock3DCrossAttn",
135
+ up_block_types: Tuple[str] = (
136
+ "UpBlock3D",
137
+ "CrossAttnUpBlock3D",
138
+ "CrossAttnUpBlock3D",
139
+ "CrossAttnUpBlock3D",
140
+ ),
141
+ only_cross_attention: Union[bool, Tuple[bool]] = False,
142
+ block_out_channels: Tuple[int] = (320, 640, 1280, 1280),
143
+ layers_per_block: int = 2,
144
+ downsample_padding: int = 1,
145
+ mid_block_scale_factor: float = 1,
146
+ act_fn: str = "silu",
147
+ norm_num_groups: int = 32,
148
+ norm_eps: float = 1e-5,
149
+ cross_attention_dim: int = 1280,
150
+ attention_head_dim: Union[int, Tuple[int]] = 8,
151
+ dual_cross_attention: bool = False,
152
+ use_linear_projection: bool = False,
153
+ class_embed_type: Optional[str] = None,
154
+ num_class_embeds: Optional[int] = None,
155
+ upcast_attention: bool = False,
156
+ resnet_time_scale_shift: str = "default",
157
+ use_inflated_groupnorm=False,
158
+ # Additional
159
+ use_motion_module=False,
160
+ motion_module_resolutions=(1, 2, 4, 8),
161
+ motion_module_mid_block=False,
162
+ motion_module_decoder_only=False,
163
+ motion_module_type=None,
164
+ motion_module_kwargs=None,
165
+ unet_use_cross_frame_attention=None,
166
+ unet_use_temporal_attention=None,
167
+ # audio
168
+ use_audio_module=False,
169
+ audio_attention_dim=768,
170
+ stack_enable_blocks_name=None,
171
+ stack_enable_blocks_depth=None,
172
+ ):
173
+ super().__init__()
174
+
175
+ self.sample_size = sample_size
176
+ time_embed_dim = block_out_channels[0] * 4
177
+
178
+ # input
179
+ self.conv_in = InflatedConv3d(
180
+ in_channels, block_out_channels[0], kernel_size=3, padding=(1, 1)
181
+ )
182
+
183
+ # time
184
+ self.time_proj = Timesteps(
185
+ block_out_channels[0], flip_sin_to_cos, freq_shift)
186
+ timestep_input_dim = block_out_channels[0]
187
+
188
+ self.time_embedding = TimestepEmbedding(
189
+ timestep_input_dim, time_embed_dim)
190
+
191
+ # class embedding
192
+ if class_embed_type is None and num_class_embeds is not None:
193
+ self.class_embedding = nn.Embedding(
194
+ num_class_embeds, time_embed_dim)
195
+ elif class_embed_type == "timestep":
196
+ self.class_embedding = TimestepEmbedding(
197
+ timestep_input_dim, time_embed_dim)
198
+ elif class_embed_type == "identity":
199
+ self.class_embedding = nn.Identity(time_embed_dim, time_embed_dim)
200
+ else:
201
+ self.class_embedding = None
202
+
203
+ self.down_blocks = nn.ModuleList([])
204
+ self.mid_block = None
205
+ self.up_blocks = nn.ModuleList([])
206
+
207
+ if isinstance(only_cross_attention, bool):
208
+ only_cross_attention = [
209
+ only_cross_attention] * len(down_block_types)
210
+
211
+ if isinstance(attention_head_dim, int):
212
+ attention_head_dim = (attention_head_dim,) * len(down_block_types)
213
+
214
+ # down
215
+ output_channel = block_out_channels[0]
216
+ for i, down_block_type in enumerate(down_block_types):
217
+ res = 2**i
218
+ input_channel = output_channel
219
+ output_channel = block_out_channels[i]
220
+ is_final_block = i == len(block_out_channels) - 1
221
+
222
+ down_block = get_down_block(
223
+ down_block_type,
224
+ num_layers=layers_per_block,
225
+ in_channels=input_channel,
226
+ out_channels=output_channel,
227
+ temb_channels=time_embed_dim,
228
+ add_downsample=not is_final_block,
229
+ resnet_eps=norm_eps,
230
+ resnet_act_fn=act_fn,
231
+ resnet_groups=norm_num_groups,
232
+ cross_attention_dim=cross_attention_dim,
233
+ attn_num_head_channels=attention_head_dim[i],
234
+ downsample_padding=downsample_padding,
235
+ dual_cross_attention=dual_cross_attention,
236
+ use_linear_projection=use_linear_projection,
237
+ only_cross_attention=only_cross_attention[i],
238
+ upcast_attention=upcast_attention,
239
+ resnet_time_scale_shift=resnet_time_scale_shift,
240
+ unet_use_cross_frame_attention=unet_use_cross_frame_attention,
241
+ unet_use_temporal_attention=unet_use_temporal_attention,
242
+ use_inflated_groupnorm=use_inflated_groupnorm,
243
+ use_motion_module=use_motion_module
244
+ and (res in motion_module_resolutions)
245
+ and (not motion_module_decoder_only),
246
+ motion_module_type=motion_module_type,
247
+ motion_module_kwargs=motion_module_kwargs,
248
+ use_audio_module=use_audio_module,
249
+ audio_attention_dim=audio_attention_dim,
250
+ depth=i,
251
+ stack_enable_blocks_name=stack_enable_blocks_name,
252
+ stack_enable_blocks_depth=stack_enable_blocks_depth,
253
+ )
254
+ self.down_blocks.append(down_block)
255
+
256
+ # mid
257
+ if mid_block_type == "UNetMidBlock3DCrossAttn":
258
+ self.mid_block = UNetMidBlock3DCrossAttn(
259
+ in_channels=block_out_channels[-1],
260
+ temb_channels=time_embed_dim,
261
+ resnet_eps=norm_eps,
262
+ resnet_act_fn=act_fn,
263
+ output_scale_factor=mid_block_scale_factor,
264
+ resnet_time_scale_shift=resnet_time_scale_shift,
265
+ cross_attention_dim=cross_attention_dim,
266
+ attn_num_head_channels=attention_head_dim[-1],
267
+ resnet_groups=norm_num_groups,
268
+ dual_cross_attention=dual_cross_attention,
269
+ use_linear_projection=use_linear_projection,
270
+ upcast_attention=upcast_attention,
271
+ unet_use_cross_frame_attention=unet_use_cross_frame_attention,
272
+ unet_use_temporal_attention=unet_use_temporal_attention,
273
+ use_inflated_groupnorm=use_inflated_groupnorm,
274
+ use_motion_module=use_motion_module and motion_module_mid_block,
275
+ motion_module_type=motion_module_type,
276
+ motion_module_kwargs=motion_module_kwargs,
277
+ use_audio_module=use_audio_module,
278
+ audio_attention_dim=audio_attention_dim,
279
+ depth=3,
280
+ stack_enable_blocks_name=stack_enable_blocks_name,
281
+ stack_enable_blocks_depth=stack_enable_blocks_depth,
282
+ )
283
+ else:
284
+ raise ValueError(f"unknown mid_block_type : {mid_block_type}")
285
+
286
+ # count how many layers upsample the videos
287
+ self.num_upsamplers = 0
288
+
289
+ # up
290
+ reversed_block_out_channels = list(reversed(block_out_channels))
291
+ reversed_attention_head_dim = list(reversed(attention_head_dim))
292
+ only_cross_attention = list(reversed(only_cross_attention))
293
+ output_channel = reversed_block_out_channels[0]
294
+ for i, up_block_type in enumerate(up_block_types):
295
+ res = 2 ** (3 - i)
296
+ is_final_block = i == len(block_out_channels) - 1
297
+
298
+ prev_output_channel = output_channel
299
+ output_channel = reversed_block_out_channels[i]
300
+ input_channel = reversed_block_out_channels[
301
+ min(i + 1, len(block_out_channels) - 1)
302
+ ]
303
+
304
+ # add upsample block for all BUT final layer
305
+ if not is_final_block:
306
+ add_upsample = True
307
+ self.num_upsamplers += 1
308
+ else:
309
+ add_upsample = False
310
+
311
+ up_block = get_up_block(
312
+ up_block_type,
313
+ num_layers=layers_per_block + 1,
314
+ in_channels=input_channel,
315
+ out_channels=output_channel,
316
+ prev_output_channel=prev_output_channel,
317
+ temb_channels=time_embed_dim,
318
+ add_upsample=add_upsample,
319
+ resnet_eps=norm_eps,
320
+ resnet_act_fn=act_fn,
321
+ resnet_groups=norm_num_groups,
322
+ cross_attention_dim=cross_attention_dim,
323
+ attn_num_head_channels=reversed_attention_head_dim[i],
324
+ dual_cross_attention=dual_cross_attention,
325
+ use_linear_projection=use_linear_projection,
326
+ only_cross_attention=only_cross_attention[i],
327
+ upcast_attention=upcast_attention,
328
+ resnet_time_scale_shift=resnet_time_scale_shift,
329
+ unet_use_cross_frame_attention=unet_use_cross_frame_attention,
330
+ unet_use_temporal_attention=unet_use_temporal_attention,
331
+ use_inflated_groupnorm=use_inflated_groupnorm,
332
+ use_motion_module=use_motion_module
333
+ and (res in motion_module_resolutions),
334
+ motion_module_type=motion_module_type,
335
+ motion_module_kwargs=motion_module_kwargs,
336
+ use_audio_module=use_audio_module,
337
+ audio_attention_dim=audio_attention_dim,
338
+ depth=3-i,
339
+ stack_enable_blocks_name=stack_enable_blocks_name,
340
+ stack_enable_blocks_depth=stack_enable_blocks_depth,
341
+ )
342
+ self.up_blocks.append(up_block)
343
+ prev_output_channel = output_channel
344
+
345
+ # out
346
+ if use_inflated_groupnorm:
347
+ self.conv_norm_out = InflatedGroupNorm(
348
+ num_channels=block_out_channels[0],
349
+ num_groups=norm_num_groups,
350
+ eps=norm_eps,
351
+ )
352
+ else:
353
+ self.conv_norm_out = nn.GroupNorm(
354
+ num_channels=block_out_channels[0],
355
+ num_groups=norm_num_groups,
356
+ eps=norm_eps,
357
+ )
358
+ self.conv_act = nn.SiLU()
359
+ self.conv_out = InflatedConv3d(
360
+ block_out_channels[0], out_channels, kernel_size=3, padding=1
361
+ )
362
+
363
+ @property
364
+ # Copied from diffusers.models.unet_2d_condition.UNet2DConditionModel.attn_processors
365
+ def attn_processors(self) -> Dict[str, AttentionProcessor]:
366
+ r"""
367
+ Returns:
368
+ `dict` of attention processors: A dictionary containing all attention processors used in the model with
369
+ indexed by its weight name.
370
+ """
371
+ # set recursively
372
+ processors = {}
373
+
374
+ def fn_recursive_add_processors(
375
+ name: str,
376
+ module: torch.nn.Module,
377
+ processors: Dict[str, AttentionProcessor],
378
+ ):
379
+ if hasattr(module, "set_processor"):
380
+ processors[f"{name}.processor"] = module.processor
381
+
382
+ for sub_name, child in module.named_children():
383
+ if "temporal_transformer" not in sub_name:
384
+ fn_recursive_add_processors(
385
+ f"{name}.{sub_name}", child, processors)
386
+
387
+ return processors
388
+
389
+ for name, module in self.named_children():
390
+ if "temporal_transformer" not in name:
391
+ fn_recursive_add_processors(name, module, processors)
392
+
393
+ return processors
394
+
395
+ def set_attention_slice(self, slice_size):
396
+ r"""
397
+ Enable sliced attention computation.
398
+
399
+ When this option is enabled, the attention module will split the input tensor in slices, to compute attention
400
+ in several steps. This is useful to save some memory in exchange for a small speed decrease.
401
+
402
+ Args:
403
+ slice_size (`str` or `int` or `list(int)`, *optional*, defaults to `"auto"`):
404
+ When `"auto"`, halves the input to the attention heads, so attention will be computed in two steps. If
405
+ `"max"`, maxium amount of memory will be saved by running only one slice at a time. If a number is
406
+ provided, uses as many slices as `attention_head_dim // slice_size`. In this case, `attention_head_dim`
407
+ must be a multiple of `slice_size`.
408
+ """
409
+ sliceable_head_dims = []
410
+
411
+ def fn_recursive_retrieve_slicable_dims(module: torch.nn.Module):
412
+ if hasattr(module, "set_attention_slice"):
413
+ sliceable_head_dims.append(module.sliceable_head_dim)
414
+
415
+ for child in module.children():
416
+ fn_recursive_retrieve_slicable_dims(child)
417
+
418
+ # retrieve number of attention layers
419
+ for module in self.children():
420
+ fn_recursive_retrieve_slicable_dims(module)
421
+
422
+ num_slicable_layers = len(sliceable_head_dims)
423
+
424
+ if slice_size == "auto":
425
+ # half the attention head size is usually a good trade-off between
426
+ # speed and memory
427
+ slice_size = [dim // 2 for dim in sliceable_head_dims]
428
+ elif slice_size == "max":
429
+ # make smallest slice possible
430
+ slice_size = num_slicable_layers * [1]
431
+
432
+ slice_size = (
433
+ num_slicable_layers * [slice_size]
434
+ if not isinstance(slice_size, list)
435
+ else slice_size
436
+ )
437
+
438
+ if len(slice_size) != len(sliceable_head_dims):
439
+ raise ValueError(
440
+ f"You have provided {len(slice_size)}, but {self.config} has {len(sliceable_head_dims)} different"
441
+ f" attention layers. Make sure to match `len(slice_size)` to be {len(sliceable_head_dims)}."
442
+ )
443
+
444
+ for i, size in enumerate(slice_size):
445
+ dim = sliceable_head_dims[i]
446
+ if size is not None and size > dim:
447
+ raise ValueError(
448
+ f"size {size} has to be smaller or equal to {dim}.")
449
+
450
+ # Recursively walk through all the children.
451
+ # Any children which exposes the set_attention_slice method
452
+ # gets the message
453
+ def fn_recursive_set_attention_slice(
454
+ module: torch.nn.Module, slice_size: List[int]
455
+ ):
456
+ if hasattr(module, "set_attention_slice"):
457
+ module.set_attention_slice(slice_size.pop())
458
+
459
+ for child in module.children():
460
+ fn_recursive_set_attention_slice(child, slice_size)
461
+
462
+ reversed_slice_size = list(reversed(slice_size))
463
+ for module in self.children():
464
+ fn_recursive_set_attention_slice(module, reversed_slice_size)
465
+
466
+ def _set_gradient_checkpointing(self, module, value=False):
467
+ if hasattr(module, "gradient_checkpointing"):
468
+ module.gradient_checkpointing = value
469
+
470
+ # Copied from diffusers.models.unet_2d_condition.UNet2DConditionModel.set_attn_processor
471
+ def set_attn_processor(
472
+ self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]
473
+ ):
474
+ r"""
475
+ Sets the attention processor to use to compute attention.
476
+
477
+ Parameters:
478
+ processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
479
+ The instantiated processor class or a dictionary of processor classes that will be set as the processor
480
+ for **all** `Attention` layers.
481
+
482
+ If `processor` is a dict, the key needs to define the path to the corresponding cross attention
483
+ processor. This is strongly recommended when setting trainable attention processors.
484
+
485
+ """
486
+ count = len(self.attn_processors.keys())
487
+
488
+ if isinstance(processor, dict) and len(processor) != count:
489
+ raise ValueError(
490
+ f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
491
+ f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
492
+ )
493
+
494
+ def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
495
+ if hasattr(module, "set_processor"):
496
+ if not isinstance(processor, dict):
497
+ module.set_processor(processor)
498
+ else:
499
+ module.set_processor(processor.pop(f"{name}.processor"))
500
+
501
+ for sub_name, child in module.named_children():
502
+ if "temporal_transformer" not in sub_name:
503
+ fn_recursive_attn_processor(
504
+ f"{name}.{sub_name}", child, processor)
505
+
506
+ for name, module in self.named_children():
507
+ if "temporal_transformer" not in name:
508
+ fn_recursive_attn_processor(name, module, processor)
509
+
510
+ def forward(
511
+ self,
512
+ sample: torch.FloatTensor,
513
+ timestep: Union[torch.Tensor, float, int],
514
+ encoder_hidden_states: torch.Tensor,
515
+ audio_embedding: Optional[torch.Tensor] = None,
516
+ class_labels: Optional[torch.Tensor] = None,
517
+ mask_cond_fea: Optional[torch.Tensor] = None,
518
+ attention_mask: Optional[torch.Tensor] = None,
519
+ full_mask: Optional[torch.Tensor] = None,
520
+ face_mask: Optional[torch.Tensor] = None,
521
+ lip_mask: Optional[torch.Tensor] = None,
522
+ motion_scale: Optional[torch.Tensor] = None,
523
+ down_block_additional_residuals: Optional[Tuple[torch.Tensor]] = None,
524
+ mid_block_additional_residual: Optional[torch.Tensor] = None,
525
+ return_dict: bool = True,
526
+ # start: bool = False,
527
+ ) -> Union[UNet3DConditionOutput, Tuple]:
528
+ r"""
529
+ Args:
530
+ sample (`torch.FloatTensor`): (batch, channel, height, width) noisy inputs tensor
531
+ timestep (`torch.FloatTensor` or `float` or `int`): (batch) timesteps
532
+ encoder_hidden_states (`torch.FloatTensor`): (batch, sequence_length, feature_dim) encoder hidden states
533
+ return_dict (`bool`, *optional*, defaults to `True`):
534
+ Whether or not to return a [`models.unet_2d_condition.UNet2DConditionOutput`] instead of a plain tuple.
535
+
536
+ Returns:
537
+ [`~models.unet_2d_condition.UNet2DConditionOutput`] or `tuple`:
538
+ [`~models.unet_2d_condition.UNet2DConditionOutput`] if `return_dict` is True, otherwise a `tuple`. When
539
+ returning a tuple, the first element is the sample tensor.
540
+ """
541
+ # By default samples have to be AT least a multiple of the overall upsampling factor.
542
+ # The overall upsampling factor is equal to 2 ** (# num of upsampling layears).
543
+ # However, the upsampling interpolation output size can be forced to fit any upsampling size
544
+ # on the fly if necessary.
545
+ default_overall_up_factor = 2**self.num_upsamplers
546
+
547
+ # upsample size should be forwarded when sample is not a multiple of `default_overall_up_factor`
548
+ forward_upsample_size = False
549
+ upsample_size = None
550
+
551
+ if any(s % default_overall_up_factor != 0 for s in sample.shape[-2:]):
552
+ logger.info(
553
+ "Forward upsample size to force interpolation output size.")
554
+ forward_upsample_size = True
555
+
556
+ # prepare attention_mask
557
+ if attention_mask is not None:
558
+ attention_mask = (1 - attention_mask.to(sample.dtype)) * -10000.0
559
+ attention_mask = attention_mask.unsqueeze(1)
560
+
561
+ # center input if necessary
562
+ if self.config.center_input_sample:
563
+ sample = 2 * sample - 1.0
564
+
565
+ # time
566
+ timesteps = timestep
567
+ if not torch.is_tensor(timesteps):
568
+ # This would be a good case for the `match` statement (Python 3.10+)
569
+ is_mps = sample.device.type == "mps"
570
+ if isinstance(timestep, float):
571
+ dtype = torch.float32 if is_mps else torch.float64
572
+ else:
573
+ dtype = torch.int32 if is_mps else torch.int64
574
+ timesteps = torch.tensor(
575
+ [timesteps], dtype=dtype, device=sample.device)
576
+ elif len(timesteps.shape) == 0:
577
+ timesteps = timesteps[None].to(sample.device)
578
+
579
+ # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
580
+ timesteps = timesteps.expand(sample.shape[0])
581
+
582
+ t_emb = self.time_proj(timesteps)
583
+
584
+ # timesteps does not contain any weights and will always return f32 tensors
585
+ # but time_embedding might actually be running in fp16. so we need to cast here.
586
+ # there might be better ways to encapsulate this.
587
+ t_emb = t_emb.to(dtype=self.dtype)
588
+ emb = self.time_embedding(t_emb)
589
+
590
+ if self.class_embedding is not None:
591
+ if class_labels is None:
592
+ raise ValueError(
593
+ "class_labels should be provided when num_class_embeds > 0"
594
+ )
595
+
596
+ if self.config.class_embed_type == "timestep":
597
+ class_labels = self.time_proj(class_labels)
598
+
599
+ class_emb = self.class_embedding(class_labels).to(dtype=self.dtype)
600
+ emb = emb + class_emb
601
+
602
+ # pre-process
603
+ sample = self.conv_in(sample)
604
+ if mask_cond_fea is not None:
605
+ sample = sample + mask_cond_fea
606
+
607
+ # down
608
+ down_block_res_samples = (sample,)
609
+ for downsample_block in self.down_blocks:
610
+ if (
611
+ hasattr(downsample_block, "has_cross_attention")
612
+ and downsample_block.has_cross_attention
613
+ ):
614
+ sample, res_samples = downsample_block(
615
+ hidden_states=sample,
616
+ temb=emb,
617
+ encoder_hidden_states=encoder_hidden_states,
618
+ attention_mask=attention_mask,
619
+ full_mask=full_mask,
620
+ face_mask=face_mask,
621
+ lip_mask=lip_mask,
622
+ audio_embedding=audio_embedding,
623
+ motion_scale=motion_scale,
624
+ )
625
+ # print("")
626
+ else:
627
+ sample, res_samples = downsample_block(
628
+ hidden_states=sample,
629
+ temb=emb,
630
+ encoder_hidden_states=encoder_hidden_states,
631
+ # audio_embedding=audio_embedding,
632
+ )
633
+ # print("")
634
+
635
+ down_block_res_samples += res_samples
636
+
637
+ if down_block_additional_residuals is not None:
638
+ new_down_block_res_samples = ()
639
+
640
+ for down_block_res_sample, down_block_additional_residual in zip(
641
+ down_block_res_samples, down_block_additional_residuals
642
+ ):
643
+ down_block_res_sample = (
644
+ down_block_res_sample + down_block_additional_residual
645
+ )
646
+ new_down_block_res_samples += (down_block_res_sample,)
647
+
648
+ down_block_res_samples = new_down_block_res_samples
649
+
650
+ # mid
651
+ sample = self.mid_block(
652
+ sample,
653
+ emb,
654
+ encoder_hidden_states=encoder_hidden_states,
655
+ attention_mask=attention_mask,
656
+ full_mask=full_mask,
657
+ face_mask=face_mask,
658
+ lip_mask=lip_mask,
659
+ audio_embedding=audio_embedding,
660
+ motion_scale=motion_scale,
661
+ )
662
+
663
+ if mid_block_additional_residual is not None:
664
+ sample = sample + mid_block_additional_residual
665
+
666
+ # up
667
+ for i, upsample_block in enumerate(self.up_blocks):
668
+ is_final_block = i == len(self.up_blocks) - 1
669
+
670
+ res_samples = down_block_res_samples[-len(upsample_block.resnets):]
671
+ down_block_res_samples = down_block_res_samples[
672
+ : -len(upsample_block.resnets)
673
+ ]
674
+
675
+ # if we have not reached the final block and need to forward the
676
+ # upsample size, we do it here
677
+ if not is_final_block and forward_upsample_size:
678
+ upsample_size = down_block_res_samples[-1].shape[2:]
679
+
680
+ if (
681
+ hasattr(upsample_block, "has_cross_attention")
682
+ and upsample_block.has_cross_attention
683
+ ):
684
+ sample = upsample_block(
685
+ hidden_states=sample,
686
+ temb=emb,
687
+ res_hidden_states_tuple=res_samples,
688
+ encoder_hidden_states=encoder_hidden_states,
689
+ upsample_size=upsample_size,
690
+ attention_mask=attention_mask,
691
+ full_mask=full_mask,
692
+ face_mask=face_mask,
693
+ lip_mask=lip_mask,
694
+ audio_embedding=audio_embedding,
695
+ motion_scale=motion_scale,
696
+ )
697
+ else:
698
+ sample = upsample_block(
699
+ hidden_states=sample,
700
+ temb=emb,
701
+ res_hidden_states_tuple=res_samples,
702
+ upsample_size=upsample_size,
703
+ encoder_hidden_states=encoder_hidden_states,
704
+ # audio_embedding=audio_embedding,
705
+ )
706
+
707
+ # post-process
708
+ sample = self.conv_norm_out(sample)
709
+ sample = self.conv_act(sample)
710
+ sample = self.conv_out(sample)
711
+
712
+ if not return_dict:
713
+ return (sample,)
714
+
715
+ return UNet3DConditionOutput(sample=sample)
716
+
717
+ @classmethod
718
+ def from_pretrained_2d(
719
+ cls,
720
+ pretrained_model_path: PathLike,
721
+ motion_module_path: PathLike,
722
+ subfolder=None,
723
+ unet_additional_kwargs=None,
724
+ mm_zero_proj_out=False,
725
+ use_landmark=True,
726
+ ):
727
+ """
728
+ Load a pre-trained 2D UNet model from a given directory.
729
+
730
+ Parameters:
731
+ pretrained_model_path (`str` or `PathLike`):
732
+ Path to the directory containing a pre-trained 2D UNet model.
733
+ dtype (`torch.dtype`, *optional*):
734
+ The data type of the loaded model. If not provided, the default data type is used.
735
+ device (`torch.device`, *optional*):
736
+ The device on which the loaded model will be placed. If not provided, the default device is used.
737
+ **kwargs (`Any`):
738
+ Additional keyword arguments passed to the model.
739
+
740
+ Returns:
741
+ `UNet3DConditionModel`:
742
+ The loaded 2D UNet model.
743
+ """
744
+ pretrained_model_path = Path(pretrained_model_path)
745
+ motion_module_path = Path(motion_module_path)
746
+ if subfolder is not None:
747
+ pretrained_model_path = pretrained_model_path.joinpath(subfolder)
748
+ logger.info(
749
+ f"loaded temporal unet's pretrained weights from {pretrained_model_path} ..."
750
+ )
751
+
752
+ config_file = pretrained_model_path / "config.json"
753
+ if not (config_file.exists() and config_file.is_file()):
754
+ raise RuntimeError(
755
+ f"{config_file} does not exist or is not a file")
756
+
757
+ unet_config = cls.load_config(config_file)
758
+ unet_config["_class_name"] = cls.__name__
759
+ unet_config["down_block_types"] = [
760
+ "CrossAttnDownBlock3D",
761
+ "CrossAttnDownBlock3D",
762
+ "CrossAttnDownBlock3D",
763
+ "DownBlock3D",
764
+ ]
765
+ unet_config["up_block_types"] = [
766
+ "UpBlock3D",
767
+ "CrossAttnUpBlock3D",
768
+ "CrossAttnUpBlock3D",
769
+ "CrossAttnUpBlock3D",
770
+ ]
771
+ unet_config["mid_block_type"] = "UNetMidBlock3DCrossAttn"
772
+ if use_landmark:
773
+ unet_config["in_channels"] = 8
774
+ unet_config["out_channels"] = 8
775
+
776
+ model = cls.from_config(unet_config, **unet_additional_kwargs)
777
+ # load the vanilla weights
778
+ if pretrained_model_path.joinpath(SAFETENSORS_WEIGHTS_NAME).exists():
779
+ logger.debug(
780
+ f"loading safeTensors weights from {pretrained_model_path} ..."
781
+ )
782
+ state_dict = load_file(
783
+ pretrained_model_path.joinpath(SAFETENSORS_WEIGHTS_NAME), device="cpu"
784
+ )
785
+
786
+ elif pretrained_model_path.joinpath(WEIGHTS_NAME).exists():
787
+ logger.debug(f"loading weights from {pretrained_model_path} ...")
788
+ state_dict = torch.load(
789
+ pretrained_model_path.joinpath(WEIGHTS_NAME),
790
+ map_location="cpu",
791
+ weights_only=True,
792
+ )
793
+ else:
794
+ raise FileNotFoundError(
795
+ f"no weights file found in {pretrained_model_path}")
796
+
797
+ # load the motion module weights
798
+ if motion_module_path.exists() and motion_module_path.is_file():
799
+ if motion_module_path.suffix.lower() in [".pth", ".pt", ".ckpt"]:
800
+ print(
801
+ f"Load motion module params from {motion_module_path}")
802
+ motion_state_dict = torch.load(
803
+ motion_module_path, map_location="cpu", weights_only=True
804
+ )
805
+ elif motion_module_path.suffix.lower() == ".safetensors":
806
+ motion_state_dict = load_file(motion_module_path, device="cpu")
807
+ else:
808
+ raise RuntimeError(
809
+ f"unknown file format for motion module weights: {motion_module_path.suffix}"
810
+ )
811
+ if mm_zero_proj_out:
812
+ logger.info(
813
+ "Zero initialize proj_out layers in motion module...")
814
+ new_motion_state_dict = OrderedDict()
815
+ for k in motion_state_dict:
816
+ if "proj_out" in k:
817
+ continue
818
+ new_motion_state_dict[k] = motion_state_dict[k]
819
+ motion_state_dict = new_motion_state_dict
820
+
821
+ # merge the state dicts
822
+ state_dict.update(motion_state_dict)
823
+
824
+ model_state_dict = model.state_dict()
825
+ for k in state_dict:
826
+ if k in model_state_dict:
827
+ if state_dict[k].shape != model_state_dict[k].shape:
828
+ state_dict[k] = model_state_dict[k]
829
+ # load the weights into the model
830
+ m, u = model.load_state_dict(state_dict, strict=False)
831
+ logger.debug(
832
+ f"### missing keys: {len(m)}; \n### unexpected keys: {len(u)};")
833
+
834
+ params = [
835
+ p.numel() if "temporal" in n else 0 for n, p in model.named_parameters()
836
+ ]
837
+ logger.info(f"Loaded {sum(params) / 1e6}M-parameter motion module")
838
+
839
+ return model
hallo/models/unet_3d_blocks.py ADDED
@@ -0,0 +1,1401 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # pylint: disable=R0801
2
+ # src/models/unet_3d_blocks.py
3
+
4
+ """
5
+ This module defines various 3D UNet blocks used in the video model.
6
+
7
+ The blocks include:
8
+ - UNetMidBlock3DCrossAttn: The middle block of the UNet with cross attention.
9
+ - CrossAttnDownBlock3D: The downsampling block with cross attention.
10
+ - DownBlock3D: The standard downsampling block without cross attention.
11
+ - CrossAttnUpBlock3D: The upsampling block with cross attention.
12
+ - UpBlock3D: The standard upsampling block without cross attention.
13
+
14
+ These blocks are used to construct the 3D UNet architecture for video-related tasks.
15
+ """
16
+
17
+ import torch
18
+ from einops import rearrange
19
+ from torch import nn
20
+
21
+ from .motion_module import get_motion_module
22
+ from .resnet import Downsample3D, ResnetBlock3D, Upsample3D
23
+ from .transformer_3d import Transformer3DModel
24
+
25
+
26
+ def get_down_block(
27
+ down_block_type,
28
+ num_layers,
29
+ in_channels,
30
+ out_channels,
31
+ temb_channels,
32
+ add_downsample,
33
+ resnet_eps,
34
+ resnet_act_fn,
35
+ attn_num_head_channels,
36
+ resnet_groups=None,
37
+ cross_attention_dim=None,
38
+ audio_attention_dim=None,
39
+ downsample_padding=None,
40
+ dual_cross_attention=False,
41
+ use_linear_projection=False,
42
+ only_cross_attention=False,
43
+ upcast_attention=False,
44
+ resnet_time_scale_shift="default",
45
+ unet_use_cross_frame_attention=None,
46
+ unet_use_temporal_attention=None,
47
+ use_inflated_groupnorm=None,
48
+ use_motion_module=None,
49
+ motion_module_type=None,
50
+ motion_module_kwargs=None,
51
+ use_audio_module=None,
52
+ depth=0,
53
+ stack_enable_blocks_name=None,
54
+ stack_enable_blocks_depth=None,
55
+ ):
56
+ """
57
+ Factory function to instantiate a down-block module for the 3D UNet architecture.
58
+
59
+ Down blocks are used in the downsampling part of the U-Net to reduce the spatial dimensions
60
+ of the feature maps while increasing the depth. This function can create blocks with or without
61
+ cross attention based on the specified parameters.
62
+
63
+ Parameters:
64
+ - down_block_type (str): The type of down block to instantiate.
65
+ - num_layers (int): The number of layers in the block.
66
+ - in_channels (int): The number of input channels.
67
+ - out_channels (int): The number of output channels.
68
+ - temb_channels (int): The number of token embedding channels.
69
+ - add_downsample (bool): Flag to add a downsampling layer.
70
+ - resnet_eps (float): Epsilon for residual block stability.
71
+ - resnet_act_fn (callable): Activation function for the residual block.
72
+ - ... (remaining parameters): Additional parameters for configuring the block.
73
+
74
+ Returns:
75
+ - nn.Module: An instance of a down-sampling block module.
76
+ """
77
+ down_block_type = (
78
+ down_block_type[7:]
79
+ if down_block_type.startswith("UNetRes")
80
+ else down_block_type
81
+ )
82
+ if down_block_type == "DownBlock3D":
83
+ return DownBlock3D(
84
+ num_layers=num_layers,
85
+ in_channels=in_channels,
86
+ out_channels=out_channels,
87
+ temb_channels=temb_channels,
88
+ add_downsample=add_downsample,
89
+ resnet_eps=resnet_eps,
90
+ resnet_act_fn=resnet_act_fn,
91
+ resnet_groups=resnet_groups,
92
+ downsample_padding=downsample_padding,
93
+ resnet_time_scale_shift=resnet_time_scale_shift,
94
+ use_inflated_groupnorm=use_inflated_groupnorm,
95
+ use_motion_module=use_motion_module,
96
+ motion_module_type=motion_module_type,
97
+ motion_module_kwargs=motion_module_kwargs,
98
+ )
99
+
100
+ if down_block_type == "CrossAttnDownBlock3D":
101
+ if cross_attention_dim is None:
102
+ raise ValueError(
103
+ "cross_attention_dim must be specified for CrossAttnDownBlock3D"
104
+ )
105
+ return CrossAttnDownBlock3D(
106
+ num_layers=num_layers,
107
+ in_channels=in_channels,
108
+ out_channels=out_channels,
109
+ temb_channels=temb_channels,
110
+ add_downsample=add_downsample,
111
+ resnet_eps=resnet_eps,
112
+ resnet_act_fn=resnet_act_fn,
113
+ resnet_groups=resnet_groups,
114
+ downsample_padding=downsample_padding,
115
+ cross_attention_dim=cross_attention_dim,
116
+ audio_attention_dim=audio_attention_dim,
117
+ attn_num_head_channels=attn_num_head_channels,
118
+ dual_cross_attention=dual_cross_attention,
119
+ use_linear_projection=use_linear_projection,
120
+ only_cross_attention=only_cross_attention,
121
+ upcast_attention=upcast_attention,
122
+ resnet_time_scale_shift=resnet_time_scale_shift,
123
+ unet_use_cross_frame_attention=unet_use_cross_frame_attention,
124
+ unet_use_temporal_attention=unet_use_temporal_attention,
125
+ use_inflated_groupnorm=use_inflated_groupnorm,
126
+ use_motion_module=use_motion_module,
127
+ motion_module_type=motion_module_type,
128
+ motion_module_kwargs=motion_module_kwargs,
129
+ use_audio_module=use_audio_module,
130
+ depth=depth,
131
+ stack_enable_blocks_name=stack_enable_blocks_name,
132
+ stack_enable_blocks_depth=stack_enable_blocks_depth,
133
+ )
134
+ raise ValueError(f"{down_block_type} does not exist.")
135
+
136
+
137
+ def get_up_block(
138
+ up_block_type,
139
+ num_layers,
140
+ in_channels,
141
+ out_channels,
142
+ prev_output_channel,
143
+ temb_channels,
144
+ add_upsample,
145
+ resnet_eps,
146
+ resnet_act_fn,
147
+ attn_num_head_channels,
148
+ resnet_groups=None,
149
+ cross_attention_dim=None,
150
+ audio_attention_dim=None,
151
+ dual_cross_attention=False,
152
+ use_linear_projection=False,
153
+ only_cross_attention=False,
154
+ upcast_attention=False,
155
+ resnet_time_scale_shift="default",
156
+ unet_use_cross_frame_attention=None,
157
+ unet_use_temporal_attention=None,
158
+ use_inflated_groupnorm=None,
159
+ use_motion_module=None,
160
+ motion_module_type=None,
161
+ motion_module_kwargs=None,
162
+ use_audio_module=None,
163
+ depth=0,
164
+ stack_enable_blocks_name=None,
165
+ stack_enable_blocks_depth=None,
166
+ ):
167
+ """
168
+ Factory function to instantiate an up-block module for the 3D UNet architecture.
169
+
170
+ Up blocks are used in the upsampling part of the U-Net to increase the spatial dimensions
171
+ of the feature maps while decreasing the depth. This function can create blocks with or without
172
+ cross attention based on the specified parameters.
173
+
174
+ Parameters:
175
+ - up_block_type (str): The type of up block to instantiate.
176
+ - num_layers (int): The number of layers in the block.
177
+ - in_channels (int): The number of input channels.
178
+ - out_channels (int): The number of output channels.
179
+ - prev_output_channel (int): The number of channels from the previous layer's output.
180
+ - temb_channels (int): The number of token embedding channels.
181
+ - add_upsample (bool): Flag to add an upsampling layer.
182
+ - resnet_eps (float): Epsilon for residual block stability.
183
+ - resnet_act_fn (callable): Activation function for the residual block.
184
+ - ... (remaining parameters): Additional parameters for configuring the block.
185
+
186
+ Returns:
187
+ - nn.Module: An instance of an up-sampling block module.
188
+ """
189
+ up_block_type = (
190
+ up_block_type[7:] if up_block_type.startswith("UNetRes") else up_block_type
191
+ )
192
+ if up_block_type == "UpBlock3D":
193
+ return UpBlock3D(
194
+ num_layers=num_layers,
195
+ in_channels=in_channels,
196
+ out_channels=out_channels,
197
+ prev_output_channel=prev_output_channel,
198
+ temb_channels=temb_channels,
199
+ add_upsample=add_upsample,
200
+ resnet_eps=resnet_eps,
201
+ resnet_act_fn=resnet_act_fn,
202
+ resnet_groups=resnet_groups,
203
+ resnet_time_scale_shift=resnet_time_scale_shift,
204
+ use_inflated_groupnorm=use_inflated_groupnorm,
205
+ use_motion_module=use_motion_module,
206
+ motion_module_type=motion_module_type,
207
+ motion_module_kwargs=motion_module_kwargs,
208
+ )
209
+
210
+ if up_block_type == "CrossAttnUpBlock3D":
211
+ if cross_attention_dim is None:
212
+ raise ValueError(
213
+ "cross_attention_dim must be specified for CrossAttnUpBlock3D"
214
+ )
215
+ return CrossAttnUpBlock3D(
216
+ num_layers=num_layers,
217
+ in_channels=in_channels,
218
+ out_channels=out_channels,
219
+ prev_output_channel=prev_output_channel,
220
+ temb_channels=temb_channels,
221
+ add_upsample=add_upsample,
222
+ resnet_eps=resnet_eps,
223
+ resnet_act_fn=resnet_act_fn,
224
+ resnet_groups=resnet_groups,
225
+ cross_attention_dim=cross_attention_dim,
226
+ audio_attention_dim=audio_attention_dim,
227
+ attn_num_head_channels=attn_num_head_channels,
228
+ dual_cross_attention=dual_cross_attention,
229
+ use_linear_projection=use_linear_projection,
230
+ only_cross_attention=only_cross_attention,
231
+ upcast_attention=upcast_attention,
232
+ resnet_time_scale_shift=resnet_time_scale_shift,
233
+ unet_use_cross_frame_attention=unet_use_cross_frame_attention,
234
+ unet_use_temporal_attention=unet_use_temporal_attention,
235
+ use_inflated_groupnorm=use_inflated_groupnorm,
236
+ use_motion_module=use_motion_module,
237
+ motion_module_type=motion_module_type,
238
+ motion_module_kwargs=motion_module_kwargs,
239
+ use_audio_module=use_audio_module,
240
+ depth=depth,
241
+ stack_enable_blocks_name=stack_enable_blocks_name,
242
+ stack_enable_blocks_depth=stack_enable_blocks_depth,
243
+ )
244
+ raise ValueError(f"{up_block_type} does not exist.")
245
+
246
+
247
+ class UNetMidBlock3DCrossAttn(nn.Module):
248
+ """
249
+ A 3D UNet middle block with cross attention mechanism. This block is part of the U-Net architecture
250
+ and is used for feature extraction in the middle of the downsampling path.
251
+
252
+ Parameters:
253
+ - in_channels (int): Number of input channels.
254
+ - temb_channels (int): Number of token embedding channels.
255
+ - dropout (float): Dropout rate.
256
+ - num_layers (int): Number of layers in the block.
257
+ - resnet_eps (float): Epsilon for residual block.
258
+ - resnet_time_scale_shift (str): Time scale shift for time embedding normalization.
259
+ - resnet_act_fn (str): Activation function for the residual block.
260
+ - resnet_groups (int): Number of groups for the convolutions in the residual block.
261
+ - resnet_pre_norm (bool): Whether to use pre-normalization in the residual block.
262
+ - attn_num_head_channels (int): Number of attention heads.
263
+ - cross_attention_dim (int): Dimensionality of the cross attention layers.
264
+ - audio_attention_dim (int): Dimensionality of the audio attention layers.
265
+ - dual_cross_attention (bool): Whether to use dual cross attention.
266
+ - use_linear_projection (bool): Whether to use linear projection in attention.
267
+ - upcast_attention (bool): Whether to upcast attention to the original input dimension.
268
+ - unet_use_cross_frame_attention (bool): Whether to use cross frame attention in U-Net.
269
+ - unet_use_temporal_attention (bool): Whether to use temporal attention in U-Net.
270
+ - use_inflated_groupnorm (bool): Whether to use inflated group normalization.
271
+ - use_motion_module (bool): Whether to use motion module.
272
+ - motion_module_type (str): Type of motion module.
273
+ - motion_module_kwargs (dict): Keyword arguments for the motion module.
274
+ - use_audio_module (bool): Whether to use audio module.
275
+ - depth (int): Depth of the block in the network.
276
+ - stack_enable_blocks_name (str): Name of the stack enable blocks.
277
+ - stack_enable_blocks_depth (int): Depth of the stack enable blocks.
278
+
279
+ Forward method:
280
+ The forward method applies the residual blocks, cross attention, and optional motion and audio modules
281
+ to the input hidden states. It returns the transformed hidden states.
282
+ """
283
+ def __init__(
284
+ self,
285
+ in_channels: int,
286
+ temb_channels: int,
287
+ dropout: float = 0.0,
288
+ num_layers: int = 1,
289
+ resnet_eps: float = 1e-6,
290
+ resnet_time_scale_shift: str = "default",
291
+ resnet_act_fn: str = "swish",
292
+ resnet_groups: int = 32,
293
+ resnet_pre_norm: bool = True,
294
+ attn_num_head_channels=1,
295
+ output_scale_factor=1.0,
296
+ cross_attention_dim=1280,
297
+ audio_attention_dim=1024,
298
+ dual_cross_attention=False,
299
+ use_linear_projection=False,
300
+ upcast_attention=False,
301
+ unet_use_cross_frame_attention=None,
302
+ unet_use_temporal_attention=None,
303
+ use_inflated_groupnorm=None,
304
+ use_motion_module=None,
305
+ motion_module_type=None,
306
+ motion_module_kwargs=None,
307
+ use_audio_module=None,
308
+ depth=0,
309
+ stack_enable_blocks_name=None,
310
+ stack_enable_blocks_depth=None,
311
+ ):
312
+ super().__init__()
313
+
314
+ self.has_cross_attention = True
315
+ self.attn_num_head_channels = attn_num_head_channels
316
+ resnet_groups = (
317
+ resnet_groups if resnet_groups is not None else min(in_channels // 4, 32)
318
+ )
319
+
320
+ # there is always at least one resnet
321
+ resnets = [
322
+ ResnetBlock3D(
323
+ in_channels=in_channels,
324
+ out_channels=in_channels,
325
+ temb_channels=temb_channels,
326
+ eps=resnet_eps,
327
+ groups=resnet_groups,
328
+ dropout=dropout,
329
+ time_embedding_norm=resnet_time_scale_shift,
330
+ non_linearity=resnet_act_fn,
331
+ output_scale_factor=output_scale_factor,
332
+ pre_norm=resnet_pre_norm,
333
+ use_inflated_groupnorm=use_inflated_groupnorm,
334
+ )
335
+ ]
336
+ attentions = []
337
+ motion_modules = []
338
+ audio_modules = []
339
+
340
+ for _ in range(num_layers):
341
+ if dual_cross_attention:
342
+ raise NotImplementedError
343
+ attentions.append(
344
+ Transformer3DModel(
345
+ attn_num_head_channels,
346
+ in_channels // attn_num_head_channels,
347
+ in_channels=in_channels,
348
+ num_layers=1,
349
+ cross_attention_dim=cross_attention_dim,
350
+ norm_num_groups=resnet_groups,
351
+ use_linear_projection=use_linear_projection,
352
+ upcast_attention=upcast_attention,
353
+ unet_use_cross_frame_attention=unet_use_cross_frame_attention,
354
+ unet_use_temporal_attention=unet_use_temporal_attention,
355
+ )
356
+ )
357
+ audio_modules.append(
358
+ Transformer3DModel(
359
+ attn_num_head_channels,
360
+ in_channels // attn_num_head_channels,
361
+ in_channels=in_channels,
362
+ num_layers=1,
363
+ cross_attention_dim=audio_attention_dim,
364
+ norm_num_groups=resnet_groups,
365
+ use_linear_projection=use_linear_projection,
366
+ upcast_attention=upcast_attention,
367
+ use_audio_module=use_audio_module,
368
+ depth=depth,
369
+ unet_block_name="mid",
370
+ stack_enable_blocks_name=stack_enable_blocks_name,
371
+ stack_enable_blocks_depth=stack_enable_blocks_depth,
372
+ )
373
+ if use_audio_module
374
+ else None
375
+ )
376
+
377
+ motion_modules.append(
378
+ get_motion_module(
379
+ in_channels=in_channels,
380
+ motion_module_type=motion_module_type,
381
+ motion_module_kwargs=motion_module_kwargs,
382
+ )
383
+ if use_motion_module
384
+ else None
385
+ )
386
+ resnets.append(
387
+ ResnetBlock3D(
388
+ in_channels=in_channels,
389
+ out_channels=in_channels,
390
+ temb_channels=temb_channels,
391
+ eps=resnet_eps,
392
+ groups=resnet_groups,
393
+ dropout=dropout,
394
+ time_embedding_norm=resnet_time_scale_shift,
395
+ non_linearity=resnet_act_fn,
396
+ output_scale_factor=output_scale_factor,
397
+ pre_norm=resnet_pre_norm,
398
+ use_inflated_groupnorm=use_inflated_groupnorm,
399
+ )
400
+ )
401
+
402
+ self.attentions = nn.ModuleList(attentions)
403
+ self.resnets = nn.ModuleList(resnets)
404
+ self.audio_modules = nn.ModuleList(audio_modules)
405
+ self.motion_modules = nn.ModuleList(motion_modules)
406
+
407
+ def forward(
408
+ self,
409
+ hidden_states,
410
+ temb=None,
411
+ encoder_hidden_states=None,
412
+ attention_mask=None,
413
+ full_mask=None,
414
+ face_mask=None,
415
+ lip_mask=None,
416
+ audio_embedding=None,
417
+ motion_scale=None,
418
+ ):
419
+ """
420
+ Forward pass for the UNetMidBlock3DCrossAttn class.
421
+
422
+ Args:
423
+ self (UNetMidBlock3DCrossAttn): An instance of the UNetMidBlock3DCrossAttn class.
424
+ hidden_states (Tensor): The input hidden states tensor.
425
+ temb (Tensor, optional): The input temporal embedding tensor. Defaults to None.
426
+ encoder_hidden_states (Tensor, optional): The encoder hidden states tensor. Defaults to None.
427
+ attention_mask (Tensor, optional): The attention mask tensor. Defaults to None.
428
+ full_mask (Tensor, optional): The full mask tensor. Defaults to None.
429
+ face_mask (Tensor, optional): The face mask tensor. Defaults to None.
430
+ lip_mask (Tensor, optional): The lip mask tensor. Defaults to None.
431
+ audio_embedding (Tensor, optional): The audio embedding tensor. Defaults to None.
432
+
433
+ Returns:
434
+ Tensor: The output tensor after passing through the UNetMidBlock3DCrossAttn layers.
435
+ """
436
+ hidden_states = self.resnets[0](hidden_states, temb)
437
+ for attn, resnet, audio_module, motion_module in zip(
438
+ self.attentions, self.resnets[1:], self.audio_modules, self.motion_modules
439
+ ):
440
+ hidden_states, motion_frame = attn(
441
+ hidden_states,
442
+ encoder_hidden_states=encoder_hidden_states,
443
+ return_dict=False,
444
+ ) # .sample
445
+ if len(motion_frame[0]) > 0:
446
+ # if motion_frame[0][0].numel() > 0:
447
+ motion_frames = motion_frame[0][0]
448
+ motion_frames = rearrange(
449
+ motion_frames,
450
+ "b f (d1 d2) c -> b c f d1 d2",
451
+ d1=hidden_states.size(-1),
452
+ )
453
+
454
+ else:
455
+ motion_frames = torch.zeros(
456
+ hidden_states.shape[0],
457
+ hidden_states.shape[1],
458
+ 4,
459
+ hidden_states.shape[3],
460
+ hidden_states.shape[4],
461
+ )
462
+
463
+ n_motion_frames = motion_frames.size(2)
464
+ if audio_module is not None:
465
+ hidden_states = (
466
+ audio_module(
467
+ hidden_states,
468
+ encoder_hidden_states=audio_embedding,
469
+ attention_mask=attention_mask,
470
+ full_mask=full_mask,
471
+ face_mask=face_mask,
472
+ lip_mask=lip_mask,
473
+ motion_scale=motion_scale,
474
+ return_dict=False,
475
+ )
476
+ )[0] # .sample
477
+ if motion_module is not None:
478
+ motion_frames = motion_frames.to(
479
+ device=hidden_states.device, dtype=hidden_states.dtype
480
+ )
481
+
482
+ _hidden_states = (
483
+ torch.cat([motion_frames, hidden_states], dim=2)
484
+ if n_motion_frames > 0
485
+ else hidden_states
486
+ )
487
+ hidden_states = motion_module(
488
+ _hidden_states, encoder_hidden_states=encoder_hidden_states
489
+ )
490
+ hidden_states = hidden_states[:, :, n_motion_frames:]
491
+
492
+ hidden_states = resnet(hidden_states, temb)
493
+
494
+ return hidden_states
495
+
496
+
497
+ class CrossAttnDownBlock3D(nn.Module):
498
+ """
499
+ A 3D downsampling block with cross attention for the U-Net architecture.
500
+
501
+ Parameters:
502
+ - (same as above, refer to the constructor for details)
503
+
504
+ Forward method:
505
+ The forward method downsamples the input hidden states using residual blocks and cross attention.
506
+ It also applies optional motion and audio modules. The method supports gradient checkpointing
507
+ to save memory during training.
508
+ """
509
+ def __init__(
510
+ self,
511
+ in_channels: int,
512
+ out_channels: int,
513
+ temb_channels: int,
514
+ dropout: float = 0.0,
515
+ num_layers: int = 1,
516
+ resnet_eps: float = 1e-6,
517
+ resnet_time_scale_shift: str = "default",
518
+ resnet_act_fn: str = "swish",
519
+ resnet_groups: int = 32,
520
+ resnet_pre_norm: bool = True,
521
+ attn_num_head_channels=1,
522
+ cross_attention_dim=1280,
523
+ audio_attention_dim=1024,
524
+ output_scale_factor=1.0,
525
+ downsample_padding=1,
526
+ add_downsample=True,
527
+ dual_cross_attention=False,
528
+ use_linear_projection=False,
529
+ only_cross_attention=False,
530
+ upcast_attention=False,
531
+ unet_use_cross_frame_attention=None,
532
+ unet_use_temporal_attention=None,
533
+ use_inflated_groupnorm=None,
534
+ use_motion_module=None,
535
+ motion_module_type=None,
536
+ motion_module_kwargs=None,
537
+ use_audio_module=None,
538
+ depth=0,
539
+ stack_enable_blocks_name=None,
540
+ stack_enable_blocks_depth=None,
541
+ ):
542
+ super().__init__()
543
+ resnets = []
544
+ attentions = []
545
+ audio_modules = []
546
+ motion_modules = []
547
+
548
+ self.has_cross_attention = True
549
+ self.attn_num_head_channels = attn_num_head_channels
550
+
551
+ for i in range(num_layers):
552
+ in_channels = in_channels if i == 0 else out_channels
553
+ resnets.append(
554
+ ResnetBlock3D(
555
+ in_channels=in_channels,
556
+ out_channels=out_channels,
557
+ temb_channels=temb_channels,
558
+ eps=resnet_eps,
559
+ groups=resnet_groups,
560
+ dropout=dropout,
561
+ time_embedding_norm=resnet_time_scale_shift,
562
+ non_linearity=resnet_act_fn,
563
+ output_scale_factor=output_scale_factor,
564
+ pre_norm=resnet_pre_norm,
565
+ use_inflated_groupnorm=use_inflated_groupnorm,
566
+ )
567
+ )
568
+ if dual_cross_attention:
569
+ raise NotImplementedError
570
+ attentions.append(
571
+ Transformer3DModel(
572
+ attn_num_head_channels,
573
+ out_channels // attn_num_head_channels,
574
+ in_channels=out_channels,
575
+ num_layers=1,
576
+ cross_attention_dim=cross_attention_dim,
577
+ norm_num_groups=resnet_groups,
578
+ use_linear_projection=use_linear_projection,
579
+ only_cross_attention=only_cross_attention,
580
+ upcast_attention=upcast_attention,
581
+ unet_use_cross_frame_attention=unet_use_cross_frame_attention,
582
+ unet_use_temporal_attention=unet_use_temporal_attention,
583
+ )
584
+ )
585
+ # TODO:检查维度
586
+ audio_modules.append(
587
+ Transformer3DModel(
588
+ attn_num_head_channels,
589
+ in_channels // attn_num_head_channels,
590
+ in_channels=out_channels,
591
+ num_layers=1,
592
+ cross_attention_dim=audio_attention_dim,
593
+ norm_num_groups=resnet_groups,
594
+ use_linear_projection=use_linear_projection,
595
+ only_cross_attention=only_cross_attention,
596
+ upcast_attention=upcast_attention,
597
+ use_audio_module=use_audio_module,
598
+ depth=depth,
599
+ unet_block_name="down",
600
+ stack_enable_blocks_name=stack_enable_blocks_name,
601
+ stack_enable_blocks_depth=stack_enable_blocks_depth,
602
+ )
603
+ if use_audio_module
604
+ else None
605
+ )
606
+ motion_modules.append(
607
+ get_motion_module(
608
+ in_channels=out_channels,
609
+ motion_module_type=motion_module_type,
610
+ motion_module_kwargs=motion_module_kwargs,
611
+ )
612
+ if use_motion_module
613
+ else None
614
+ )
615
+
616
+ self.attentions = nn.ModuleList(attentions)
617
+ self.resnets = nn.ModuleList(resnets)
618
+ self.audio_modules = nn.ModuleList(audio_modules)
619
+ self.motion_modules = nn.ModuleList(motion_modules)
620
+
621
+ if add_downsample:
622
+ self.downsamplers = nn.ModuleList(
623
+ [
624
+ Downsample3D(
625
+ out_channels,
626
+ use_conv=True,
627
+ out_channels=out_channels,
628
+ padding=downsample_padding,
629
+ name="op",
630
+ )
631
+ ]
632
+ )
633
+ else:
634
+ self.downsamplers = None
635
+
636
+ self.gradient_checkpointing = False
637
+
638
+ def forward(
639
+ self,
640
+ hidden_states,
641
+ temb=None,
642
+ encoder_hidden_states=None,
643
+ attention_mask=None,
644
+ full_mask=None,
645
+ face_mask=None,
646
+ lip_mask=None,
647
+ audio_embedding=None,
648
+ motion_scale=None,
649
+ ):
650
+ """
651
+ Defines the forward pass for the CrossAttnDownBlock3D class.
652
+
653
+ Parameters:
654
+ - hidden_states : torch.Tensor
655
+ The input tensor to the block.
656
+ temb : torch.Tensor, optional
657
+ The token embeddings from the previous block.
658
+ encoder_hidden_states : torch.Tensor, optional
659
+ The hidden states from the encoder.
660
+ attention_mask : torch.Tensor, optional
661
+ The attention mask for the cross-attention mechanism.
662
+ full_mask : torch.Tensor, optional
663
+ The full mask for the cross-attention mechanism.
664
+ face_mask : torch.Tensor, optional
665
+ The face mask for the cross-attention mechanism.
666
+ lip_mask : torch.Tensor, optional
667
+ The lip mask for the cross-attention mechanism.
668
+ audio_embedding : torch.Tensor, optional
669
+ The audio embedding for the cross-attention mechanism.
670
+
671
+ Returns:
672
+ -- torch.Tensor
673
+ The output tensor from the block.
674
+ """
675
+ output_states = ()
676
+
677
+ for _, (resnet, attn, audio_module, motion_module) in enumerate(
678
+ zip(self.resnets, self.attentions, self.audio_modules, self.motion_modules)
679
+ ):
680
+ # self.gradient_checkpointing = False
681
+ if self.training and self.gradient_checkpointing:
682
+
683
+ def create_custom_forward(module, return_dict=None):
684
+ def custom_forward(*inputs):
685
+ if return_dict is not None:
686
+ return module(*inputs, return_dict=return_dict)
687
+
688
+ return module(*inputs)
689
+
690
+ return custom_forward
691
+
692
+ hidden_states = torch.utils.checkpoint.checkpoint(
693
+ create_custom_forward(resnet), hidden_states, temb
694
+ )
695
+
696
+ motion_frames = []
697
+ hidden_states, motion_frame = torch.utils.checkpoint.checkpoint(
698
+ create_custom_forward(attn, return_dict=False),
699
+ hidden_states,
700
+ encoder_hidden_states,
701
+ )
702
+ if len(motion_frame[0]) > 0:
703
+ motion_frames = motion_frame[0][0]
704
+ # motion_frames = torch.cat(motion_frames, dim=0)
705
+ motion_frames = rearrange(
706
+ motion_frames,
707
+ "b f (d1 d2) c -> b c f d1 d2",
708
+ d1=hidden_states.size(-1),
709
+ )
710
+
711
+ else:
712
+ motion_frames = torch.zeros(
713
+ hidden_states.shape[0],
714
+ hidden_states.shape[1],
715
+ 4,
716
+ hidden_states.shape[3],
717
+ hidden_states.shape[4],
718
+ )
719
+
720
+ n_motion_frames = motion_frames.size(2)
721
+
722
+ if audio_module is not None:
723
+ # audio_embedding = audio_embedding
724
+ hidden_states = torch.utils.checkpoint.checkpoint(
725
+ create_custom_forward(audio_module, return_dict=False),
726
+ hidden_states,
727
+ audio_embedding,
728
+ attention_mask,
729
+ full_mask,
730
+ face_mask,
731
+ lip_mask,
732
+ motion_scale,
733
+ )[0]
734
+
735
+ # add motion module
736
+ if motion_module is not None:
737
+ motion_frames = motion_frames.to(
738
+ device=hidden_states.device, dtype=hidden_states.dtype
739
+ )
740
+ _hidden_states = torch.cat(
741
+ [motion_frames, hidden_states], dim=2
742
+ ) # if n_motion_frames > 0 else hidden_states
743
+ hidden_states = torch.utils.checkpoint.checkpoint(
744
+ create_custom_forward(motion_module),
745
+ _hidden_states,
746
+ encoder_hidden_states,
747
+ )
748
+ hidden_states = hidden_states[:, :, n_motion_frames:]
749
+
750
+ else:
751
+ hidden_states = resnet(hidden_states, temb)
752
+ hidden_states = attn(
753
+ hidden_states,
754
+ encoder_hidden_states=encoder_hidden_states,
755
+ ).sample
756
+ if audio_module is not None:
757
+ hidden_states = audio_module(
758
+ hidden_states,
759
+ audio_embedding,
760
+ attention_mask=attention_mask,
761
+ full_mask=full_mask,
762
+ face_mask=face_mask,
763
+ lip_mask=lip_mask,
764
+ return_dict=False,
765
+ )[0]
766
+ # add motion module
767
+ if motion_module is not None:
768
+ hidden_states = motion_module(
769
+ hidden_states, encoder_hidden_states=encoder_hidden_states
770
+ )
771
+
772
+ output_states += (hidden_states,)
773
+
774
+ if self.downsamplers is not None:
775
+ for downsampler in self.downsamplers:
776
+ hidden_states = downsampler(hidden_states)
777
+
778
+ output_states += (hidden_states,)
779
+
780
+ return hidden_states, output_states
781
+
782
+
783
+ class DownBlock3D(nn.Module):
784
+ """
785
+ A 3D downsampling block for the U-Net architecture. This block performs downsampling operations
786
+ using residual blocks and an optional motion module.
787
+
788
+ Parameters:
789
+ - in_channels (int): Number of input channels.
790
+ - out_channels (int): Number of output channels.
791
+ - temb_channels (int): Number of token embedding channels.
792
+ - dropout (float): Dropout rate for the block.
793
+ - num_layers (int): Number of layers in the block.
794
+ - resnet_eps (float): Epsilon for residual block stability.
795
+ - resnet_time_scale_shift (str): Time scale shift for the residual block's time embedding.
796
+ - resnet_act_fn (str): Activation function used in the residual block.
797
+ - resnet_groups (int): Number of groups for the convolutions in the residual block.
798
+ - resnet_pre_norm (bool): Whether to use pre-normalization in the residual block.
799
+ - output_scale_factor (float): Scaling factor for the block's output.
800
+ - add_downsample (bool): Whether to add a downsampling layer.
801
+ - downsample_padding (int): Padding for the downsampling layer.
802
+ - use_inflated_groupnorm (bool): Whether to use inflated group normalization.
803
+ - use_motion_module (bool): Whether to include a motion module.
804
+ - motion_module_type (str): Type of motion module to use.
805
+ - motion_module_kwargs (dict): Keyword arguments for the motion module.
806
+
807
+ Forward method:
808
+ The forward method processes the input hidden states through the residual blocks and optional
809
+ motion modules, followed by an optional downsampling step. It supports gradient checkpointing
810
+ during training to reduce memory usage.
811
+ """
812
+ def __init__(
813
+ self,
814
+ in_channels: int,
815
+ out_channels: int,
816
+ temb_channels: int,
817
+ dropout: float = 0.0,
818
+ num_layers: int = 1,
819
+ resnet_eps: float = 1e-6,
820
+ resnet_time_scale_shift: str = "default",
821
+ resnet_act_fn: str = "swish",
822
+ resnet_groups: int = 32,
823
+ resnet_pre_norm: bool = True,
824
+ output_scale_factor=1.0,
825
+ add_downsample=True,
826
+ downsample_padding=1,
827
+ use_inflated_groupnorm=None,
828
+ use_motion_module=None,
829
+ motion_module_type=None,
830
+ motion_module_kwargs=None,
831
+ ):
832
+ super().__init__()
833
+ resnets = []
834
+ motion_modules = []
835
+
836
+ # use_motion_module = False
837
+ for i in range(num_layers):
838
+ in_channels = in_channels if i == 0 else out_channels
839
+ resnets.append(
840
+ ResnetBlock3D(
841
+ in_channels=in_channels,
842
+ out_channels=out_channels,
843
+ temb_channels=temb_channels,
844
+ eps=resnet_eps,
845
+ groups=resnet_groups,
846
+ dropout=dropout,
847
+ time_embedding_norm=resnet_time_scale_shift,
848
+ non_linearity=resnet_act_fn,
849
+ output_scale_factor=output_scale_factor,
850
+ pre_norm=resnet_pre_norm,
851
+ use_inflated_groupnorm=use_inflated_groupnorm,
852
+ )
853
+ )
854
+ motion_modules.append(
855
+ get_motion_module(
856
+ in_channels=out_channels,
857
+ motion_module_type=motion_module_type,
858
+ motion_module_kwargs=motion_module_kwargs,
859
+ )
860
+ if use_motion_module
861
+ else None
862
+ )
863
+
864
+ self.resnets = nn.ModuleList(resnets)
865
+ self.motion_modules = nn.ModuleList(motion_modules)
866
+
867
+ if add_downsample:
868
+ self.downsamplers = nn.ModuleList(
869
+ [
870
+ Downsample3D(
871
+ out_channels,
872
+ use_conv=True,
873
+ out_channels=out_channels,
874
+ padding=downsample_padding,
875
+ name="op",
876
+ )
877
+ ]
878
+ )
879
+ else:
880
+ self.downsamplers = None
881
+
882
+ self.gradient_checkpointing = False
883
+
884
+ def forward(
885
+ self,
886
+ hidden_states,
887
+ temb=None,
888
+ encoder_hidden_states=None,
889
+ ):
890
+ """
891
+ forward method for the DownBlock3D class.
892
+
893
+ Args:
894
+ hidden_states (Tensor): The input tensor to the DownBlock3D layer.
895
+ temb (Tensor, optional): The token embeddings, if using transformer.
896
+ encoder_hidden_states (Tensor, optional): The hidden states from the encoder.
897
+
898
+ Returns:
899
+ Tensor: The output tensor after passing through the DownBlock3D layer.
900
+ """
901
+ output_states = ()
902
+
903
+ for resnet, motion_module in zip(self.resnets, self.motion_modules):
904
+ # print(f"DownBlock3D {self.gradient_checkpointing = }")
905
+ if self.training and self.gradient_checkpointing:
906
+
907
+ def create_custom_forward(module):
908
+ def custom_forward(*inputs):
909
+ return module(*inputs)
910
+
911
+ return custom_forward
912
+
913
+ hidden_states = torch.utils.checkpoint.checkpoint(
914
+ create_custom_forward(resnet), hidden_states, temb
915
+ )
916
+
917
+ else:
918
+ hidden_states = resnet(hidden_states, temb)
919
+
920
+ # add motion module
921
+ hidden_states = (
922
+ motion_module(
923
+ hidden_states, encoder_hidden_states=encoder_hidden_states
924
+ )
925
+ if motion_module is not None
926
+ else hidden_states
927
+ )
928
+
929
+ output_states += (hidden_states,)
930
+
931
+ if self.downsamplers is not None:
932
+ for downsampler in self.downsamplers:
933
+ hidden_states = downsampler(hidden_states)
934
+
935
+ output_states += (hidden_states,)
936
+
937
+ return hidden_states, output_states
938
+
939
+
940
+ class CrossAttnUpBlock3D(nn.Module):
941
+ """
942
+ Standard 3D downsampling block for the U-Net architecture. This block performs downsampling
943
+ operations in the U-Net using residual blocks and an optional motion module.
944
+
945
+ Parameters:
946
+ - in_channels (int): Number of input channels.
947
+ - out_channels (int): Number of output channels.
948
+ - temb_channels (int): Number of channels for the temporal embedding.
949
+ - dropout (float): Dropout rate for the block.
950
+ - num_layers (int): Number of layers in the block.
951
+ - resnet_eps (float): Epsilon for residual block stability.
952
+ - resnet_time_scale_shift (str): Time scale shift for the residual block's time embedding.
953
+ - resnet_act_fn (str): Activation function used in the residual block.
954
+ - resnet_groups (int): Number of groups for the convolutions in the residual block.
955
+ - resnet_pre_norm (bool): Whether to use pre-normalization in the residual block.
956
+ - output_scale_factor (float): Scaling factor for the block's output.
957
+ - add_downsample (bool): Whether to add a downsampling layer.
958
+ - downsample_padding (int): Padding for the downsampling layer.
959
+ - use_inflated_groupnorm (bool): Whether to use inflated group normalization.
960
+ - use_motion_module (bool): Whether to include a motion module.
961
+ - motion_module_type (str): Type of motion module to use.
962
+ - motion_module_kwargs (dict): Keyword arguments for the motion module.
963
+
964
+ Forward method:
965
+ The forward method processes the input hidden states through the residual blocks and optional
966
+ motion modules, followed by an optional downsampling step. It supports gradient checkpointing
967
+ during training to reduce memory usage.
968
+ """
969
+ def __init__(
970
+ self,
971
+ in_channels: int,
972
+ out_channels: int,
973
+ prev_output_channel: int,
974
+ temb_channels: int,
975
+ dropout: float = 0.0,
976
+ num_layers: int = 1,
977
+ resnet_eps: float = 1e-6,
978
+ resnet_time_scale_shift: str = "default",
979
+ resnet_act_fn: str = "swish",
980
+ resnet_groups: int = 32,
981
+ resnet_pre_norm: bool = True,
982
+ attn_num_head_channels=1,
983
+ cross_attention_dim=1280,
984
+ audio_attention_dim=1024,
985
+ output_scale_factor=1.0,
986
+ add_upsample=True,
987
+ dual_cross_attention=False,
988
+ use_linear_projection=False,
989
+ only_cross_attention=False,
990
+ upcast_attention=False,
991
+ unet_use_cross_frame_attention=None,
992
+ unet_use_temporal_attention=None,
993
+ use_motion_module=None,
994
+ use_inflated_groupnorm=None,
995
+ motion_module_type=None,
996
+ motion_module_kwargs=None,
997
+ use_audio_module=None,
998
+ depth=0,
999
+ stack_enable_blocks_name=None,
1000
+ stack_enable_blocks_depth=None,
1001
+ ):
1002
+ super().__init__()
1003
+ resnets = []
1004
+ attentions = []
1005
+ audio_modules = []
1006
+ motion_modules = []
1007
+
1008
+ self.has_cross_attention = True
1009
+ self.attn_num_head_channels = attn_num_head_channels
1010
+
1011
+ for i in range(num_layers):
1012
+ res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
1013
+ resnet_in_channels = prev_output_channel if i == 0 else out_channels
1014
+
1015
+ resnets.append(
1016
+ ResnetBlock3D(
1017
+ in_channels=resnet_in_channels + res_skip_channels,
1018
+ out_channels=out_channels,
1019
+ temb_channels=temb_channels,
1020
+ eps=resnet_eps,
1021
+ groups=resnet_groups,
1022
+ dropout=dropout,
1023
+ time_embedding_norm=resnet_time_scale_shift,
1024
+ non_linearity=resnet_act_fn,
1025
+ output_scale_factor=output_scale_factor,
1026
+ pre_norm=resnet_pre_norm,
1027
+ use_inflated_groupnorm=use_inflated_groupnorm,
1028
+ )
1029
+ )
1030
+
1031
+ if dual_cross_attention:
1032
+ raise NotImplementedError
1033
+ attentions.append(
1034
+ Transformer3DModel(
1035
+ attn_num_head_channels,
1036
+ out_channels // attn_num_head_channels,
1037
+ in_channels=out_channels,
1038
+ num_layers=1,
1039
+ cross_attention_dim=cross_attention_dim,
1040
+ norm_num_groups=resnet_groups,
1041
+ use_linear_projection=use_linear_projection,
1042
+ only_cross_attention=only_cross_attention,
1043
+ upcast_attention=upcast_attention,
1044
+ unet_use_cross_frame_attention=unet_use_cross_frame_attention,
1045
+ unet_use_temporal_attention=unet_use_temporal_attention,
1046
+ )
1047
+ )
1048
+ audio_modules.append(
1049
+ Transformer3DModel(
1050
+ attn_num_head_channels,
1051
+ in_channels // attn_num_head_channels,
1052
+ in_channels=out_channels,
1053
+ num_layers=1,
1054
+ cross_attention_dim=audio_attention_dim,
1055
+ norm_num_groups=resnet_groups,
1056
+ use_linear_projection=use_linear_projection,
1057
+ only_cross_attention=only_cross_attention,
1058
+ upcast_attention=upcast_attention,
1059
+ use_audio_module=use_audio_module,
1060
+ depth=depth,
1061
+ unet_block_name="up",
1062
+ stack_enable_blocks_name=stack_enable_blocks_name,
1063
+ stack_enable_blocks_depth=stack_enable_blocks_depth,
1064
+ )
1065
+ if use_audio_module
1066
+ else None
1067
+ )
1068
+ motion_modules.append(
1069
+ get_motion_module(
1070
+ in_channels=out_channels,
1071
+ motion_module_type=motion_module_type,
1072
+ motion_module_kwargs=motion_module_kwargs,
1073
+ )
1074
+ if use_motion_module
1075
+ else None
1076
+ )
1077
+
1078
+ self.attentions = nn.ModuleList(attentions)
1079
+ self.resnets = nn.ModuleList(resnets)
1080
+ self.audio_modules = nn.ModuleList(audio_modules)
1081
+ self.motion_modules = nn.ModuleList(motion_modules)
1082
+
1083
+ if add_upsample:
1084
+ self.upsamplers = nn.ModuleList(
1085
+ [Upsample3D(out_channels, use_conv=True, out_channels=out_channels)]
1086
+ )
1087
+ else:
1088
+ self.upsamplers = None
1089
+
1090
+ self.gradient_checkpointing = False
1091
+
1092
+ def forward(
1093
+ self,
1094
+ hidden_states,
1095
+ res_hidden_states_tuple,
1096
+ temb=None,
1097
+ encoder_hidden_states=None,
1098
+ upsample_size=None,
1099
+ attention_mask=None,
1100
+ full_mask=None,
1101
+ face_mask=None,
1102
+ lip_mask=None,
1103
+ audio_embedding=None,
1104
+ motion_scale=None,
1105
+ ):
1106
+ """
1107
+ Forward pass for the CrossAttnUpBlock3D class.
1108
+
1109
+ Args:
1110
+ self (CrossAttnUpBlock3D): An instance of the CrossAttnUpBlock3D class.
1111
+ hidden_states (Tensor): The input hidden states tensor.
1112
+ res_hidden_states_tuple (Tuple[Tensor]): A tuple of residual hidden states tensors.
1113
+ temb (Tensor, optional): The token embeddings tensor. Defaults to None.
1114
+ encoder_hidden_states (Tensor, optional): The encoder hidden states tensor. Defaults to None.
1115
+ upsample_size (int, optional): The upsample size. Defaults to None.
1116
+ attention_mask (Tensor, optional): The attention mask tensor. Defaults to None.
1117
+ full_mask (Tensor, optional): The full mask tensor. Defaults to None.
1118
+ face_mask (Tensor, optional): The face mask tensor. Defaults to None.
1119
+ lip_mask (Tensor, optional): The lip mask tensor. Defaults to None.
1120
+ audio_embedding (Tensor, optional): The audio embedding tensor. Defaults to None.
1121
+
1122
+ Returns:
1123
+ Tensor: The output tensor after passing through the CrossAttnUpBlock3D.
1124
+ """
1125
+ for _, (resnet, attn, audio_module, motion_module) in enumerate(
1126
+ zip(self.resnets, self.attentions, self.audio_modules, self.motion_modules)
1127
+ ):
1128
+ # pop res hidden states
1129
+ res_hidden_states = res_hidden_states_tuple[-1]
1130
+ res_hidden_states_tuple = res_hidden_states_tuple[:-1]
1131
+ hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
1132
+
1133
+ if self.training and self.gradient_checkpointing:
1134
+
1135
+ def create_custom_forward(module, return_dict=None):
1136
+ def custom_forward(*inputs):
1137
+ if return_dict is not None:
1138
+ return module(*inputs, return_dict=return_dict)
1139
+
1140
+ return module(*inputs)
1141
+
1142
+ return custom_forward
1143
+
1144
+ hidden_states = torch.utils.checkpoint.checkpoint(
1145
+ create_custom_forward(resnet), hidden_states, temb
1146
+ )
1147
+
1148
+ motion_frames = []
1149
+ hidden_states, motion_frame = torch.utils.checkpoint.checkpoint(
1150
+ create_custom_forward(attn, return_dict=False),
1151
+ hidden_states,
1152
+ encoder_hidden_states,
1153
+ )
1154
+ if len(motion_frame[0]) > 0:
1155
+ motion_frames = motion_frame[0][0]
1156
+ # motion_frames = torch.cat(motion_frames, dim=0)
1157
+ motion_frames = rearrange(
1158
+ motion_frames,
1159
+ "b f (d1 d2) c -> b c f d1 d2",
1160
+ d1=hidden_states.size(-1),
1161
+ )
1162
+ else:
1163
+ motion_frames = torch.zeros(
1164
+ hidden_states.shape[0],
1165
+ hidden_states.shape[1],
1166
+ 4,
1167
+ hidden_states.shape[3],
1168
+ hidden_states.shape[4],
1169
+ )
1170
+
1171
+ n_motion_frames = motion_frames.size(2)
1172
+
1173
+ if audio_module is not None:
1174
+ # audio_embedding = audio_embedding
1175
+ hidden_states = torch.utils.checkpoint.checkpoint(
1176
+ create_custom_forward(audio_module, return_dict=False),
1177
+ hidden_states,
1178
+ audio_embedding,
1179
+ attention_mask,
1180
+ full_mask,
1181
+ face_mask,
1182
+ lip_mask,
1183
+ motion_scale,
1184
+ )[0]
1185
+
1186
+ # add motion module
1187
+ if motion_module is not None:
1188
+ motion_frames = motion_frames.to(
1189
+ device=hidden_states.device, dtype=hidden_states.dtype
1190
+ )
1191
+
1192
+ _hidden_states = (
1193
+ torch.cat([motion_frames, hidden_states], dim=2)
1194
+ if n_motion_frames > 0
1195
+ else hidden_states
1196
+ )
1197
+ hidden_states = torch.utils.checkpoint.checkpoint(
1198
+ create_custom_forward(motion_module),
1199
+ _hidden_states,
1200
+ encoder_hidden_states,
1201
+ )
1202
+ hidden_states = hidden_states[:, :, n_motion_frames:]
1203
+ else:
1204
+ hidden_states = resnet(hidden_states, temb)
1205
+ hidden_states = attn(
1206
+ hidden_states,
1207
+ encoder_hidden_states=encoder_hidden_states,
1208
+ ).sample
1209
+
1210
+ if audio_module is not None:
1211
+
1212
+ hidden_states = (
1213
+ audio_module(
1214
+ hidden_states,
1215
+ encoder_hidden_states=audio_embedding,
1216
+ attention_mask=attention_mask,
1217
+ full_mask=full_mask,
1218
+ face_mask=face_mask,
1219
+ lip_mask=lip_mask,
1220
+ )
1221
+ ).sample
1222
+ # add motion module
1223
+ hidden_states = (
1224
+ motion_module(
1225
+ hidden_states, encoder_hidden_states=encoder_hidden_states
1226
+ )
1227
+ if motion_module is not None
1228
+ else hidden_states
1229
+ )
1230
+
1231
+ if self.upsamplers is not None:
1232
+ for upsampler in self.upsamplers:
1233
+ hidden_states = upsampler(hidden_states, upsample_size)
1234
+
1235
+ return hidden_states
1236
+
1237
+
1238
+ class UpBlock3D(nn.Module):
1239
+ """
1240
+ 3D upsampling block with cross attention for the U-Net architecture. This block performs
1241
+ upsampling operations and incorporates cross attention mechanisms, which allow the model to
1242
+ focus on different parts of the input when upscaling.
1243
+
1244
+ Parameters:
1245
+ - in_channels (int): Number of input channels.
1246
+ - out_channels (int): Number of output channels.
1247
+ - prev_output_channel (int): Number of channels from the previous layer's output.
1248
+ - temb_channels (int): Number of channels for the temporal embedding.
1249
+ - dropout (float): Dropout rate for the block.
1250
+ - num_layers (int): Number of layers in the block.
1251
+ - resnet_eps (float): Epsilon for residual block stability.
1252
+ - resnet_time_scale_shift (str): Time scale shift for the residual block's time embedding.
1253
+ - resnet_act_fn (str): Activation function used in the residual block.
1254
+ - resnet_groups (int): Number of groups for the convolutions in the residual block.
1255
+ - resnet_pre_norm (bool): Whether to use pre-normalization in the residual block.
1256
+ - attn_num_head_channels (int): Number of attention heads for the cross attention mechanism.
1257
+ - cross_attention_dim (int): Dimensionality of the cross attention layers.
1258
+ - audio_attention_dim (int): Dimensionality of the audio attention layers.
1259
+ - output_scale_factor (float): Scaling factor for the block's output.
1260
+ - add_upsample (bool): Whether to add an upsampling layer.
1261
+ - dual_cross_attention (bool): Whether to use dual cross attention (not implemented).
1262
+ - use_linear_projection (bool): Whether to use linear projection in the cross attention.
1263
+ - only_cross_attention (bool): Whether to use only cross attention (no self-attention).
1264
+ - upcast_attention (bool): Whether to upcast attention to the original input dimension.
1265
+ - unet_use_cross_frame_attention (bool): Whether to use cross frame attention in U-Net.
1266
+ - unet_use_temporal_attention (bool): Whether to use temporal attention in U-Net.
1267
+ - use_motion_module (bool): Whether to include a motion module.
1268
+ - use_inflated_groupnorm (bool): Whether to use inflated group normalization.
1269
+ - motion_module_type (str): Type of motion module to use.
1270
+ - motion_module_kwargs (dict): Keyword arguments for the motion module.
1271
+ - use_audio_module (bool): Whether to include an audio module.
1272
+ - depth (int): Depth of the block in the network.
1273
+ - stack_enable_blocks_name (str): Name of the stack enable blocks.
1274
+ - stack_enable_blocks_depth (int): Depth of the stack enable blocks.
1275
+
1276
+ Forward method:
1277
+ The forward method upsamples the input hidden states and residual hidden states, processes
1278
+ them through the residual and cross attention blocks, and optional motion and audio modules.
1279
+ It supports gradient checkpointing during training.
1280
+ """
1281
+ def __init__(
1282
+ self,
1283
+ in_channels: int,
1284
+ prev_output_channel: int,
1285
+ out_channels: int,
1286
+ temb_channels: int,
1287
+ dropout: float = 0.0,
1288
+ num_layers: int = 1,
1289
+ resnet_eps: float = 1e-6,
1290
+ resnet_time_scale_shift: str = "default",
1291
+ resnet_act_fn: str = "swish",
1292
+ resnet_groups: int = 32,
1293
+ resnet_pre_norm: bool = True,
1294
+ output_scale_factor=1.0,
1295
+ add_upsample=True,
1296
+ use_inflated_groupnorm=None,
1297
+ use_motion_module=None,
1298
+ motion_module_type=None,
1299
+ motion_module_kwargs=None,
1300
+ ):
1301
+ super().__init__()
1302
+ resnets = []
1303
+ motion_modules = []
1304
+
1305
+ # use_motion_module = False
1306
+ for i in range(num_layers):
1307
+ res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
1308
+ resnet_in_channels = prev_output_channel if i == 0 else out_channels
1309
+
1310
+ resnets.append(
1311
+ ResnetBlock3D(
1312
+ in_channels=resnet_in_channels + res_skip_channels,
1313
+ out_channels=out_channels,
1314
+ temb_channels=temb_channels,
1315
+ eps=resnet_eps,
1316
+ groups=resnet_groups,
1317
+ dropout=dropout,
1318
+ time_embedding_norm=resnet_time_scale_shift,
1319
+ non_linearity=resnet_act_fn,
1320
+ output_scale_factor=output_scale_factor,
1321
+ pre_norm=resnet_pre_norm,
1322
+ use_inflated_groupnorm=use_inflated_groupnorm,
1323
+ )
1324
+ )
1325
+ motion_modules.append(
1326
+ get_motion_module(
1327
+ in_channels=out_channels,
1328
+ motion_module_type=motion_module_type,
1329
+ motion_module_kwargs=motion_module_kwargs,
1330
+ )
1331
+ if use_motion_module
1332
+ else None
1333
+ )
1334
+
1335
+ self.resnets = nn.ModuleList(resnets)
1336
+ self.motion_modules = nn.ModuleList(motion_modules)
1337
+
1338
+ if add_upsample:
1339
+ self.upsamplers = nn.ModuleList(
1340
+ [Upsample3D(out_channels, use_conv=True, out_channels=out_channels)]
1341
+ )
1342
+ else:
1343
+ self.upsamplers = None
1344
+
1345
+ self.gradient_checkpointing = False
1346
+
1347
+ def forward(
1348
+ self,
1349
+ hidden_states,
1350
+ res_hidden_states_tuple,
1351
+ temb=None,
1352
+ upsample_size=None,
1353
+ encoder_hidden_states=None,
1354
+ ):
1355
+ """
1356
+ Forward pass for the UpBlock3D class.
1357
+
1358
+ Args:
1359
+ self (UpBlock3D): An instance of the UpBlock3D class.
1360
+ hidden_states (Tensor): The input hidden states tensor.
1361
+ res_hidden_states_tuple (Tuple[Tensor]): A tuple of residual hidden states tensors.
1362
+ temb (Tensor, optional): The token embeddings tensor. Defaults to None.
1363
+ upsample_size (int, optional): The upsample size. Defaults to None.
1364
+ encoder_hidden_states (Tensor, optional): The encoder hidden states tensor. Defaults to None.
1365
+
1366
+ Returns:
1367
+ Tensor: The output tensor after passing through the UpBlock3D layers.
1368
+ """
1369
+ for resnet, motion_module in zip(self.resnets, self.motion_modules):
1370
+ # pop res hidden states
1371
+ res_hidden_states = res_hidden_states_tuple[-1]
1372
+ res_hidden_states_tuple = res_hidden_states_tuple[:-1]
1373
+ hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
1374
+
1375
+ # print(f"UpBlock3D {self.gradient_checkpointing = }")
1376
+ if self.training and self.gradient_checkpointing:
1377
+
1378
+ def create_custom_forward(module):
1379
+ def custom_forward(*inputs):
1380
+ return module(*inputs)
1381
+
1382
+ return custom_forward
1383
+
1384
+ hidden_states = torch.utils.checkpoint.checkpoint(
1385
+ create_custom_forward(resnet), hidden_states, temb
1386
+ )
1387
+ else:
1388
+ hidden_states = resnet(hidden_states, temb)
1389
+ hidden_states = (
1390
+ motion_module(
1391
+ hidden_states, encoder_hidden_states=encoder_hidden_states
1392
+ )
1393
+ if motion_module is not None
1394
+ else hidden_states
1395
+ )
1396
+
1397
+ if self.upsamplers is not None:
1398
+ for upsampler in self.upsamplers:
1399
+ hidden_states = upsampler(hidden_states, upsample_size)
1400
+
1401
+ return hidden_states
hallo/models/wav2vec.py ADDED
@@ -0,0 +1,209 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # pylint: disable=R0901
2
+ # src/models/wav2vec.py
3
+
4
+ """
5
+ This module defines the Wav2Vec model, which is a pre-trained model for speech recognition and understanding.
6
+ It inherits from the Wav2Vec2Model class in the transformers library and provides additional functionalities
7
+ such as feature extraction and encoding.
8
+
9
+ Classes:
10
+ Wav2VecModel: Inherits from Wav2Vec2Model and adds additional methods for feature extraction and encoding.
11
+
12
+ Functions:
13
+ linear_interpolation: Interpolates the features based on the sequence length.
14
+ """
15
+
16
+ import torch.nn.functional as F
17
+ from transformers import Wav2Vec2Model
18
+ from transformers.modeling_outputs import BaseModelOutput
19
+
20
+
21
+ class Wav2VecModel(Wav2Vec2Model):
22
+ """
23
+ Wav2VecModel is a custom model class that extends the Wav2Vec2Model class from the transformers library.
24
+ It inherits all the functionality of the Wav2Vec2Model and adds additional methods for feature extraction and encoding.
25
+ ...
26
+
27
+ Attributes:
28
+ base_model (Wav2Vec2Model): The base Wav2Vec2Model object.
29
+
30
+ Methods:
31
+ forward(input_values, seq_len, attention_mask=None, mask_time_indices=None
32
+ , output_attentions=None, output_hidden_states=None, return_dict=None):
33
+ Forward pass of the Wav2VecModel.
34
+ It takes input_values, seq_len, and other optional parameters as input and returns the output of the base model.
35
+
36
+ feature_extract(input_values, seq_len):
37
+ Extracts features from the input_values using the base model.
38
+
39
+ encode(extract_features, attention_mask=None, mask_time_indices=None, output_attentions=None, output_hidden_states=None, return_dict=None):
40
+ Encodes the extracted features using the base model and returns the encoded features.
41
+ """
42
+ def forward(
43
+ self,
44
+ input_values,
45
+ seq_len,
46
+ attention_mask=None,
47
+ mask_time_indices=None,
48
+ output_attentions=None,
49
+ output_hidden_states=None,
50
+ return_dict=None,
51
+ ):
52
+ """
53
+ Forward pass of the Wav2Vec model.
54
+
55
+ Args:
56
+ self: The instance of the model.
57
+ input_values: The input values (waveform) to the model.
58
+ seq_len: The sequence length of the input values.
59
+ attention_mask: Attention mask to be used for the model.
60
+ mask_time_indices: Mask indices to be used for the model.
61
+ output_attentions: If set to True, returns attentions.
62
+ output_hidden_states: If set to True, returns hidden states.
63
+ return_dict: If set to True, returns a BaseModelOutput instead of a tuple.
64
+
65
+ Returns:
66
+ The output of the Wav2Vec model.
67
+ """
68
+ self.config.output_attentions = True
69
+
70
+ output_hidden_states = (
71
+ output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
72
+ )
73
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
74
+
75
+ extract_features = self.feature_extractor(input_values)
76
+ extract_features = extract_features.transpose(1, 2)
77
+ extract_features = linear_interpolation(extract_features, seq_len=seq_len)
78
+
79
+ if attention_mask is not None:
80
+ # compute reduced attention_mask corresponding to feature vectors
81
+ attention_mask = self._get_feature_vector_attention_mask(
82
+ extract_features.shape[1], attention_mask, add_adapter=False
83
+ )
84
+
85
+ hidden_states, extract_features = self.feature_projection(extract_features)
86
+ hidden_states = self._mask_hidden_states(
87
+ hidden_states, mask_time_indices=mask_time_indices, attention_mask=attention_mask
88
+ )
89
+
90
+ encoder_outputs = self.encoder(
91
+ hidden_states,
92
+ attention_mask=attention_mask,
93
+ output_attentions=output_attentions,
94
+ output_hidden_states=output_hidden_states,
95
+ return_dict=return_dict,
96
+ )
97
+
98
+ hidden_states = encoder_outputs[0]
99
+
100
+ if self.adapter is not None:
101
+ hidden_states = self.adapter(hidden_states)
102
+
103
+ if not return_dict:
104
+ return (hidden_states, ) + encoder_outputs[1:]
105
+ return BaseModelOutput(
106
+ last_hidden_state=hidden_states,
107
+ hidden_states=encoder_outputs.hidden_states,
108
+ attentions=encoder_outputs.attentions,
109
+ )
110
+
111
+
112
+ def feature_extract(
113
+ self,
114
+ input_values,
115
+ seq_len,
116
+ ):
117
+ """
118
+ Extracts features from the input values and returns the extracted features.
119
+
120
+ Parameters:
121
+ input_values (torch.Tensor): The input values to be processed.
122
+ seq_len (torch.Tensor): The sequence lengths of the input values.
123
+
124
+ Returns:
125
+ extracted_features (torch.Tensor): The extracted features from the input values.
126
+ """
127
+ extract_features = self.feature_extractor(input_values)
128
+ extract_features = extract_features.transpose(1, 2)
129
+ extract_features = linear_interpolation(extract_features, seq_len=seq_len)
130
+
131
+ return extract_features
132
+
133
+ def encode(
134
+ self,
135
+ extract_features,
136
+ attention_mask=None,
137
+ mask_time_indices=None,
138
+ output_attentions=None,
139
+ output_hidden_states=None,
140
+ return_dict=None,
141
+ ):
142
+ """
143
+ Encodes the input features into the output space.
144
+
145
+ Args:
146
+ extract_features (torch.Tensor): The extracted features from the audio signal.
147
+ attention_mask (torch.Tensor, optional): Attention mask to be used for padding.
148
+ mask_time_indices (torch.Tensor, optional): Masked indices for the time dimension.
149
+ output_attentions (bool, optional): If set to True, returns the attention weights.
150
+ output_hidden_states (bool, optional): If set to True, returns all hidden states.
151
+ return_dict (bool, optional): If set to True, returns a BaseModelOutput instead of the tuple.
152
+
153
+ Returns:
154
+ The encoded output features.
155
+ """
156
+ self.config.output_attentions = True
157
+
158
+ output_hidden_states = (
159
+ output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
160
+ )
161
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
162
+
163
+ if attention_mask is not None:
164
+ # compute reduced attention_mask corresponding to feature vectors
165
+ attention_mask = self._get_feature_vector_attention_mask(
166
+ extract_features.shape[1], attention_mask, add_adapter=False
167
+ )
168
+
169
+ hidden_states, extract_features = self.feature_projection(extract_features)
170
+ hidden_states = self._mask_hidden_states(
171
+ hidden_states, mask_time_indices=mask_time_indices, attention_mask=attention_mask
172
+ )
173
+
174
+ encoder_outputs = self.encoder(
175
+ hidden_states,
176
+ attention_mask=attention_mask,
177
+ output_attentions=output_attentions,
178
+ output_hidden_states=output_hidden_states,
179
+ return_dict=return_dict,
180
+ )
181
+
182
+ hidden_states = encoder_outputs[0]
183
+
184
+ if self.adapter is not None:
185
+ hidden_states = self.adapter(hidden_states)
186
+
187
+ if not return_dict:
188
+ return (hidden_states, ) + encoder_outputs[1:]
189
+ return BaseModelOutput(
190
+ last_hidden_state=hidden_states,
191
+ hidden_states=encoder_outputs.hidden_states,
192
+ attentions=encoder_outputs.attentions,
193
+ )
194
+
195
+
196
+ def linear_interpolation(features, seq_len):
197
+ """
198
+ Transpose the features to interpolate linearly.
199
+
200
+ Args:
201
+ features (torch.Tensor): The extracted features to be interpolated.
202
+ seq_len (torch.Tensor): The sequence lengths of the features.
203
+
204
+ Returns:
205
+ torch.Tensor: The interpolated features.
206
+ """
207
+ features = features.transpose(1, 2)
208
+ output_features = F.interpolate(features, size=seq_len, align_corners=True, mode='linear')
209
+ return output_features.transpose(1, 2)
hallo/utils/__init__.py ADDED
File without changes
hallo/utils/util.py ADDED
@@ -0,0 +1,616 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # pylint: disable=C0116
2
+ # pylint: disable=W0718
3
+ # pylint: disable=R1732
4
+ """
5
+ utils.py
6
+
7
+ This module provides utility functions for various tasks such as setting random seeds,
8
+ importing modules from files, managing checkpoint files, and saving video files from
9
+ sequences of PIL images.
10
+
11
+ Functions:
12
+ seed_everything(seed)
13
+ import_filename(filename)
14
+ delete_additional_ckpt(base_path, num_keep)
15
+ save_videos_from_pil(pil_images, path, fps=8)
16
+
17
+ Dependencies:
18
+ importlib
19
+ os
20
+ os.path as osp
21
+ random
22
+ shutil
23
+ sys
24
+ pathlib.Path
25
+ av
26
+ cv2
27
+ mediapipe as mp
28
+ numpy as np
29
+ torch
30
+ torchvision
31
+ einops.rearrange
32
+ moviepy.editor.AudioFileClip, VideoClip
33
+ PIL.Image
34
+
35
+ Examples:
36
+ seed_everything(42)
37
+ imported_module = import_filename('path/to/your/module.py')
38
+ delete_additional_ckpt('path/to/checkpoints', 1)
39
+ save_videos_from_pil(pil_images, 'output/video.mp4', fps=12)
40
+
41
+ The functions in this module ensure reproducibility of experiments by seeding random number
42
+ generators, allow dynamic importing of modules, manage checkpoint files by deleting extra ones,
43
+ and provide a way to save sequences of images as video files.
44
+
45
+ Function Details:
46
+ seed_everything(seed)
47
+ Seeds all random number generators to ensure reproducibility.
48
+
49
+ import_filename(filename)
50
+ Imports a module from a given file location.
51
+
52
+ delete_additional_ckpt(base_path, num_keep)
53
+ Deletes additional checkpoint files in the given directory.
54
+
55
+ save_videos_from_pil(pil_images, path, fps=8)
56
+ Saves a sequence of images as a video using the Pillow library.
57
+
58
+ Attributes:
59
+ _ (str): Placeholder for static type checking
60
+ """
61
+
62
+ import importlib
63
+ import os
64
+ import os.path as osp
65
+ import random
66
+ import shutil
67
+ import subprocess
68
+ import sys
69
+ from pathlib import Path
70
+
71
+ import av
72
+ import cv2
73
+ import mediapipe as mp
74
+ import numpy as np
75
+ import torch
76
+ import torchvision
77
+ from einops import rearrange
78
+ from moviepy.editor import AudioFileClip, VideoClip
79
+ from PIL import Image
80
+
81
+
82
+ def seed_everything(seed):
83
+ """
84
+ Seeds all random number generators to ensure reproducibility.
85
+
86
+ Args:
87
+ seed (int): The seed value to set for all random number generators.
88
+ """
89
+ torch.manual_seed(seed)
90
+ torch.cuda.manual_seed_all(seed)
91
+ np.random.seed(seed % (2**32))
92
+ random.seed(seed)
93
+
94
+
95
+ def import_filename(filename):
96
+ """
97
+ Import a module from a given file location.
98
+
99
+ Args:
100
+ filename (str): The path to the file containing the module to be imported.
101
+
102
+ Returns:
103
+ module: The imported module.
104
+
105
+ Raises:
106
+ ImportError: If the module cannot be imported.
107
+
108
+ Example:
109
+ >>> imported_module = import_filename('path/to/your/module.py')
110
+ """
111
+ spec = importlib.util.spec_from_file_location("mymodule", filename)
112
+ module = importlib.util.module_from_spec(spec)
113
+ sys.modules[spec.name] = module
114
+ spec.loader.exec_module(module)
115
+ return module
116
+
117
+
118
+ def delete_additional_ckpt(base_path, num_keep):
119
+ """
120
+ Deletes additional checkpoint files in the given directory.
121
+
122
+ Args:
123
+ base_path (str): The path to the directory containing the checkpoint files.
124
+ num_keep (int): The number of most recent checkpoint files to keep.
125
+
126
+ Returns:
127
+ None
128
+
129
+ Raises:
130
+ FileNotFoundError: If the base_path does not exist.
131
+
132
+ Example:
133
+ >>> delete_additional_ckpt('path/to/checkpoints', 1)
134
+ # This will delete all but the most recent checkpoint file in 'path/to/checkpoints'.
135
+ """
136
+ dirs = []
137
+ for d in os.listdir(base_path):
138
+ if d.startswith("checkpoint-"):
139
+ dirs.append(d)
140
+ num_tot = len(dirs)
141
+ if num_tot <= num_keep:
142
+ return
143
+ # ensure ckpt is sorted and delete the ealier!
144
+ del_dirs = sorted(dirs, key=lambda x: int(
145
+ x.split("-")[-1]))[: num_tot - num_keep]
146
+ for d in del_dirs:
147
+ path_to_dir = osp.join(base_path, d)
148
+ if osp.exists(path_to_dir):
149
+ shutil.rmtree(path_to_dir)
150
+
151
+
152
+ def save_videos_from_pil(pil_images, path, fps=8):
153
+ """
154
+ Save a sequence of images as a video using the Pillow library.
155
+
156
+ Args:
157
+ pil_images (List[PIL.Image]): A list of PIL.Image objects representing the frames of the video.
158
+ path (str): The output file path for the video.
159
+ fps (int, optional): The frames per second rate of the video. Defaults to 8.
160
+
161
+ Returns:
162
+ None
163
+
164
+ Raises:
165
+ ValueError: If the save format is not supported.
166
+
167
+ This function takes a list of PIL.Image objects and saves them as a video file with a specified frame rate.
168
+ The output file format is determined by the file extension of the provided path. Supported formats include
169
+ .mp4, .avi, and .mkv. The function uses the Pillow library to handle the image processing and video
170
+ creation.
171
+ """
172
+ save_fmt = Path(path).suffix
173
+ os.makedirs(os.path.dirname(path), exist_ok=True)
174
+ width, height = pil_images[0].size
175
+
176
+ if save_fmt == ".mp4":
177
+ codec = "libx264"
178
+ container = av.open(path, "w")
179
+ stream = container.add_stream(codec, rate=fps)
180
+
181
+ stream.width = width
182
+ stream.height = height
183
+
184
+ for pil_image in pil_images:
185
+ # pil_image = Image.fromarray(image_arr).convert("RGB")
186
+ av_frame = av.VideoFrame.from_image(pil_image)
187
+ container.mux(stream.encode(av_frame))
188
+ container.mux(stream.encode())
189
+ container.close()
190
+
191
+ elif save_fmt == ".gif":
192
+ pil_images[0].save(
193
+ fp=path,
194
+ format="GIF",
195
+ append_images=pil_images[1:],
196
+ save_all=True,
197
+ duration=(1 / fps * 1000),
198
+ loop=0,
199
+ )
200
+ else:
201
+ raise ValueError("Unsupported file type. Use .mp4 or .gif.")
202
+
203
+
204
+ def save_videos_grid(videos: torch.Tensor, path: str, rescale=False, n_rows=6, fps=8):
205
+ """
206
+ Save a grid of videos as an animation or video.
207
+
208
+ Args:
209
+ videos (torch.Tensor): A tensor of shape (batch_size, channels, time, height, width)
210
+ containing the videos to save.
211
+ path (str): The path to save the video grid. Supported formats are .mp4, .avi, and .gif.
212
+ rescale (bool, optional): If True, rescale the video to the original resolution.
213
+ Defaults to False.
214
+ n_rows (int, optional): The number of rows in the video grid. Defaults to 6.
215
+ fps (int, optional): The frame rate of the saved video. Defaults to 8.
216
+
217
+ Raises:
218
+ ValueError: If the video format is not supported.
219
+
220
+ Returns:
221
+ None
222
+ """
223
+ videos = rearrange(videos, "b c t h w -> t b c h w")
224
+ # height, width = videos.shape[-2:]
225
+ outputs = []
226
+
227
+ for x in videos:
228
+ x = torchvision.utils.make_grid(x, nrow=n_rows) # (c h w)
229
+ x = x.transpose(0, 1).transpose(1, 2).squeeze(-1) # (h w c)
230
+ if rescale:
231
+ x = (x + 1.0) / 2.0 # -1,1 -> 0,1
232
+ x = (x * 255).numpy().astype(np.uint8)
233
+ x = Image.fromarray(x)
234
+
235
+ outputs.append(x)
236
+
237
+ os.makedirs(os.path.dirname(path), exist_ok=True)
238
+
239
+ save_videos_from_pil(outputs, path, fps)
240
+
241
+
242
+ def read_frames(video_path):
243
+ """
244
+ Reads video frames from a given video file.
245
+
246
+ Args:
247
+ video_path (str): The path to the video file.
248
+
249
+ Returns:
250
+ container (av.container.InputContainer): The input container object
251
+ containing the video stream.
252
+
253
+ Raises:
254
+ FileNotFoundError: If the video file is not found.
255
+ RuntimeError: If there is an error in reading the video stream.
256
+
257
+ The function reads the video frames from the specified video file using the
258
+ Python AV library (av). It returns an input container object that contains
259
+ the video stream. If the video file is not found, it raises a FileNotFoundError,
260
+ and if there is an error in reading the video stream, it raises a RuntimeError.
261
+ """
262
+ container = av.open(video_path)
263
+
264
+ video_stream = next(s for s in container.streams if s.type == "video")
265
+ frames = []
266
+ for packet in container.demux(video_stream):
267
+ for frame in packet.decode():
268
+ image = Image.frombytes(
269
+ "RGB",
270
+ (frame.width, frame.height),
271
+ frame.to_rgb().to_ndarray(),
272
+ )
273
+ frames.append(image)
274
+
275
+ return frames
276
+
277
+
278
+ def get_fps(video_path):
279
+ """
280
+ Get the frame rate (FPS) of a video file.
281
+
282
+ Args:
283
+ video_path (str): The path to the video file.
284
+
285
+ Returns:
286
+ int: The frame rate (FPS) of the video file.
287
+ """
288
+ container = av.open(video_path)
289
+ video_stream = next(s for s in container.streams if s.type == "video")
290
+ fps = video_stream.average_rate
291
+ container.close()
292
+ return fps
293
+
294
+
295
+ def tensor_to_video(tensor, output_video_file, audio_source, fps=25):
296
+ """
297
+ Converts a Tensor with shape [c, f, h, w] into a video and adds an audio track from the specified audio file.
298
+
299
+ Args:
300
+ tensor (Tensor): The Tensor to be converted, shaped [c, f, h, w].
301
+ output_video_file (str): The file path where the output video will be saved.
302
+ audio_source (str): The path to the audio file (WAV file) that contains the audio track to be added.
303
+ fps (int): The frame rate of the output video. Default is 25 fps.
304
+ """
305
+ tensor = tensor.permute(1, 2, 3, 0).cpu(
306
+ ).numpy() # convert to [f, h, w, c]
307
+ tensor = np.clip(tensor * 255, 0, 255).astype(
308
+ np.uint8
309
+ ) # to [0, 255]
310
+
311
+ def make_frame(t):
312
+ # get index
313
+ frame_index = min(int(t * fps), tensor.shape[0] - 1)
314
+ return tensor[frame_index]
315
+ new_video_clip = VideoClip(make_frame, duration=tensor.shape[0] / fps)
316
+ audio_clip = AudioFileClip(audio_source).subclip(0, tensor.shape[0] / fps)
317
+ new_video_clip = new_video_clip.set_audio(audio_clip)
318
+ new_video_clip.write_videofile(output_video_file, fps=fps)
319
+
320
+
321
+ silhouette_ids = [
322
+ 10, 338, 297, 332, 284, 251, 389, 356, 454, 323, 361, 288,
323
+ 397, 365, 379, 378, 400, 377, 152, 148, 176, 149, 150, 136,
324
+ 172, 58, 132, 93, 234, 127, 162, 21, 54, 103, 67, 109
325
+ ]
326
+ lip_ids = [61, 185, 40, 39, 37, 0, 267, 269, 270, 409, 291,
327
+ 146, 91, 181, 84, 17, 314, 405, 321, 375]
328
+
329
+
330
+ def compute_face_landmarks(detection_result, h, w):
331
+ """
332
+ Compute face landmarks from a detection result.
333
+
334
+ Args:
335
+ detection_result (mediapipe.solutions.face_mesh.FaceMesh): The detection result containing face landmarks.
336
+ h (int): The height of the video frame.
337
+ w (int): The width of the video frame.
338
+
339
+ Returns:
340
+ face_landmarks_list (list): A list of face landmarks.
341
+ """
342
+ face_landmarks_list = detection_result.face_landmarks
343
+ if len(face_landmarks_list) != 1:
344
+ print("#face is invalid:", len(face_landmarks_list))
345
+ return []
346
+ return [[p.x * w, p.y * h] for p in face_landmarks_list[0]]
347
+
348
+
349
+ def get_landmark(file):
350
+ """
351
+ This function takes a file as input and returns the facial landmarks detected in the file.
352
+
353
+ Args:
354
+ file (str): The path to the file containing the video or image to be processed.
355
+
356
+ Returns:
357
+ Tuple[List[float], List[float]]: A tuple containing two lists of floats representing the x and y coordinates of the facial landmarks.
358
+ """
359
+ model_path = "pretrained_models/face_analysis/models/face_landmarker_v2_with_blendshapes.task"
360
+ BaseOptions = mp.tasks.BaseOptions
361
+ FaceLandmarker = mp.tasks.vision.FaceLandmarker
362
+ FaceLandmarkerOptions = mp.tasks.vision.FaceLandmarkerOptions
363
+ VisionRunningMode = mp.tasks.vision.RunningMode
364
+ # Create a face landmarker instance with the video mode:
365
+ options = FaceLandmarkerOptions(
366
+ base_options=BaseOptions(model_asset_path=model_path),
367
+ running_mode=VisionRunningMode.IMAGE,
368
+ )
369
+
370
+ with FaceLandmarker.create_from_options(options) as landmarker:
371
+ image = mp.Image.create_from_file(str(file))
372
+ height, width = image.height, image.width
373
+ face_landmarker_result = landmarker.detect(image)
374
+ face_landmark = compute_face_landmarks(
375
+ face_landmarker_result, height, width)
376
+
377
+ return np.array(face_landmark), height, width
378
+
379
+
380
+ def get_lip_mask(landmarks, height, width, out_path):
381
+ """
382
+ Extracts the lip region from the given landmarks and saves it as an image.
383
+
384
+ Parameters:
385
+ landmarks (numpy.ndarray): Array of facial landmarks.
386
+ height (int): Height of the output lip mask image.
387
+ width (int): Width of the output lip mask image.
388
+ out_path (pathlib.Path): Path to save the lip mask image.
389
+ """
390
+ lip_landmarks = np.take(landmarks, lip_ids, 0)
391
+ min_xy_lip = np.round(np.min(lip_landmarks, 0))
392
+ max_xy_lip = np.round(np.max(lip_landmarks, 0))
393
+ min_xy_lip[0], max_xy_lip[0], min_xy_lip[1], max_xy_lip[1] = expand_region(
394
+ [min_xy_lip[0], max_xy_lip[0], min_xy_lip[1], max_xy_lip[1]], width, height, 2.0)
395
+ lip_mask = np.zeros((height, width), dtype=np.uint8)
396
+ lip_mask[round(min_xy_lip[1]):round(max_xy_lip[1]),
397
+ round(min_xy_lip[0]):round(max_xy_lip[0])] = 255
398
+ cv2.imwrite(str(out_path), lip_mask)
399
+
400
+
401
+ def get_face_mask(landmarks, height, width, out_path, expand_ratio):
402
+ """
403
+ Generate a face mask based on the given landmarks.
404
+
405
+ Args:
406
+ landmarks (numpy.ndarray): The landmarks of the face.
407
+ height (int): The height of the output face mask image.
408
+ width (int): The width of the output face mask image.
409
+ out_path (pathlib.Path): The path to save the face mask image.
410
+
411
+ Returns:
412
+ None. The face mask image is saved at the specified path.
413
+ """
414
+ face_landmarks = np.take(landmarks, silhouette_ids, 0)
415
+ min_xy_face = np.round(np.min(face_landmarks, 0))
416
+ max_xy_face = np.round(np.max(face_landmarks, 0))
417
+ min_xy_face[0], max_xy_face[0], min_xy_face[1], max_xy_face[1] = expand_region(
418
+ [min_xy_face[0], max_xy_face[0], min_xy_face[1], max_xy_face[1]], width, height, expand_ratio)
419
+ face_mask = np.zeros((height, width), dtype=np.uint8)
420
+ face_mask[round(min_xy_face[1]):round(max_xy_face[1]),
421
+ round(min_xy_face[0]):round(max_xy_face[0])] = 255
422
+ cv2.imwrite(str(out_path), face_mask)
423
+
424
+
425
+ def get_mask(file, cache_dir, face_expand_raio):
426
+ """
427
+ Generate a face mask based on the given landmarks and save it to the specified cache directory.
428
+
429
+ Args:
430
+ file (str): The path to the file containing the landmarks.
431
+ cache_dir (str): The directory to save the generated face mask.
432
+
433
+ Returns:
434
+ None
435
+ """
436
+ landmarks, height, width = get_landmark(file)
437
+ file_name = os.path.basename(file).split(".")[0]
438
+ get_lip_mask(landmarks, height, width, os.path.join(
439
+ cache_dir, f"{file_name}_lip_mask.png"))
440
+ get_face_mask(landmarks, height, width, os.path.join(
441
+ cache_dir, f"{file_name}_face_mask.png"), face_expand_raio)
442
+ get_blur_mask(os.path.join(
443
+ cache_dir, f"{file_name}_face_mask.png"), os.path.join(
444
+ cache_dir, f"{file_name}_face_mask_blur.png"), kernel_size=(51, 51))
445
+ get_blur_mask(os.path.join(
446
+ cache_dir, f"{file_name}_lip_mask.png"), os.path.join(
447
+ cache_dir, f"{file_name}_sep_lip.png"), kernel_size=(31, 31))
448
+ get_background_mask(os.path.join(
449
+ cache_dir, f"{file_name}_face_mask_blur.png"), os.path.join(
450
+ cache_dir, f"{file_name}_sep_background.png"))
451
+ get_sep_face_mask(os.path.join(
452
+ cache_dir, f"{file_name}_face_mask_blur.png"), os.path.join(
453
+ cache_dir, f"{file_name}_sep_lip.png"), os.path.join(
454
+ cache_dir, f"{file_name}_sep_face.png"))
455
+
456
+
457
+ def expand_region(region, image_w, image_h, expand_ratio=1.0):
458
+ """
459
+ Expand the given region by a specified ratio.
460
+ Args:
461
+ region (tuple): A tuple containing the coordinates (min_x, max_x, min_y, max_y) of the region.
462
+ image_w (int): The width of the image.
463
+ image_h (int): The height of the image.
464
+ expand_ratio (float, optional): The ratio by which the region should be expanded. Defaults to 1.0.
465
+
466
+ Returns:
467
+ tuple: A tuple containing the expanded coordinates (min_x, max_x, min_y, max_y) of the region.
468
+ """
469
+
470
+ min_x, max_x, min_y, max_y = region
471
+ mid_x = (max_x + min_x) // 2
472
+ side_len_x = (max_x - min_x) * expand_ratio
473
+ mid_y = (max_y + min_y) // 2
474
+ side_len_y = (max_y - min_y) * expand_ratio
475
+ min_x = mid_x - side_len_x // 2
476
+ max_x = mid_x + side_len_x // 2
477
+ min_y = mid_y - side_len_y // 2
478
+ max_y = mid_y + side_len_y // 2
479
+ if min_x < 0:
480
+ max_x -= min_x
481
+ min_x = 0
482
+ if max_x > image_w:
483
+ min_x -= max_x - image_w
484
+ max_x = image_w
485
+ if min_y < 0:
486
+ max_y -= min_y
487
+ min_y = 0
488
+ if max_y > image_h:
489
+ min_y -= max_y - image_h
490
+ max_y = image_h
491
+
492
+ return round(min_x), round(max_x), round(min_y), round(max_y)
493
+
494
+
495
+ def get_blur_mask(file_path, output_file_path, resize_dim=(64, 64), kernel_size=(101, 101)):
496
+ """
497
+ Read, resize, blur, normalize, and save an image.
498
+
499
+ Parameters:
500
+ file_path (str): Path to the input image file.
501
+ output_dir (str): Path to the output directory to save blurred images.
502
+ resize_dim (tuple): Dimensions to resize the images to.
503
+ kernel_size (tuple): Size of the kernel to use for Gaussian blur.
504
+ """
505
+ # Read the mask image
506
+ mask = cv2.imread(file_path, cv2.IMREAD_GRAYSCALE)
507
+
508
+ # Check if the image is loaded successfully
509
+ if mask is not None:
510
+ # Resize the mask image
511
+ resized_mask = cv2.resize(mask, resize_dim)
512
+ # Apply Gaussian blur to the resized mask image
513
+ blurred_mask = cv2.GaussianBlur(resized_mask, kernel_size, 0)
514
+ # Normalize the blurred image
515
+ normalized_mask = cv2.normalize(
516
+ blurred_mask, None, 0, 255, cv2.NORM_MINMAX)
517
+ # Save the normalized mask image
518
+ cv2.imwrite(output_file_path, normalized_mask)
519
+ return f"Processed, normalized, and saved: {output_file_path}"
520
+ return f"Failed to load image: {file_path}"
521
+
522
+
523
+ def get_background_mask(file_path, output_file_path):
524
+ """
525
+ Read an image, invert its values, and save the result.
526
+
527
+ Parameters:
528
+ file_path (str): Path to the input image file.
529
+ output_dir (str): Path to the output directory to save the inverted image.
530
+ """
531
+ # Read the image
532
+ image = cv2.imread(file_path, cv2.IMREAD_GRAYSCALE)
533
+
534
+ if image is None:
535
+ print(f"Failed to load image: {file_path}")
536
+ return
537
+
538
+ # Invert the image
539
+ inverted_image = 1.0 - (
540
+ image / 255.0
541
+ ) # Assuming the image values are in [0, 255] range
542
+ # Convert back to uint8
543
+ inverted_image = (inverted_image * 255).astype(np.uint8)
544
+
545
+ # Save the inverted image
546
+ cv2.imwrite(output_file_path, inverted_image)
547
+ print(f"Processed and saved: {output_file_path}")
548
+
549
+
550
+ def get_sep_face_mask(file_path1, file_path2, output_file_path):
551
+ """
552
+ Read two images, subtract the second one from the first, and save the result.
553
+
554
+ Parameters:
555
+ output_dir (str): Path to the output directory to save the subtracted image.
556
+ """
557
+
558
+ # Read the images
559
+ mask1 = cv2.imread(file_path1, cv2.IMREAD_GRAYSCALE)
560
+ mask2 = cv2.imread(file_path2, cv2.IMREAD_GRAYSCALE)
561
+
562
+ if mask1 is None or mask2 is None:
563
+ print(f"Failed to load images: {file_path1}")
564
+ return
565
+
566
+ # Ensure the images are the same size
567
+ if mask1.shape != mask2.shape:
568
+ print(
569
+ f"Image shapes do not match for {file_path1}: {mask1.shape} vs {mask2.shape}"
570
+ )
571
+ return
572
+
573
+ # Subtract the second mask from the first
574
+ result_mask = cv2.subtract(mask1, mask2)
575
+
576
+ # Save the result mask image
577
+ cv2.imwrite(output_file_path, result_mask)
578
+ print(f"Processed and saved: {output_file_path}")
579
+
580
+ def resample_audio(input_audio_file: str, output_audio_file: str, sample_rate: int):
581
+ p = subprocess.Popen([
582
+ "ffmpeg", "-y", "-v", "error", "-i", input_audio_file, "-ar", str(sample_rate), output_audio_file
583
+ ])
584
+ ret = p.wait()
585
+ assert ret == 0, "Resample audio failed!"
586
+ return output_audio_file
587
+
588
+ def get_face_region(image_path: str, detector):
589
+ try:
590
+ image = cv2.imread(image_path)
591
+ if image is None:
592
+ print(f"Failed to open image: {image_path}. Skipping...")
593
+ return None, None
594
+
595
+ mp_image = mp.Image(image_format=mp.ImageFormat.SRGB, data=image)
596
+ detection_result = detector.detect(mp_image)
597
+
598
+ # Adjust mask creation for the three-channel image
599
+ mask = np.zeros_like(image, dtype=np.uint8)
600
+
601
+ for detection in detection_result.detections:
602
+ bbox = detection.bounding_box
603
+ start_point = (int(bbox.origin_x), int(bbox.origin_y))
604
+ end_point = (int(bbox.origin_x + bbox.width),
605
+ int(bbox.origin_y + bbox.height))
606
+ cv2.rectangle(mask, start_point, end_point,
607
+ (255, 255, 255), thickness=-1)
608
+
609
+ save_path = image_path.replace("images", "face_masks")
610
+ os.makedirs(os.path.dirname(save_path), exist_ok=True)
611
+ cv2.imwrite(save_path, mask)
612
+ # print(f"Processed and saved {save_path}")
613
+ return image_path, mask
614
+ except Exception as e:
615
+ print(f"Error processing image {image_path}: {e}")
616
+ return None, None
requirements.txt ADDED
@@ -0,0 +1,30 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ accelerate==0.28.0
2
+ audio-separator==0.17.2
3
+ av==12.1.0
4
+ bitsandbytes==0.43.1
5
+ decord==0.6.0
6
+ diffusers==0.27.2
7
+ einops==0.8.0
8
+ insightface==0.7.3
9
+ librosa==0.10.2.post1
10
+ mediapipe[vision]==0.10.14
11
+ mlflow==2.13.1
12
+ moviepy==1.0.3
13
+ numpy==1.26.4
14
+ omegaconf==2.3.0
15
+ onnx2torch==1.5.14
16
+ onnx==1.16.1
17
+ onnxruntime==1.18.0
18
+ opencv-contrib-python==4.9.0.80
19
+ opencv-python-headless==4.9.0.80
20
+ opencv-python==4.9.0.80
21
+ pillow==10.3.0
22
+ setuptools==70.0.0
23
+ torch==2.2.2
24
+ torchvision==0.17.2
25
+ tqdm==4.66.4
26
+ transformers==4.39.2
27
+ xformers==0.0.25.post1
28
+ isort==5.13.2
29
+ pylint==3.2.2
30
+ pre-commit==3.7.1
scripts/inference.py ADDED
@@ -0,0 +1,372 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # pylint: disable=E1101
2
+ # scripts/inference.py
3
+
4
+ """
5
+ This script contains the main inference pipeline for processing audio and image inputs to generate a video output.
6
+
7
+ The script imports necessary packages and classes, defines a neural network model,
8
+ and contains functions for processing audio embeddings and performing inference.
9
+
10
+ The main inference process is outlined in the following steps:
11
+ 1. Initialize the configuration.
12
+ 2. Set up runtime variables.
13
+ 3. Prepare the input data for inference (source image, face mask, and face embeddings).
14
+ 4. Process the audio embeddings.
15
+ 5. Build and freeze the model and scheduler.
16
+ 6. Run the inference loop and save the result.
17
+
18
+ Usage:
19
+ This script can be run from the command line with the following arguments:
20
+ - audio_path: Path to the audio file.
21
+ - image_path: Path to the source image.
22
+ - face_mask_path: Path to the face mask image.
23
+ - face_emb_path: Path to the face embeddings file.
24
+ - output_path: Path to save the output video.
25
+
26
+ Example:
27
+ python scripts/inference.py --audio_path audio.wav --image_path image.jpg
28
+ --face_mask_path face_mask.png --face_emb_path face_emb.pt --output_path output.mp4
29
+ """
30
+
31
+ import argparse
32
+ import os
33
+
34
+ import torch
35
+ from diffusers import AutoencoderKL, DDIMScheduler
36
+ from omegaconf import OmegaConf
37
+ from torch import nn
38
+
39
+ from hallo.animate.face_animate import FaceAnimatePipeline
40
+ from hallo.datasets.audio_processor import AudioProcessor
41
+ from hallo.datasets.image_processor import ImageProcessor
42
+ from hallo.models.audio_proj import AudioProjModel
43
+ from hallo.models.face_locator import FaceLocator
44
+ from hallo.models.image_proj import ImageProjModel
45
+ from hallo.models.unet_2d_condition import UNet2DConditionModel
46
+ from hallo.models.unet_3d import UNet3DConditionModel
47
+ from hallo.utils.util import tensor_to_video
48
+
49
+
50
+ class Net(nn.Module):
51
+ """
52
+ The Net class combines all the necessary modules for the inference process.
53
+
54
+ Args:
55
+ reference_unet (UNet2DConditionModel): The UNet2DConditionModel used as a reference for inference.
56
+ denoising_unet (UNet3DConditionModel): The UNet3DConditionModel used for denoising the input audio.
57
+ face_locator (FaceLocator): The FaceLocator model used to locate the face in the input image.
58
+ imageproj (nn.Module): The ImageProjector model used to project the source image onto the face.
59
+ audioproj (nn.Module): The AudioProjector model used to project the audio embeddings onto the face.
60
+ """
61
+ def __init__(
62
+ self,
63
+ reference_unet: UNet2DConditionModel,
64
+ denoising_unet: UNet3DConditionModel,
65
+ face_locator: FaceLocator,
66
+ imageproj,
67
+ audioproj,
68
+ ):
69
+ super().__init__()
70
+ self.reference_unet = reference_unet
71
+ self.denoising_unet = denoising_unet
72
+ self.face_locator = face_locator
73
+ self.imageproj = imageproj
74
+ self.audioproj = audioproj
75
+
76
+ def forward(self,):
77
+ """
78
+ empty function to override abstract function of nn Module
79
+ """
80
+
81
+ def get_modules(self):
82
+ """
83
+ Simple method to avoid too-few-public-methods pylint error
84
+ """
85
+ return {
86
+ "reference_unet": self.reference_unet,
87
+ "denoising_unet": self.denoising_unet,
88
+ "face_locator": self.face_locator,
89
+ "imageproj": self.imageproj,
90
+ "audioproj": self.audioproj,
91
+ }
92
+
93
+
94
+ def process_audio_emb(audio_emb):
95
+ """
96
+ Process the audio embedding to concatenate with other tensors.
97
+
98
+ Parameters:
99
+ audio_emb (torch.Tensor): The audio embedding tensor to process.
100
+
101
+ Returns:
102
+ concatenated_tensors (List[torch.Tensor]): The concatenated tensor list.
103
+ """
104
+ concatenated_tensors = []
105
+
106
+ for i in range(audio_emb.shape[0]):
107
+ vectors_to_concat = [
108
+ audio_emb[max(min(i + j, audio_emb.shape[0]-1), 0)]for j in range(-2, 3)]
109
+ concatenated_tensors.append(torch.stack(vectors_to_concat, dim=0))
110
+
111
+ audio_emb = torch.stack(concatenated_tensors, dim=0)
112
+
113
+ return audio_emb
114
+
115
+
116
+
117
+ def inference_process(args: argparse.Namespace):
118
+ """
119
+ Perform inference processing.
120
+
121
+ Args:
122
+ args (argparse.Namespace): Command-line arguments.
123
+
124
+ This function initializes the configuration for the inference process. It sets up the necessary
125
+ modules and variables to prepare for the upcoming inference steps.
126
+ """
127
+ # 1. init config
128
+ config = OmegaConf.load(args.config)
129
+ config = OmegaConf.merge(config, vars(args))
130
+ source_image_path = config.source_image
131
+ driving_audio_path = config.driving_audio
132
+ save_path = config.save_path
133
+ if not os.path.exists(save_path):
134
+ os.makedirs(save_path)
135
+ motion_scale = [config.pose_weight, config.face_weight, config.lip_weight]
136
+ if args.checkpoint is not None:
137
+ config.audio_ckpt_dir = args.checkpoint
138
+ # 2. runtime variables
139
+ device = torch.device(
140
+ "cuda") if torch.cuda.is_available() else torch.device("cpu")
141
+ if config.weight_dtype == "fp16":
142
+ weight_dtype = torch.float16
143
+ elif config.weight_dtype == "bf16":
144
+ weight_dtype = torch.bfloat16
145
+ elif config.weight_dtype == "fp32":
146
+ weight_dtype = torch.float32
147
+ else:
148
+ weight_dtype = torch.float32
149
+
150
+ # 3. prepare inference data
151
+ # 3.1 prepare source image, face mask, face embeddings
152
+ img_size = (config.data.source_image.width,
153
+ config.data.source_image.height)
154
+ clip_length = config.data.n_sample_frames
155
+ face_analysis_model_path = config.face_analysis.model_path
156
+ with ImageProcessor(img_size, face_analysis_model_path) as image_processor:
157
+ source_image_pixels, \
158
+ source_image_face_region, \
159
+ source_image_face_emb, \
160
+ source_image_full_mask, \
161
+ source_image_face_mask, \
162
+ source_image_lip_mask = image_processor.preprocess(
163
+ source_image_path, save_path, config.face_expand_ratio)
164
+
165
+ # 3.2 prepare audio embeddings
166
+ sample_rate = config.data.driving_audio.sample_rate
167
+ assert sample_rate == 16000, "audio sample rate must be 16000"
168
+ fps = config.data.export_video.fps
169
+ wav2vec_model_path = config.wav2vec.model_path
170
+ wav2vec_only_last_features = config.wav2vec.features == "last"
171
+ audio_separator_model_file = config.audio_separator.model_path
172
+ with AudioProcessor(
173
+ sample_rate,
174
+ fps,
175
+ wav2vec_model_path,
176
+ wav2vec_only_last_features,
177
+ os.path.dirname(audio_separator_model_file),
178
+ os.path.basename(audio_separator_model_file),
179
+ os.path.join(save_path, "audio_preprocess")
180
+ ) as audio_processor:
181
+ audio_emb = audio_processor.preprocess(driving_audio_path)
182
+
183
+ # 4. build modules
184
+ sched_kwargs = OmegaConf.to_container(config.noise_scheduler_kwargs)
185
+ if config.enable_zero_snr:
186
+ sched_kwargs.update(
187
+ rescale_betas_zero_snr=True,
188
+ timestep_spacing="trailing",
189
+ prediction_type="v_prediction",
190
+ )
191
+ val_noise_scheduler = DDIMScheduler(**sched_kwargs)
192
+ sched_kwargs.update({"beta_schedule": "scaled_linear"})
193
+
194
+ vae = AutoencoderKL.from_pretrained(config.vae.model_path)
195
+ reference_unet = UNet2DConditionModel.from_pretrained(
196
+ config.base_model_path, subfolder="unet")
197
+ denoising_unet = UNet3DConditionModel.from_pretrained_2d(
198
+ config.base_model_path,
199
+ config.motion_module_path,
200
+ subfolder="unet",
201
+ unet_additional_kwargs=OmegaConf.to_container(
202
+ config.unet_additional_kwargs),
203
+ use_landmark=False,
204
+ )
205
+ face_locator = FaceLocator(conditioning_embedding_channels=320)
206
+ image_proj = ImageProjModel(
207
+ cross_attention_dim=denoising_unet.config.cross_attention_dim,
208
+ clip_embeddings_dim=512,
209
+ clip_extra_context_tokens=4,
210
+ )
211
+
212
+ audio_proj = AudioProjModel(
213
+ seq_len=5,
214
+ blocks=12, # use 12 layers' hidden states of wav2vec
215
+ channels=768, # audio embedding channel
216
+ intermediate_dim=512,
217
+ output_dim=768,
218
+ context_tokens=32,
219
+ ).to(device=device, dtype=weight_dtype)
220
+
221
+ audio_ckpt_dir = config.audio_ckpt_dir
222
+
223
+
224
+ # Freeze
225
+ vae.requires_grad_(False)
226
+ image_proj.requires_grad_(False)
227
+ reference_unet.requires_grad_(False)
228
+ denoising_unet.requires_grad_(False)
229
+ face_locator.requires_grad_(False)
230
+ audio_proj.requires_grad_(False)
231
+
232
+ reference_unet.enable_gradient_checkpointing()
233
+ denoising_unet.enable_gradient_checkpointing()
234
+
235
+ net = Net(
236
+ reference_unet,
237
+ denoising_unet,
238
+ face_locator,
239
+ image_proj,
240
+ audio_proj,
241
+ )
242
+
243
+ m,u = net.load_state_dict(
244
+ torch.load(
245
+ os.path.join(audio_ckpt_dir, "net.pth"),
246
+ map_location="cpu",
247
+ ),
248
+ )
249
+ assert len(m) == 0 and len(u) == 0, "Fail to load correct checkpoint."
250
+ print("loaded weight from ", os.path.join(audio_ckpt_dir, "net.pth"))
251
+
252
+ # 5. inference
253
+ pipeline = FaceAnimatePipeline(
254
+ vae=vae,
255
+ reference_unet=net.reference_unet,
256
+ denoising_unet=net.denoising_unet,
257
+ face_locator=net.face_locator,
258
+ scheduler=val_noise_scheduler,
259
+ image_proj=net.imageproj,
260
+ )
261
+ pipeline.to(device=device, dtype=weight_dtype)
262
+
263
+ audio_emb = process_audio_emb(audio_emb)
264
+
265
+ source_image_pixels = source_image_pixels.unsqueeze(0)
266
+ source_image_face_region = source_image_face_region.unsqueeze(0)
267
+ source_image_face_emb = source_image_face_emb.reshape(1, -1)
268
+ source_image_face_emb = torch.tensor(source_image_face_emb)
269
+
270
+ source_image_full_mask = [
271
+ (mask.repeat(clip_length, 1))
272
+ for mask in source_image_full_mask
273
+ ]
274
+ source_image_face_mask = [
275
+ (mask.repeat(clip_length, 1))
276
+ for mask in source_image_face_mask
277
+ ]
278
+ source_image_lip_mask = [
279
+ (mask.repeat(clip_length, 1))
280
+ for mask in source_image_lip_mask
281
+ ]
282
+
283
+
284
+ times = audio_emb.shape[0] // clip_length
285
+
286
+ tensor_result = []
287
+
288
+ generator = torch.manual_seed(42)
289
+
290
+ for t in range(times):
291
+
292
+ if len(tensor_result) == 0:
293
+ # The first iteration
294
+ motion_zeros = source_image_pixels.repeat(
295
+ config.data.n_motion_frames, 1, 1, 1)
296
+ motion_zeros = motion_zeros.to(
297
+ dtype=source_image_pixels.dtype, device=source_image_pixels.device)
298
+ pixel_values_ref_img = torch.cat(
299
+ [source_image_pixels, motion_zeros], dim=0) # concat the ref image and the first motion frames
300
+ else:
301
+ motion_frames = tensor_result[-1][0]
302
+ motion_frames = motion_frames.permute(1, 0, 2, 3)
303
+ motion_frames = motion_frames[0-config.data.n_motion_frames:]
304
+ motion_frames = motion_frames * 2.0 - 1.0
305
+ motion_frames = motion_frames.to(
306
+ dtype=source_image_pixels.dtype, device=source_image_pixels.device)
307
+ pixel_values_ref_img = torch.cat(
308
+ [source_image_pixels, motion_frames], dim=0) # concat the ref image and the motion frames
309
+
310
+ pixel_values_ref_img = pixel_values_ref_img.unsqueeze(0)
311
+
312
+ audio_tensor = audio_emb[
313
+ t * clip_length: min((t + 1) * clip_length, audio_emb.shape[0])
314
+ ]
315
+ audio_tensor = audio_tensor.unsqueeze(0)
316
+ audio_tensor = audio_tensor.to(
317
+ device=net.audioproj.device, dtype=net.audioproj.dtype)
318
+ audio_tensor = net.audioproj(audio_tensor)
319
+
320
+ pipeline_output = pipeline(
321
+ ref_image=pixel_values_ref_img,
322
+ audio_tensor=audio_tensor,
323
+ face_emb=source_image_face_emb,
324
+ face_mask=source_image_face_region,
325
+ pixel_values_full_mask=source_image_full_mask,
326
+ pixel_values_face_mask=source_image_face_mask,
327
+ pixel_values_lip_mask=source_image_lip_mask,
328
+ width=img_size[0],
329
+ height=img_size[1],
330
+ video_length=clip_length,
331
+ num_inference_steps=config.inference_steps,
332
+ guidance_scale=config.cfg_scale,
333
+ generator=generator,
334
+ motion_scale=motion_scale,
335
+ )
336
+
337
+ tensor_result.append(pipeline_output.videos)
338
+
339
+ tensor_result = torch.cat(tensor_result, dim=2)
340
+ tensor_result = tensor_result.squeeze(0)
341
+
342
+ output_file = config.output
343
+ # save the result after all iteration
344
+ tensor_to_video(tensor_result, output_file, driving_audio_path)
345
+
346
+
347
+ if __name__ == "__main__":
348
+ parser = argparse.ArgumentParser()
349
+
350
+ parser.add_argument(
351
+ "-c", "--config", default="configs/inference/default.yaml")
352
+ parser.add_argument("--source_image", type=str, required=False,
353
+ help="source image", default="test_data/source_images/6.jpg")
354
+ parser.add_argument("--driving_audio", type=str, required=False,
355
+ help="driving audio", default="test_data/driving_audios/singing/sing_4.wav")
356
+ parser.add_argument(
357
+ "--output", type=str, help="output video file name", default=".cache/output.mp4")
358
+ parser.add_argument(
359
+ "--pose_weight", type=float, help="weight of pose", default=1.0)
360
+ parser.add_argument(
361
+ "--face_weight", type=float, help="weight of face", default=1.0)
362
+ parser.add_argument(
363
+ "--lip_weight", type=float, help="weight of lip", default=1.0)
364
+ parser.add_argument(
365
+ "--face_expand_ratio", type=float, help="face region", default=1.2)
366
+ parser.add_argument(
367
+ "--checkpoint", type=str, help="which checkpoint", default=None)
368
+
369
+
370
+ command_line_args = parser.parse_args()
371
+
372
+ inference_process(command_line_args)
setup.py ADDED
@@ -0,0 +1,55 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """
2
+ setup.py
3
+ ----
4
+ This is the main setup file for the hallo face animation project. It defines the package
5
+ metadata, required dependencies, and provides the entry point for installing the package.
6
+
7
+ """
8
+
9
+ # -*- coding: utf-8 -*-
10
+ from setuptools import setup
11
+
12
+ packages = \
13
+ ['hallo', 'hallo.datasets', 'hallo.models', 'hallo.animate', 'hallo.utils']
14
+
15
+ package_data = \
16
+ {'': ['*']}
17
+
18
+ install_requires = \
19
+ ['accelerate==0.28.0',
20
+ 'audio-separator>=0.17.2,<0.18.0',
21
+ 'av==12.1.0',
22
+ 'bitsandbytes==0.43.1',
23
+ 'decord==0.6.0',
24
+ 'diffusers==0.27.2',
25
+ 'einops>=0.8.0,<0.9.0',
26
+ 'insightface>=0.7.3,<0.8.0',
27
+ 'mediapipe[vision]>=0.10.14,<0.11.0',
28
+ 'mlflow==2.13.1',
29
+ 'moviepy>=1.0.3,<2.0.0',
30
+ 'omegaconf>=2.3.0,<3.0.0',
31
+ 'opencv-python>=4.9.0.80,<5.0.0.0',
32
+ 'pillow>=10.3.0,<11.0.0',
33
+ 'torch==2.2.2',
34
+ 'torchvision==0.17.2',
35
+ 'transformers==4.39.2',
36
+ 'xformers==0.0.25.post1']
37
+
38
+ setup_kwargs = {
39
+ 'name': 'anna',
40
+ 'version': '0.1.0',
41
+ 'description': '',
42
+ 'long_description': '# Anna face animation',
43
+ 'author': 'Your Name',
44
+ 'author_email': 'you@example.com',
45
+ 'maintainer': 'None',
46
+ 'maintainer_email': 'None',
47
+ 'url': 'None',
48
+ 'packages': packages,
49
+ 'package_data': package_data,
50
+ 'install_requires': install_requires,
51
+ 'python_requires': '>=3.10,<4.0',
52
+ }
53
+
54
+
55
+ setup(**setup_kwargs)