File size: 13,995 Bytes
c83dd81
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
# Adapted from https://github.com/magic-research/magic-animate/blob/main/magicanimate/models/mutual_self_attention.py
from typing import Any, Dict, Optional

import torch
from einops import rearrange

from src.models.attention import TemporalBasicTransformerBlock

from .attention import BasicTransformerBlock


def torch_dfs(model: torch.nn.Module):
    result = [model]
    for child in model.children():
        result += torch_dfs(child)
    return result


class ReferenceAttentionControl:
    def __init__(

        self,

        unet,

        mode="write",

        do_classifier_free_guidance=False,

        attention_auto_machine_weight=float("inf"),

        gn_auto_machine_weight=1.0,

        style_fidelity=1.0,

        reference_attn=True,

        reference_adain=False,

        fusion_blocks="midup",

        batch_size=1,

    ) -> None:
        # 10. Modify self attention and group norm
        self.unet = unet
        assert mode in ["read", "write"]
        assert fusion_blocks in ["midup", "full"]
        self.reference_attn = reference_attn
        self.reference_adain = reference_adain
        self.fusion_blocks = fusion_blocks
        self.register_reference_hooks(
            mode,
            do_classifier_free_guidance,
            attention_auto_machine_weight,
            gn_auto_machine_weight,
            style_fidelity,
            reference_attn,
            reference_adain,
            fusion_blocks,
            batch_size=batch_size,
        )

    def register_reference_hooks(

        self,

        mode,

        do_classifier_free_guidance,

        attention_auto_machine_weight,

        gn_auto_machine_weight,

        style_fidelity,

        reference_attn,

        reference_adain,

        dtype=torch.float16,

        batch_size=1,

        num_images_per_prompt=1,

        device=torch.device("cuda"),

        fusion_blocks="midup",

    ):
        MODE = mode
        do_classifier_free_guidance = do_classifier_free_guidance
        attention_auto_machine_weight = attention_auto_machine_weight
        gn_auto_machine_weight = gn_auto_machine_weight
        style_fidelity = style_fidelity
        reference_attn = reference_attn
        reference_adain = reference_adain
        fusion_blocks = fusion_blocks
        num_images_per_prompt = num_images_per_prompt
        dtype = dtype
        if do_classifier_free_guidance:
            uc_mask = (
                torch.Tensor(
                    [1] * batch_size * num_images_per_prompt * 16
                    + [0] * batch_size * num_images_per_prompt * 16
                )
                .to(device)
                .bool()
            )
        else:
            uc_mask = (
                torch.Tensor([0] * batch_size * num_images_per_prompt * 2)
                .to(device)
                .bool()
            )

        def hacked_basic_transformer_inner_forward(

            self,

            hidden_states: torch.FloatTensor,

            attention_mask: Optional[torch.FloatTensor] = None,

            audio_cond_fea: Optional[torch.FloatTensor] = None,

            encoder_hidden_states: Optional[torch.FloatTensor] = None,

            encoder_attention_mask: Optional[torch.FloatTensor] = None,

            timestep: Optional[torch.LongTensor] = None,

            cross_attention_kwargs: Dict[str, Any] = None,

            class_labels: Optional[torch.LongTensor] = None,

            video_length=None,

            audio_feature_ratio = 3.0

        ):
            if self.use_ada_layer_norm:  # False
                norm_hidden_states = self.norm1(hidden_states, timestep)
            elif self.use_ada_layer_norm_zero:
                (
                    norm_hidden_states,
                    gate_msa,
                    shift_mlp,
                    scale_mlp,
                    gate_mlp,
                ) = self.norm1(
                    hidden_states,
                    timestep,
                    class_labels,
                    hidden_dtype=hidden_states.dtype,
                )
            else:
                norm_hidden_states = self.norm1(hidden_states)

            # 1. Self-Attention
            # self.only_cross_attention = False
            cross_attention_kwargs = (
                cross_attention_kwargs if cross_attention_kwargs is not None else {}
            )
            if self.only_cross_attention:
                attn_output = self.attn1(
                    norm_hidden_states,
                    attention_mask=attention_mask,
                    **cross_attention_kwargs,
                )
            else:
                if MODE == "write":
                    self.bank.append(norm_hidden_states.clone())
                    attn_output = self.attn1(
                        norm_hidden_states,
                        attention_mask=attention_mask,
                        **cross_attention_kwargs,
                    )
                if MODE == "read":
                    bank_feas = [
                        rearrange(
                            d.unsqueeze(1).repeat(1, video_length, 1, 1),
                            "b t l c -> (b t) l c",
                        )
                        for d in self.bank
                    ]
                    modify_norm_hidden_states = torch.cat(
                        [norm_hidden_states] + bank_feas, dim=1
                    )
                    # print(f"modify_norm_hidden_states:{modify_norm_hidden_states.shape}")

                    hidden_states_uc = (
                        self.attn1(
                            norm_hidden_states,
                            encoder_hidden_states=modify_norm_hidden_states,
                            attention_mask=attention_mask,
                        )
                        + hidden_states
                    )
                    if do_classifier_free_guidance:
                        hidden_states_c = hidden_states_uc.clone()
                        _uc_mask = uc_mask.clone()
                        # print(hidden_states_c.shape, _uc_mask.shape)
                        if hidden_states.shape[0] != _uc_mask.shape[0]:
                            _uc_mask = (
                                torch.Tensor(
                                    [1] * (hidden_states.shape[0] // 2)
                                    + [0] * (hidden_states.shape[0] // 2)
                                )
                                .to(device)
                                .bool()
                            )
                        # print(hidden_states_c.shape, norm_hidden_states.shape, hidden_states.shape, _uc_mask.shape)
                        hidden_states_c[_uc_mask] = (
                            self.attn1(
                                norm_hidden_states[_uc_mask],
                                encoder_hidden_states=norm_hidden_states[_uc_mask], # B * 4096 * 768
                                attention_mask=attention_mask,
                            )
                            + hidden_states[_uc_mask]
                        )
                        hidden_states = hidden_states_c.clone()
                    else:
                        hidden_states = hidden_states_uc

                    # self.bank.clear()
                    if self.attn2 is not None:
                        # Ref Cross-Attention
                        norm_hidden_states = (
                            self.norm2(hidden_states, timestep)
                            if self.use_ada_layer_norm
                            else self.norm2(hidden_states)
                        )
                        # print("Audio Cross-Attention shapes:", norm_hidden_states.shape, audio_cond_fea.shape)
                        if audio_feature_ratio > 0:
                            # print('#'*5, norm_hidden_states.shape, audio_cond_fea.shape)
                            hidden_states = (
                                self.attn2(
                                    norm_hidden_states,
                                    encoder_hidden_states=audio_cond_fea, # B * 50 * 768,
                                    attention_mask=attention_mask,
                                ) * audio_feature_ratio
                                + hidden_states
                            )
                        # print("Audio Cross-Attention max after:", hidden_states.max())


                    # Feed-forward
                    hidden_states = self.ff(self.norm3(hidden_states)) + hidden_states

                    # Temporal-Attention
                    return hidden_states

            if self.use_ada_layer_norm_zero:
                attn_output = gate_msa.unsqueeze(1) * attn_output
            hidden_states = attn_output + hidden_states

            # 3. Feed-forward
            norm_hidden_states = self.norm3(hidden_states)

            if self.use_ada_layer_norm_zero:
                norm_hidden_states = (
                    norm_hidden_states * (1 + scale_mlp[:, None]) + shift_mlp[:, None]
                )

            ff_output = self.ff(norm_hidden_states)

            if self.use_ada_layer_norm_zero:
                ff_output = gate_mlp.unsqueeze(1) * ff_output

            hidden_states = ff_output + hidden_states

            return hidden_states

        if self.reference_attn:
            if self.fusion_blocks == "midup":
                attn_modules = [
                    module
                    for module in (
                        torch_dfs(self.unet.mid_block) + torch_dfs(self.unet.up_blocks)
                    )
                    if isinstance(module, BasicTransformerBlock)
                    or isinstance(module, TemporalBasicTransformerBlock)
                ]
            elif self.fusion_blocks == "full":
                attn_modules = [
                    module
                    for module in torch_dfs(self.unet)
                    if isinstance(module, BasicTransformerBlock)
                    or isinstance(module, TemporalBasicTransformerBlock)
                ]
            attn_modules = sorted(
                attn_modules, key=lambda x: -x.norm1.normalized_shape[0]
            )

            for i, module in enumerate(attn_modules):
                module._original_inner_forward = module.forward
                if isinstance(module, BasicTransformerBlock):
                    module.forward = hacked_basic_transformer_inner_forward.__get__(
                        module, BasicTransformerBlock
                    )
                if isinstance(module, TemporalBasicTransformerBlock):
                    module.forward = hacked_basic_transformer_inner_forward.__get__(
                        module, TemporalBasicTransformerBlock
                    )

                module.bank = []
                module.attn_weight = float(i) / float(len(attn_modules))

    def update(self, writer, do_classifier_free_guidance=False, dtype=torch.float16):
        if self.reference_attn:
            if self.fusion_blocks == "midup":
                reader_attn_modules = [
                    module
                    for module in (
                        torch_dfs(self.unet.mid_block) + torch_dfs(self.unet.up_blocks)
                    )
                    if isinstance(module, TemporalBasicTransformerBlock)
                ]
                writer_attn_modules = [
                    module
                    for module in (
                        torch_dfs(writer.unet.mid_block)
                        + torch_dfs(writer.unet.up_blocks)
                    )
                    if isinstance(module, BasicTransformerBlock)
                ]
            elif self.fusion_blocks == "full":
                reader_attn_modules = [
                    module
                    for module in torch_dfs(self.unet)
                    if isinstance(module, TemporalBasicTransformerBlock)
                ]
                writer_attn_modules = [
                    module
                    for module in torch_dfs(writer.unet)
                    if isinstance(module, BasicTransformerBlock)
                ]
            reader_attn_modules = sorted(
                reader_attn_modules, key=lambda x: -x.norm1.normalized_shape[0]
            )
            writer_attn_modules = sorted(
                writer_attn_modules, key=lambda x: -x.norm1.normalized_shape[0]
            )
            for r, w in zip(reader_attn_modules, writer_attn_modules):
                if do_classifier_free_guidance:
                    r.bank = [torch.cat([v, v]).to(dtype) for v in w.bank]
                else:
                    r.bank = [v.clone().to(dtype) for v in w.bank]

    def clear(self):
        if self.reference_attn:
            if self.fusion_blocks == "midup":
                reader_attn_modules = [
                    module
                    for module in (
                        torch_dfs(self.unet.mid_block) + torch_dfs(self.unet.up_blocks)
                    )
                    if isinstance(module, BasicTransformerBlock)
                    or isinstance(module, TemporalBasicTransformerBlock)
                ]
            elif self.fusion_blocks == "full":
                reader_attn_modules = [
                    module
                    for module in torch_dfs(self.unet)
                    if isinstance(module, BasicTransformerBlock)
                    or isinstance(module, TemporalBasicTransformerBlock)
                ]
            reader_attn_modules = sorted(
                reader_attn_modules, key=lambda x: -x.norm1.normalized_shape[0]
            )
            for r in reader_attn_modules:
                r.bank.clear()