File size: 11,667 Bytes
87c1b97
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
from contextlib import contextmanager
from typing import Any, Dict, List, Tuple, Union

import pytorch_lightning as pl
import torch
from omegaconf import ListConfig, OmegaConf
from safetensors.torch import load_file as load_safetensors
from torch.optim.lr_scheduler import LambdaLR

from ..modules import UNCONDITIONAL_CONFIG
from ..modules.diffusionmodules.wrappers import OPENAIUNETWRAPPER
from ..modules.ema import LitEma
from ..util import (
    default,
    disabled_train,
    get_obj_from_str,
    instantiate_from_config,
    log_txt_as_img,
)


class DiffusionEngine(pl.LightningModule):
    def __init__(

        self,

        network_config,

        denoiser_config,

        first_stage_config,

        conditioner_config: Union[None, Dict, ListConfig, OmegaConf] = None,

        sampler_config: Union[None, Dict, ListConfig, OmegaConf] = None,

        optimizer_config: Union[None, Dict, ListConfig, OmegaConf] = None,

        scheduler_config: Union[None, Dict, ListConfig, OmegaConf] = None,

        loss_fn_config: Union[None, Dict, ListConfig, OmegaConf] = None,

        network_wrapper: Union[None, str] = None,

        ckpt_path: Union[None, str] = None,

        use_ema: bool = False,

        ema_decay_rate: float = 0.9999,

        scale_factor: float = 1.0,

        disable_first_stage_autocast=False,

        input_key: str = "jpg",

        log_keys: Union[List, None] = None,

        no_cond_log: bool = False,

        compile_model: bool = False,

    ):
        super().__init__()
        self.log_keys = log_keys
        self.input_key = input_key
        self.optimizer_config = default(
            optimizer_config, {"target": "torch.optim.AdamW"}
        )
        model = instantiate_from_config(network_config)
        self.model = get_obj_from_str(default(network_wrapper, OPENAIUNETWRAPPER))(
            model, compile_model=compile_model
        )

        self.denoiser = instantiate_from_config(denoiser_config)
        self.sampler = (
            instantiate_from_config(sampler_config)
            if sampler_config is not None
            else None
        )
        self.conditioner = instantiate_from_config(
            default(conditioner_config, UNCONDITIONAL_CONFIG)
        )
        self.scheduler_config = scheduler_config
        self._init_first_stage(first_stage_config)

        self.loss_fn = (
            instantiate_from_config(loss_fn_config)
            if loss_fn_config is not None
            else None
        )

        self.use_ema = use_ema
        if self.use_ema:
            self.model_ema = LitEma(self.model, decay=ema_decay_rate)
            print(f"Keeping EMAs of {len(list(self.model_ema.buffers()))}.")

        self.scale_factor = scale_factor
        self.disable_first_stage_autocast = disable_first_stage_autocast
        self.no_cond_log = no_cond_log

        if ckpt_path is not None:
            self.init_from_ckpt(ckpt_path)

    def init_from_ckpt(

        self,

        path: str,

    ) -> None:
        if path.endswith("ckpt"):
            sd = torch.load(path, map_location="cpu")["state_dict"]
        elif path.endswith("safetensors"):
            sd = load_safetensors(path)
        else:
            raise NotImplementedError

        missing, unexpected = self.load_state_dict(sd, strict=False)
        print(
            f"Restored from {path} with {len(missing)} missing and {len(unexpected)} unexpected keys"
        )
        if len(missing) > 0:
            print(f"Missing Keys: {missing}")
        if len(unexpected) > 0:
            print(f"Unexpected Keys: {unexpected}")

    def _init_first_stage(self, config):
        model = instantiate_from_config(config).eval()
        model.train = disabled_train
        for param in model.parameters():
            param.requires_grad = False
        self.first_stage_model = model

    def get_input(self, batch):
        # assuming unified data format, dataloader returns a dict.
        # image tensors should be scaled to -1 ... 1 and in bchw format
        return batch[self.input_key]

    @torch.no_grad()
    def decode_first_stage(self, z):
        z = 1.0 / self.scale_factor * z
        with torch.autocast("cuda", enabled=not self.disable_first_stage_autocast):
            out = self.first_stage_model.decode(z)
        return out

    @torch.no_grad()
    def encode_first_stage(self, x):
        with torch.autocast("cuda", enabled=not self.disable_first_stage_autocast):
            z = self.first_stage_model.encode(x)
        z = self.scale_factor * z
        return z

    def forward(self, x, batch):
        loss = self.loss_fn(self.model, self.denoiser, self.conditioner, x, batch)
        loss_mean = loss.mean()
        loss_dict = {"loss": loss_mean}
        return loss_mean, loss_dict

    def shared_step(self, batch: Dict) -> Any:
        x = self.get_input(batch)
        x = self.encode_first_stage(x)
        batch["global_step"] = self.global_step
        loss, loss_dict = self(x, batch)
        return loss, loss_dict

    def training_step(self, batch, batch_idx):
        loss, loss_dict = self.shared_step(batch)

        self.log_dict(
            loss_dict, prog_bar=True, logger=True, on_step=True, on_epoch=False
        )

        self.log(
            "global_step",
            self.global_step,
            prog_bar=True,
            logger=True,
            on_step=True,
            on_epoch=False,
        )

        # if self.scheduler_config is not None:
        lr = self.optimizers().param_groups[0]["lr"]
        self.log(
            "lr_abs", lr, prog_bar=True, logger=True, on_step=True, on_epoch=False
        )

        return loss

    def on_train_start(self, *args, **kwargs):
        if self.sampler is None or self.loss_fn is None:
            raise ValueError("Sampler and loss function need to be set for training.")

    def on_train_batch_end(self, *args, **kwargs):
        if self.use_ema:
            self.model_ema(self.model)

    @contextmanager
    def ema_scope(self, context=None):
        if self.use_ema:
            self.model_ema.store(self.model.parameters())
            self.model_ema.copy_to(self.model)
            if context is not None:
                print(f"{context}: Switched to EMA weights")
        try:
            yield None
        finally:
            if self.use_ema:
                self.model_ema.restore(self.model.parameters())
                if context is not None:
                    print(f"{context}: Restored training weights")

    def instantiate_optimizer_from_config(self, params, lr, cfg):
        return get_obj_from_str(cfg["target"])(
            params, lr=lr, **cfg.get("params", dict())
        )

    def configure_optimizers(self):
        lr = self.learning_rate
        params = list(self.model.parameters())
        for embedder in self.conditioner.embedders:
            if embedder.is_trainable:
                params = params + list(embedder.parameters())
        opt = self.instantiate_optimizer_from_config(params, lr, self.optimizer_config)
        if self.scheduler_config is not None:
            scheduler = instantiate_from_config(self.scheduler_config)
            print("Setting up LambdaLR scheduler...")
            scheduler = [
                {
                    "scheduler": LambdaLR(opt, lr_lambda=scheduler.schedule),
                    "interval": "step",
                    "frequency": 1,
                }
            ]
            return [opt], scheduler
        return opt

    @torch.no_grad()
    def sample(

        self,

        cond: Dict,

        uc: Union[Dict, None] = None,

        batch_size: int = 16,

        shape: Union[None, Tuple, List] = None,

        **kwargs,

    ):
        randn = torch.randn(batch_size, *shape).to(self.device)

        denoiser = lambda input, sigma, c: self.denoiser(
            self.model, input, sigma, c, **kwargs
        )
        samples = self.sampler(denoiser, randn, cond, uc=uc)
        return samples

    @torch.no_grad()
    def log_conditionings(self, batch: Dict, n: int) -> Dict:
        """

        Defines heuristics to log different conditionings.

        These can be lists of strings (text-to-image), tensors, ints, ...

        """
        image_h, image_w = batch[self.input_key].shape[2:]
        log = dict()

        for embedder in self.conditioner.embedders:
            if (
                (self.log_keys is None) or (embedder.input_key in self.log_keys)
            ) and not self.no_cond_log:
                x = batch[embedder.input_key][:n]
                if isinstance(x, torch.Tensor):
                    if x.dim() == 1:
                        # class-conditional, convert integer to string
                        x = [str(x[i].item()) for i in range(x.shape[0])]
                        xc = log_txt_as_img((image_h, image_w), x, size=image_h // 4)
                    elif x.dim() == 2:
                        # size and crop cond and the like
                        x = [
                            "x".join([str(xx) for xx in x[i].tolist()])
                            for i in range(x.shape[0])
                        ]
                        xc = log_txt_as_img((image_h, image_w), x, size=image_h // 20)
                    else:
                        raise NotImplementedError()
                elif isinstance(x, (List, ListConfig)):
                    if isinstance(x[0], str):
                        # strings
                        xc = log_txt_as_img((image_h, image_w), x, size=image_h // 20)
                    else:
                        raise NotImplementedError()
                else:
                    raise NotImplementedError()
                log[embedder.input_key] = xc
        return log

    @torch.no_grad()
    def log_images(

        self,

        batch: Dict,

        N: int = 8,

        sample: bool = True,

        ucg_keys: List[str] = None,

        **kwargs,

    ) -> Dict:
        conditioner_input_keys = [e.input_key for e in self.conditioner.embedders]
        if ucg_keys:
            assert all(map(lambda x: x in conditioner_input_keys, ucg_keys)), (
                "Each defined ucg key for sampling must be in the provided conditioner input keys,"
                f"but we have {ucg_keys} vs. {conditioner_input_keys}"
            )
        else:
            ucg_keys = conditioner_input_keys
        log = dict()

        x = self.get_input(batch)

        c, uc = self.conditioner.get_unconditional_conditioning(
            batch,
            force_uc_zero_embeddings=ucg_keys
            if len(self.conditioner.embedders) > 0
            else [],
        )

        sampling_kwargs = {}

        N = min(x.shape[0], N)
        x = x.to(self.device)[:N]
        log["inputs"] = x
        z = self.encode_first_stage(x)
        log["reconstructions"] = self.decode_first_stage(z)
        log.update(self.log_conditionings(batch, N))

        for k in c:
            if isinstance(c[k], torch.Tensor):
                c[k], uc[k] = map(lambda y: y[k][:N].to(self.device), (c, uc))

        if sample:
            with self.ema_scope("Plotting"):
                samples = self.sample(
                    c, shape=z.shape[1:], uc=uc, batch_size=N, **sampling_kwargs
                )
            samples = self.decode_first_stage(samples)
            log["samples"] = samples
        return log