Fabrice-TIERCELIN commited on
Commit
38a87f5
1 Parent(s): e40d0ec

Upload modules.py

Browse files
Files changed (1) hide show
  1. sgm/modules/encoders/modules.py +1062 -0
sgm/modules/encoders/modules.py ADDED
@@ -0,0 +1,1062 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from contextlib import nullcontext
2
+ from functools import partial
3
+ from typing import Dict, List, Optional, Tuple, Union
4
+
5
+ import kornia
6
+ import numpy as np
7
+ import open_clip
8
+ import torch
9
+ import torch.nn as nn
10
+ from einops import rearrange, repeat
11
+ from omegaconf import ListConfig
12
+ from torch.utils.checkpoint import checkpoint
13
+ from transformers import (
14
+ ByT5Tokenizer,
15
+ CLIPTextModel,
16
+ CLIPTokenizer,
17
+ T5EncoderModel,
18
+ T5Tokenizer,
19
+ )
20
+
21
+ from ...modules.autoencoding.regularizers import DiagonalGaussianRegularizer
22
+ from ...modules.diffusionmodules.model import Encoder
23
+ from ...modules.diffusionmodules.openaimodel import Timestep
24
+ from ...modules.diffusionmodules.util import extract_into_tensor, make_beta_schedule
25
+ from ...modules.distributions.distributions import DiagonalGaussianDistribution
26
+ from ...util import (
27
+ autocast,
28
+ count_params,
29
+ default,
30
+ disabled_train,
31
+ expand_dims_like,
32
+ instantiate_from_config,
33
+ )
34
+
35
+ from CKPT_PTH import SDXL_CLIP1_PATH, SDXL_CLIP2_CKPT_PTH
36
+
37
+ class AbstractEmbModel(nn.Module):
38
+ def __init__(self):
39
+ super().__init__()
40
+ self._is_trainable = None
41
+ self._ucg_rate = None
42
+ self._input_key = None
43
+
44
+ @property
45
+ def is_trainable(self) -> bool:
46
+ return self._is_trainable
47
+
48
+ @property
49
+ def ucg_rate(self) -> Union[float, torch.Tensor]:
50
+ return self._ucg_rate
51
+
52
+ @property
53
+ def input_key(self) -> str:
54
+ return self._input_key
55
+
56
+ @is_trainable.setter
57
+ def is_trainable(self, value: bool):
58
+ self._is_trainable = value
59
+
60
+ @ucg_rate.setter
61
+ def ucg_rate(self, value: Union[float, torch.Tensor]):
62
+ self._ucg_rate = value
63
+
64
+ @input_key.setter
65
+ def input_key(self, value: str):
66
+ self._input_key = value
67
+
68
+ @is_trainable.deleter
69
+ def is_trainable(self):
70
+ del self._is_trainable
71
+
72
+ @ucg_rate.deleter
73
+ def ucg_rate(self):
74
+ del self._ucg_rate
75
+
76
+ @input_key.deleter
77
+ def input_key(self):
78
+ del self._input_key
79
+
80
+
81
+ class GeneralConditioner(nn.Module):
82
+ OUTPUT_DIM2KEYS = {2: "vector", 3: "crossattn", 4: "concat", 5: "concat"}
83
+ KEY2CATDIM = {"vector": 1, "crossattn": 2, "concat": 1, 'control_vector': 1}
84
+
85
+ def __init__(self, emb_models: Union[List, ListConfig]):
86
+ super().__init__()
87
+ embedders = []
88
+ for n, embconfig in enumerate(emb_models):
89
+ embedder = instantiate_from_config(embconfig)
90
+ assert isinstance(
91
+ embedder, AbstractEmbModel
92
+ ), f"embedder model {embedder.__class__.__name__} has to inherit from AbstractEmbModel"
93
+ embedder.is_trainable = embconfig.get("is_trainable", False)
94
+ embedder.ucg_rate = embconfig.get("ucg_rate", 0.0)
95
+ if not embedder.is_trainable:
96
+ embedder.train = disabled_train
97
+ for param in embedder.parameters():
98
+ param.requires_grad = False
99
+ embedder.eval()
100
+ print(
101
+ f"Initialized embedder #{n}: {embedder.__class__.__name__} "
102
+ f"with {count_params(embedder, False)} params. Trainable: {embedder.is_trainable}"
103
+ )
104
+
105
+ if "input_key" in embconfig:
106
+ embedder.input_key = embconfig["input_key"]
107
+ elif "input_keys" in embconfig:
108
+ embedder.input_keys = embconfig["input_keys"]
109
+ else:
110
+ raise KeyError(
111
+ f"need either 'input_key' or 'input_keys' for embedder {embedder.__class__.__name__}"
112
+ )
113
+
114
+ embedder.legacy_ucg_val = embconfig.get("legacy_ucg_value", None)
115
+ if embedder.legacy_ucg_val is not None:
116
+ embedder.ucg_prng = np.random.RandomState()
117
+
118
+ embedders.append(embedder)
119
+ self.embedders = nn.ModuleList(embedders)
120
+
121
+ def possibly_get_ucg_val(self, embedder: AbstractEmbModel, batch: Dict) -> Dict:
122
+ assert embedder.legacy_ucg_val is not None
123
+ p = embedder.ucg_rate
124
+ val = embedder.legacy_ucg_val
125
+ for i in range(len(batch[embedder.input_key])):
126
+ if embedder.ucg_prng.choice(2, p=[1 - p, p]):
127
+ batch[embedder.input_key][i] = val
128
+ return batch
129
+
130
+ def forward(
131
+ self, batch: Dict, force_zero_embeddings: Optional[List] = None
132
+ ) -> Dict:
133
+ output = dict()
134
+ if force_zero_embeddings is None:
135
+ force_zero_embeddings = []
136
+ for embedder in self.embedders:
137
+ embedding_context = nullcontext if embedder.is_trainable else torch.no_grad
138
+ with embedding_context():
139
+ if hasattr(embedder, "input_key") and (embedder.input_key is not None):
140
+ if embedder.legacy_ucg_val is not None:
141
+ batch = self.possibly_get_ucg_val(embedder, batch)
142
+ emb_out = embedder(batch[embedder.input_key])
143
+ elif hasattr(embedder, "input_keys"):
144
+ emb_out = embedder(*[batch[k] for k in embedder.input_keys])
145
+ assert isinstance(
146
+ emb_out, (torch.Tensor, list, tuple)
147
+ ), f"encoder outputs must be tensors or a sequence, but got {type(emb_out)}"
148
+ if not isinstance(emb_out, (list, tuple)):
149
+ emb_out = [emb_out]
150
+ for emb in emb_out:
151
+ out_key = self.OUTPUT_DIM2KEYS[emb.dim()]
152
+ if embedder.ucg_rate > 0.0 and embedder.legacy_ucg_val is None:
153
+ emb = (
154
+ expand_dims_like(
155
+ torch.bernoulli(
156
+ (1.0 - embedder.ucg_rate)
157
+ * torch.ones(emb.shape[0], device=emb.device)
158
+ ),
159
+ emb,
160
+ )
161
+ * emb
162
+ )
163
+ if (
164
+ hasattr(embedder, "input_key")
165
+ and embedder.input_key in force_zero_embeddings
166
+ ):
167
+ emb = torch.zeros_like(emb)
168
+ if out_key in output:
169
+ output[out_key] = torch.cat(
170
+ (output[out_key], emb), self.KEY2CATDIM[out_key]
171
+ )
172
+ else:
173
+ output[out_key] = emb
174
+ return output
175
+
176
+ def get_unconditional_conditioning(
177
+ self, batch_c, batch_uc=None, force_uc_zero_embeddings=None
178
+ ):
179
+ if force_uc_zero_embeddings is None:
180
+ force_uc_zero_embeddings = []
181
+ ucg_rates = list()
182
+ for embedder in self.embedders:
183
+ ucg_rates.append(embedder.ucg_rate)
184
+ embedder.ucg_rate = 0.0
185
+ c = self(batch_c)
186
+ uc = self(batch_c if batch_uc is None else batch_uc, force_uc_zero_embeddings)
187
+
188
+ for embedder, rate in zip(self.embedders, ucg_rates):
189
+ embedder.ucg_rate = rate
190
+ return c, uc
191
+
192
+
193
+ class GeneralConditionerWithControl(GeneralConditioner):
194
+ def forward(
195
+ self, batch: Dict, force_zero_embeddings: Optional[List] = None
196
+ ) -> Dict:
197
+ output = dict()
198
+ if force_zero_embeddings is None:
199
+ force_zero_embeddings = []
200
+ for embedder in self.embedders:
201
+ embedding_context = nullcontext if embedder.is_trainable else torch.no_grad
202
+ with embedding_context():
203
+ if hasattr(embedder, "input_key") and (embedder.input_key is not None):
204
+ if embedder.legacy_ucg_val is not None:
205
+ batch = self.possibly_get_ucg_val(embedder, batch)
206
+ emb_out = embedder(batch[embedder.input_key])
207
+ elif hasattr(embedder, "input_keys"):
208
+ emb_out = embedder(*[batch[k] for k in embedder.input_keys])
209
+ assert isinstance(
210
+ emb_out, (torch.Tensor, list, tuple)
211
+ ), f"encoder outputs must be tensors or a sequence, but got {type(emb_out)}"
212
+ if not isinstance(emb_out, (list, tuple)):
213
+ emb_out = [emb_out]
214
+ for emb in emb_out:
215
+ if 'control_vector' in embedder.input_key:
216
+ out_key = 'control_vector'
217
+ else:
218
+ out_key = self.OUTPUT_DIM2KEYS[emb.dim()]
219
+ if embedder.ucg_rate > 0.0 and embedder.legacy_ucg_val is None:
220
+ emb = (
221
+ expand_dims_like(
222
+ torch.bernoulli(
223
+ (1.0 - embedder.ucg_rate)
224
+ * torch.ones(emb.shape[0], device=emb.device)
225
+ ),
226
+ emb,
227
+ )
228
+ * emb
229
+ )
230
+ if (
231
+ hasattr(embedder, "input_key")
232
+ and embedder.input_key in force_zero_embeddings
233
+ ):
234
+ emb = torch.zeros_like(emb)
235
+ if out_key in output:
236
+ output[out_key] = torch.cat(
237
+ (output[out_key], emb), self.KEY2CATDIM[out_key]
238
+ )
239
+ else:
240
+ output[out_key] = emb
241
+
242
+ output["control"] = batch["control"]
243
+ return output
244
+
245
+
246
+ class PreparedConditioner(nn.Module):
247
+ def __init__(self, cond_pth, un_cond_pth=None):
248
+ super().__init__()
249
+ conditions = torch.load(cond_pth)
250
+ for k, v in conditions.items():
251
+ self.register_buffer(k, v)
252
+ self.un_cond_pth = un_cond_pth
253
+ if un_cond_pth is not None:
254
+ un_conditions = torch.load(un_cond_pth)
255
+ for k, v in un_conditions.items():
256
+ self.register_buffer(k+'_uc', v)
257
+
258
+
259
+ @torch.no_grad()
260
+ def forward(
261
+ self, batch: Dict, return_uc=False
262
+ ) -> Dict:
263
+ output = dict()
264
+ for k, v in self.state_dict().items():
265
+ if not return_uc:
266
+ if k.endswith("_uc"):
267
+ continue
268
+ else:
269
+ output[k] = v.detach().clone().repeat(batch['control'].shape[0], *[1 for _ in range(v.ndim - 1)])
270
+ else:
271
+ if k.endswith("_uc"):
272
+ output[k[:-3]] = v.detach().clone().repeat(batch['control'].shape[0], *[1 for _ in range(v.ndim - 1)])
273
+ else:
274
+ continue
275
+ output["control"] = batch["control"]
276
+
277
+ for k, v in output.items():
278
+ if isinstance(v, torch.Tensor):
279
+ assert (torch.isnan(v).any()) is not None
280
+ return output
281
+
282
+ def get_unconditional_conditioning(
283
+ self, batch_c, batch_uc=None, force_uc_zero_embeddings=None
284
+ ):
285
+ c = self(batch_c)
286
+ if self.un_cond_pth is not None:
287
+ uc = self(batch_c, return_uc=True)
288
+ else:
289
+ uc = None
290
+ return c, uc
291
+
292
+
293
+
294
+ class InceptionV3(nn.Module):
295
+ """Wrapper around the https://github.com/mseitzer/pytorch-fid inception
296
+ port with an additional squeeze at the end"""
297
+
298
+ def __init__(self, normalize_input=False, **kwargs):
299
+ super().__init__()
300
+ from pytorch_fid import inception
301
+
302
+ kwargs["resize_input"] = True
303
+ self.model = inception.InceptionV3(normalize_input=normalize_input, **kwargs)
304
+
305
+ def forward(self, inp):
306
+ # inp = kornia.geometry.resize(inp, (299, 299),
307
+ # interpolation='bicubic',
308
+ # align_corners=False,
309
+ # antialias=True)
310
+ # inp = inp.clamp(min=-1, max=1)
311
+
312
+ outp = self.model(inp)
313
+
314
+ if len(outp) == 1:
315
+ return outp[0].squeeze()
316
+
317
+ return outp
318
+
319
+
320
+ class IdentityEncoder(AbstractEmbModel):
321
+ def encode(self, x):
322
+ return x
323
+
324
+ def forward(self, x):
325
+ return x
326
+
327
+
328
+ class ClassEmbedder(AbstractEmbModel):
329
+ def __init__(self, embed_dim, n_classes=1000, add_sequence_dim=False):
330
+ super().__init__()
331
+ self.embedding = nn.Embedding(n_classes, embed_dim)
332
+ self.n_classes = n_classes
333
+ self.add_sequence_dim = add_sequence_dim
334
+
335
+ def forward(self, c):
336
+ c = self.embedding(c)
337
+ if self.add_sequence_dim:
338
+ c = c[:, None, :]
339
+ return c
340
+
341
+ def get_unconditional_conditioning(self, bs, device="cuda"):
342
+ uc_class = (
343
+ self.n_classes - 1
344
+ ) # 1000 classes --> 0 ... 999, one extra class for ucg (class 1000)
345
+ uc = torch.ones((bs,), device=device) * uc_class
346
+ uc = {self.key: uc.long()}
347
+ return uc
348
+
349
+
350
+ class ClassEmbedderForMultiCond(ClassEmbedder):
351
+ def forward(self, batch, key=None, disable_dropout=False):
352
+ out = batch
353
+ key = default(key, self.key)
354
+ islist = isinstance(batch[key], list)
355
+ if islist:
356
+ batch[key] = batch[key][0]
357
+ c_out = super().forward(batch, key, disable_dropout)
358
+ out[key] = [c_out] if islist else c_out
359
+ return out
360
+
361
+
362
+ class FrozenT5Embedder(AbstractEmbModel):
363
+ """Uses the T5 transformer encoder for text"""
364
+
365
+ def __init__(
366
+ self, version="google/t5-v1_1-xxl", device="cuda", max_length=77, freeze=True
367
+ ): # others are google/t5-v1_1-xl and google/t5-v1_1-xxl
368
+ super().__init__()
369
+ self.tokenizer = T5Tokenizer.from_pretrained(version)
370
+ self.transformer = T5EncoderModel.from_pretrained(version)
371
+ self.device = device
372
+ self.max_length = max_length
373
+ if freeze:
374
+ self.freeze()
375
+
376
+ def freeze(self):
377
+ self.transformer = self.transformer.eval()
378
+
379
+ for param in self.parameters():
380
+ param.requires_grad = False
381
+
382
+ # @autocast
383
+ def forward(self, text):
384
+ batch_encoding = self.tokenizer(
385
+ text,
386
+ truncation=True,
387
+ max_length=self.max_length,
388
+ return_length=True,
389
+ return_overflowing_tokens=False,
390
+ padding="max_length",
391
+ return_tensors="pt",
392
+ )
393
+ tokens = batch_encoding["input_ids"].to(self.device)
394
+ with torch.autocast("cuda", enabled=False):
395
+ outputs = self.transformer(input_ids=tokens)
396
+ z = outputs.last_hidden_state
397
+ return z
398
+
399
+ def encode(self, text):
400
+ return self(text)
401
+
402
+
403
+ class FrozenByT5Embedder(AbstractEmbModel):
404
+ """
405
+ Uses the ByT5 transformer encoder for text. Is character-aware.
406
+ """
407
+
408
+ def __init__(
409
+ self, version="google/byt5-base", device="cuda", max_length=77, freeze=True
410
+ ): # others are google/t5-v1_1-xl and google/t5-v1_1-xxl
411
+ super().__init__()
412
+ self.tokenizer = ByT5Tokenizer.from_pretrained(version)
413
+ self.transformer = T5EncoderModel.from_pretrained(version)
414
+ self.device = device
415
+ self.max_length = max_length
416
+ if freeze:
417
+ self.freeze()
418
+
419
+ def freeze(self):
420
+ self.transformer = self.transformer.eval()
421
+
422
+ for param in self.parameters():
423
+ param.requires_grad = False
424
+
425
+ def forward(self, text):
426
+ batch_encoding = self.tokenizer(
427
+ text,
428
+ truncation=True,
429
+ max_length=self.max_length,
430
+ return_length=True,
431
+ return_overflowing_tokens=False,
432
+ padding="max_length",
433
+ return_tensors="pt",
434
+ )
435
+ tokens = batch_encoding["input_ids"].to(self.device)
436
+ with torch.autocast("cuda", enabled=False):
437
+ outputs = self.transformer(input_ids=tokens)
438
+ z = outputs.last_hidden_state
439
+ return z
440
+
441
+ def encode(self, text):
442
+ return self(text)
443
+
444
+
445
+ class FrozenCLIPEmbedder(AbstractEmbModel):
446
+ """Uses the CLIP transformer encoder for text (from huggingface)"""
447
+
448
+ LAYERS = ["last", "pooled", "hidden"]
449
+
450
+ def __init__(
451
+ self,
452
+ version="openai/clip-vit-large-patch14",
453
+ device="cuda",
454
+ max_length=77,
455
+ freeze=True,
456
+ layer="last",
457
+ layer_idx=None,
458
+ always_return_pooled=False,
459
+ ): # clip-vit-base-patch32
460
+ super().__init__()
461
+ assert layer in self.LAYERS
462
+ self.tokenizer = CLIPTokenizer.from_pretrained(version if SDXL_CLIP1_PATH is None else SDXL_CLIP1_PATH)
463
+ self.transformer = CLIPTextModel.from_pretrained(version if SDXL_CLIP1_PATH is None else SDXL_CLIP1_PATH)
464
+ self.device = device
465
+ self.max_length = max_length
466
+ if freeze:
467
+ self.freeze()
468
+ self.layer = layer
469
+ self.layer_idx = layer_idx
470
+ self.return_pooled = always_return_pooled
471
+ if layer == "hidden":
472
+ assert layer_idx is not None
473
+ assert 0 <= abs(layer_idx) <= 12
474
+
475
+ def freeze(self):
476
+ self.transformer = self.transformer.eval()
477
+
478
+ for param in self.parameters():
479
+ param.requires_grad = False
480
+
481
+ @autocast
482
+ def forward(self, text):
483
+ batch_encoding = self.tokenizer(
484
+ text,
485
+ truncation=True,
486
+ max_length=self.max_length,
487
+ return_length=True,
488
+ return_overflowing_tokens=False,
489
+ padding="max_length",
490
+ return_tensors="pt",
491
+ )
492
+ tokens = batch_encoding["input_ids"].to(self.device)
493
+ outputs = self.transformer(
494
+ input_ids=tokens, output_hidden_states=self.layer == "hidden"
495
+ )
496
+ if self.layer == "last":
497
+ z = outputs.last_hidden_state
498
+ elif self.layer == "pooled":
499
+ z = outputs.pooler_output[:, None, :]
500
+ else:
501
+ z = outputs.hidden_states[self.layer_idx]
502
+ if self.return_pooled:
503
+ return z, outputs.pooler_output
504
+ return z
505
+
506
+ def encode(self, text):
507
+ return self(text)
508
+
509
+
510
+ class FrozenOpenCLIPEmbedder2(AbstractEmbModel):
511
+ """
512
+ Uses the OpenCLIP transformer encoder for text
513
+ """
514
+
515
+ LAYERS = ["pooled", "last", "penultimate"]
516
+
517
+ def __init__(
518
+ self,
519
+ arch="ViT-H-14",
520
+ version="laion2b_s32b_b79k",
521
+ device="cuda",
522
+ max_length=77,
523
+ freeze=True,
524
+ layer="last",
525
+ always_return_pooled=False,
526
+ legacy=True,
527
+ ):
528
+ super().__init__()
529
+ assert layer in self.LAYERS
530
+ model, _, _ = open_clip.create_model_and_transforms(
531
+ arch,
532
+ device=torch.device("cpu"),
533
+ pretrained=version if SDXL_CLIP2_CKPT_PTH is None else SDXL_CLIP2_CKPT_PTH,
534
+ )
535
+ del model.visual
536
+ self.model = model
537
+
538
+ self.device = device
539
+ self.max_length = max_length
540
+ self.return_pooled = always_return_pooled
541
+ if freeze:
542
+ self.freeze()
543
+ self.layer = layer
544
+ if self.layer == "last":
545
+ self.layer_idx = 0
546
+ elif self.layer == "penultimate":
547
+ self.layer_idx = 1
548
+ else:
549
+ raise NotImplementedError()
550
+ self.legacy = legacy
551
+
552
+ def freeze(self):
553
+ self.model = self.model.eval()
554
+ for param in self.parameters():
555
+ param.requires_grad = False
556
+
557
+ @autocast
558
+ def forward(self, text):
559
+ tokens = open_clip.tokenize(text)
560
+ z = self.encode_with_transformer(tokens.to(self.device))
561
+ if not self.return_pooled and self.legacy:
562
+ return z
563
+ if self.return_pooled:
564
+ assert not self.legacy
565
+ return z[self.layer], z["pooled"]
566
+ return z[self.layer]
567
+
568
+ def encode_with_transformer(self, text):
569
+ x = self.model.token_embedding(text) # [batch_size, n_ctx, d_model]
570
+ x = x + self.model.positional_embedding
571
+ x = x.permute(1, 0, 2) # NLD -> LND
572
+ x = self.text_transformer_forward(x, attn_mask=self.model.attn_mask)
573
+ if self.legacy:
574
+ x = x[self.layer]
575
+ x = self.model.ln_final(x)
576
+ return x
577
+ else:
578
+ # x is a dict and will stay a dict
579
+ o = x["last"]
580
+ o = self.model.ln_final(o)
581
+ pooled = self.pool(o, text)
582
+ x["pooled"] = pooled
583
+ return x
584
+
585
+ def pool(self, x, text):
586
+ # take features from the eot embedding (eot_token is the highest number in each sequence)
587
+ x = (
588
+ x[torch.arange(x.shape[0]), text.argmax(dim=-1)]
589
+ @ self.model.text_projection
590
+ )
591
+ return x
592
+
593
+ def text_transformer_forward(self, x: torch.Tensor, attn_mask=None):
594
+ outputs = {}
595
+ for i, r in enumerate(self.model.transformer.resblocks):
596
+ if i == len(self.model.transformer.resblocks) - 1:
597
+ outputs["penultimate"] = x.permute(1, 0, 2) # LND -> NLD
598
+ if (
599
+ self.model.transformer.grad_checkpointing
600
+ and not torch.jit.is_scripting()
601
+ ):
602
+ x = checkpoint(r, x, attn_mask)
603
+ else:
604
+ x = r(x, attn_mask=attn_mask)
605
+ outputs["last"] = x.permute(1, 0, 2) # LND -> NLD
606
+ return outputs
607
+
608
+ def encode(self, text):
609
+ return self(text)
610
+
611
+
612
+ class FrozenOpenCLIPEmbedder(AbstractEmbModel):
613
+ LAYERS = [
614
+ # "pooled",
615
+ "last",
616
+ "penultimate",
617
+ ]
618
+
619
+ def __init__(
620
+ self,
621
+ arch="ViT-H-14",
622
+ version="laion2b_s32b_b79k",
623
+ device="cuda",
624
+ max_length=77,
625
+ freeze=True,
626
+ layer="last",
627
+ ):
628
+ super().__init__()
629
+ assert layer in self.LAYERS
630
+ model, _, _ = open_clip.create_model_and_transforms(
631
+ arch, device=torch.device("cpu"), pretrained=version
632
+ )
633
+ del model.visual
634
+ self.model = model
635
+
636
+ self.device = device
637
+ self.max_length = max_length
638
+ if freeze:
639
+ self.freeze()
640
+ self.layer = layer
641
+ if self.layer == "last":
642
+ self.layer_idx = 0
643
+ elif self.layer == "penultimate":
644
+ self.layer_idx = 1
645
+ else:
646
+ raise NotImplementedError()
647
+
648
+ def freeze(self):
649
+ self.model = self.model.eval()
650
+ for param in self.parameters():
651
+ param.requires_grad = False
652
+
653
+ def forward(self, text):
654
+ tokens = open_clip.tokenize(text)
655
+ z = self.encode_with_transformer(tokens.to(self.device))
656
+ return z
657
+
658
+ def encode_with_transformer(self, text):
659
+ x = self.model.token_embedding(text) # [batch_size, n_ctx, d_model]
660
+ x = x + self.model.positional_embedding
661
+ x = x.permute(1, 0, 2) # NLD -> LND
662
+ x = self.text_transformer_forward(x, attn_mask=self.model.attn_mask)
663
+ x = x.permute(1, 0, 2) # LND -> NLD
664
+ x = self.model.ln_final(x)
665
+ return x
666
+
667
+ def text_transformer_forward(self, x: torch.Tensor, attn_mask=None):
668
+ for i, r in enumerate(self.model.transformer.resblocks):
669
+ if i == len(self.model.transformer.resblocks) - self.layer_idx:
670
+ break
671
+ if (
672
+ self.model.transformer.grad_checkpointing
673
+ and not torch.jit.is_scripting()
674
+ ):
675
+ x = checkpoint(r, x, attn_mask)
676
+ else:
677
+ x = r(x, attn_mask=attn_mask)
678
+ return x
679
+
680
+ def encode(self, text):
681
+ return self(text)
682
+
683
+
684
+ class FrozenOpenCLIPImageEmbedder(AbstractEmbModel):
685
+ """
686
+ Uses the OpenCLIP vision transformer encoder for images
687
+ """
688
+
689
+ def __init__(
690
+ self,
691
+ arch="ViT-H-14",
692
+ version="laion2b_s32b_b79k",
693
+ device="cuda",
694
+ max_length=77,
695
+ freeze=True,
696
+ antialias=True,
697
+ ucg_rate=0.0,
698
+ unsqueeze_dim=False,
699
+ repeat_to_max_len=False,
700
+ num_image_crops=0,
701
+ output_tokens=False,
702
+ ):
703
+ super().__init__()
704
+ model, _, _ = open_clip.create_model_and_transforms(
705
+ arch,
706
+ device=torch.device("cpu"),
707
+ pretrained=version,
708
+ )
709
+ del model.transformer
710
+ self.model = model
711
+ self.max_crops = num_image_crops
712
+ self.pad_to_max_len = self.max_crops > 0
713
+ self.repeat_to_max_len = repeat_to_max_len and (not self.pad_to_max_len)
714
+ self.device = device
715
+ self.max_length = max_length
716
+ if freeze:
717
+ self.freeze()
718
+
719
+ self.antialias = antialias
720
+
721
+ self.register_buffer(
722
+ "mean", torch.Tensor([0.48145466, 0.4578275, 0.40821073]), persistent=False
723
+ )
724
+ self.register_buffer(
725
+ "std", torch.Tensor([0.26862954, 0.26130258, 0.27577711]), persistent=False
726
+ )
727
+ self.ucg_rate = ucg_rate
728
+ self.unsqueeze_dim = unsqueeze_dim
729
+ self.stored_batch = None
730
+ self.model.visual.output_tokens = output_tokens
731
+ self.output_tokens = output_tokens
732
+
733
+ def preprocess(self, x):
734
+ # normalize to [0,1]
735
+ x = kornia.geometry.resize(
736
+ x,
737
+ (224, 224),
738
+ interpolation="bicubic",
739
+ align_corners=True,
740
+ antialias=self.antialias,
741
+ )
742
+ x = (x + 1.0) / 2.0
743
+ # renormalize according to clip
744
+ x = kornia.enhance.normalize(x, self.mean, self.std)
745
+ return x
746
+
747
+ def freeze(self):
748
+ self.model = self.model.eval()
749
+ for param in self.parameters():
750
+ param.requires_grad = False
751
+
752
+ @autocast
753
+ def forward(self, image, no_dropout=False):
754
+ z = self.encode_with_vision_transformer(image)
755
+ tokens = None
756
+ if self.output_tokens:
757
+ z, tokens = z[0], z[1]
758
+ z = z.to(image.dtype)
759
+ if self.ucg_rate > 0.0 and not no_dropout and not (self.max_crops > 0):
760
+ z = (
761
+ torch.bernoulli(
762
+ (1.0 - self.ucg_rate) * torch.ones(z.shape[0], device=z.device)
763
+ )[:, None]
764
+ * z
765
+ )
766
+ if tokens is not None:
767
+ tokens = (
768
+ expand_dims_like(
769
+ torch.bernoulli(
770
+ (1.0 - self.ucg_rate)
771
+ * torch.ones(tokens.shape[0], device=tokens.device)
772
+ ),
773
+ tokens,
774
+ )
775
+ * tokens
776
+ )
777
+ if self.unsqueeze_dim:
778
+ z = z[:, None, :]
779
+ if self.output_tokens:
780
+ assert not self.repeat_to_max_len
781
+ assert not self.pad_to_max_len
782
+ return tokens, z
783
+ if self.repeat_to_max_len:
784
+ if z.dim() == 2:
785
+ z_ = z[:, None, :]
786
+ else:
787
+ z_ = z
788
+ return repeat(z_, "b 1 d -> b n d", n=self.max_length), z
789
+ elif self.pad_to_max_len:
790
+ assert z.dim() == 3
791
+ z_pad = torch.cat(
792
+ (
793
+ z,
794
+ torch.zeros(
795
+ z.shape[0],
796
+ self.max_length - z.shape[1],
797
+ z.shape[2],
798
+ device=z.device,
799
+ ),
800
+ ),
801
+ 1,
802
+ )
803
+ return z_pad, z_pad[:, 0, ...]
804
+ return z
805
+
806
+ def encode_with_vision_transformer(self, img):
807
+ # if self.max_crops > 0:
808
+ # img = self.preprocess_by_cropping(img)
809
+ if img.dim() == 5:
810
+ assert self.max_crops == img.shape[1]
811
+ img = rearrange(img, "b n c h w -> (b n) c h w")
812
+ img = self.preprocess(img)
813
+ if not self.output_tokens:
814
+ assert not self.model.visual.output_tokens
815
+ x = self.model.visual(img)
816
+ tokens = None
817
+ else:
818
+ assert self.model.visual.output_tokens
819
+ x, tokens = self.model.visual(img)
820
+ if self.max_crops > 0:
821
+ x = rearrange(x, "(b n) d -> b n d", n=self.max_crops)
822
+ # drop out between 0 and all along the sequence axis
823
+ x = (
824
+ torch.bernoulli(
825
+ (1.0 - self.ucg_rate)
826
+ * torch.ones(x.shape[0], x.shape[1], 1, device=x.device)
827
+ )
828
+ * x
829
+ )
830
+ if tokens is not None:
831
+ tokens = rearrange(tokens, "(b n) t d -> b t (n d)", n=self.max_crops)
832
+ print(
833
+ f"You are running very experimental token-concat in {self.__class__.__name__}. "
834
+ f"Check what you are doing, and then remove this message."
835
+ )
836
+ if self.output_tokens:
837
+ return x, tokens
838
+ return x
839
+
840
+ def encode(self, text):
841
+ return self(text)
842
+
843
+
844
+ class FrozenCLIPT5Encoder(AbstractEmbModel):
845
+ def __init__(
846
+ self,
847
+ clip_version="openai/clip-vit-large-patch14",
848
+ t5_version="google/t5-v1_1-xl",
849
+ device="cuda",
850
+ clip_max_length=77,
851
+ t5_max_length=77,
852
+ ):
853
+ super().__init__()
854
+ self.clip_encoder = FrozenCLIPEmbedder(
855
+ clip_version, device, max_length=clip_max_length
856
+ )
857
+ self.t5_encoder = FrozenT5Embedder(t5_version, device, max_length=t5_max_length)
858
+ print(
859
+ f"{self.clip_encoder.__class__.__name__} has {count_params(self.clip_encoder) * 1.e-6:.2f} M parameters, "
860
+ f"{self.t5_encoder.__class__.__name__} comes with {count_params(self.t5_encoder) * 1.e-6:.2f} M params."
861
+ )
862
+
863
+ def encode(self, text):
864
+ return self(text)
865
+
866
+ def forward(self, text):
867
+ clip_z = self.clip_encoder.encode(text)
868
+ t5_z = self.t5_encoder.encode(text)
869
+ return [clip_z, t5_z]
870
+
871
+
872
+ class SpatialRescaler(nn.Module):
873
+ def __init__(
874
+ self,
875
+ n_stages=1,
876
+ method="bilinear",
877
+ multiplier=0.5,
878
+ in_channels=3,
879
+ out_channels=None,
880
+ bias=False,
881
+ wrap_video=False,
882
+ kernel_size=1,
883
+ remap_output=False,
884
+ ):
885
+ super().__init__()
886
+ self.n_stages = n_stages
887
+ assert self.n_stages >= 0
888
+ assert method in [
889
+ "nearest",
890
+ "linear",
891
+ "bilinear",
892
+ "trilinear",
893
+ "bicubic",
894
+ "area",
895
+ ]
896
+ self.multiplier = multiplier
897
+ self.interpolator = partial(torch.nn.functional.interpolate, mode=method)
898
+ self.remap_output = out_channels is not None or remap_output
899
+ if self.remap_output:
900
+ print(
901
+ f"Spatial Rescaler mapping from {in_channels} to {out_channels} channels after resizing."
902
+ )
903
+ self.channel_mapper = nn.Conv2d(
904
+ in_channels,
905
+ out_channels,
906
+ kernel_size=kernel_size,
907
+ bias=bias,
908
+ padding=kernel_size // 2,
909
+ )
910
+ self.wrap_video = wrap_video
911
+
912
+ def forward(self, x):
913
+ if self.wrap_video and x.ndim == 5:
914
+ B, C, T, H, W = x.shape
915
+ x = rearrange(x, "b c t h w -> b t c h w")
916
+ x = rearrange(x, "b t c h w -> (b t) c h w")
917
+
918
+ for stage in range(self.n_stages):
919
+ x = self.interpolator(x, scale_factor=self.multiplier)
920
+
921
+ if self.wrap_video:
922
+ x = rearrange(x, "(b t) c h w -> b t c h w", b=B, t=T, c=C)
923
+ x = rearrange(x, "b t c h w -> b c t h w")
924
+ if self.remap_output:
925
+ x = self.channel_mapper(x)
926
+ return x
927
+
928
+ def encode(self, x):
929
+ return self(x)
930
+
931
+
932
+ class LowScaleEncoder(nn.Module):
933
+ def __init__(
934
+ self,
935
+ model_config,
936
+ linear_start,
937
+ linear_end,
938
+ timesteps=1000,
939
+ max_noise_level=250,
940
+ output_size=64,
941
+ scale_factor=1.0,
942
+ ):
943
+ super().__init__()
944
+ self.max_noise_level = max_noise_level
945
+ self.model = instantiate_from_config(model_config)
946
+ self.augmentation_schedule = self.register_schedule(
947
+ timesteps=timesteps, linear_start=linear_start, linear_end=linear_end
948
+ )
949
+ self.out_size = output_size
950
+ self.scale_factor = scale_factor
951
+
952
+ def register_schedule(
953
+ self,
954
+ beta_schedule="linear",
955
+ timesteps=1000,
956
+ linear_start=1e-4,
957
+ linear_end=2e-2,
958
+ cosine_s=8e-3,
959
+ ):
960
+ betas = make_beta_schedule(
961
+ beta_schedule,
962
+ timesteps,
963
+ linear_start=linear_start,
964
+ linear_end=linear_end,
965
+ cosine_s=cosine_s,
966
+ )
967
+ alphas = 1.0 - betas
968
+ alphas_cumprod = np.cumprod(alphas, axis=0)
969
+ alphas_cumprod_prev = np.append(1.0, alphas_cumprod[:-1])
970
+
971
+ (timesteps,) = betas.shape
972
+ self.num_timesteps = int(timesteps)
973
+ self.linear_start = linear_start
974
+ self.linear_end = linear_end
975
+ assert (
976
+ alphas_cumprod.shape[0] == self.num_timesteps
977
+ ), "alphas have to be defined for each timestep"
978
+
979
+ to_torch = partial(torch.tensor, dtype=torch.float32)
980
+
981
+ self.register_buffer("betas", to_torch(betas))
982
+ self.register_buffer("alphas_cumprod", to_torch(alphas_cumprod))
983
+ self.register_buffer("alphas_cumprod_prev", to_torch(alphas_cumprod_prev))
984
+
985
+ # calculations for diffusion q(x_t | x_{t-1}) and others
986
+ self.register_buffer("sqrt_alphas_cumprod", to_torch(np.sqrt(alphas_cumprod)))
987
+ self.register_buffer(
988
+ "sqrt_one_minus_alphas_cumprod", to_torch(np.sqrt(1.0 - alphas_cumprod))
989
+ )
990
+ self.register_buffer(
991
+ "log_one_minus_alphas_cumprod", to_torch(np.log(1.0 - alphas_cumprod))
992
+ )
993
+ self.register_buffer(
994
+ "sqrt_recip_alphas_cumprod", to_torch(np.sqrt(1.0 / alphas_cumprod))
995
+ )
996
+ self.register_buffer(
997
+ "sqrt_recipm1_alphas_cumprod", to_torch(np.sqrt(1.0 / alphas_cumprod - 1))
998
+ )
999
+
1000
+ def q_sample(self, x_start, t, noise=None):
1001
+ noise = default(noise, lambda: torch.randn_like(x_start))
1002
+ return (
1003
+ extract_into_tensor(self.sqrt_alphas_cumprod, t, x_start.shape) * x_start
1004
+ + extract_into_tensor(self.sqrt_one_minus_alphas_cumprod, t, x_start.shape)
1005
+ * noise
1006
+ )
1007
+
1008
+ def forward(self, x):
1009
+ z = self.model.encode(x)
1010
+ if isinstance(z, DiagonalGaussianDistribution):
1011
+ z = z.sample()
1012
+ z = z * self.scale_factor
1013
+ noise_level = torch.randint(
1014
+ 0, self.max_noise_level, (x.shape[0],), device=x.device
1015
+ ).long()
1016
+ z = self.q_sample(z, noise_level)
1017
+ if self.out_size is not None:
1018
+ z = torch.nn.functional.interpolate(z, size=self.out_size, mode="nearest")
1019
+ # z = z.repeat_interleave(2, -2).repeat_interleave(2, -1)
1020
+ return z, noise_level
1021
+
1022
+ def decode(self, z):
1023
+ z = z / self.scale_factor
1024
+ return self.model.decode(z)
1025
+
1026
+
1027
+ class ConcatTimestepEmbedderND(AbstractEmbModel):
1028
+ """embeds each dimension independently and concatenates them"""
1029
+
1030
+ def __init__(self, outdim):
1031
+ super().__init__()
1032
+ self.timestep = Timestep(outdim)
1033
+ self.outdim = outdim
1034
+
1035
+ def forward(self, x):
1036
+ if x.ndim == 1:
1037
+ x = x[:, None]
1038
+ assert len(x.shape) == 2
1039
+ b, dims = x.shape[0], x.shape[1]
1040
+ x = rearrange(x, "b d -> (b d)")
1041
+ emb = self.timestep(x)
1042
+ emb = rearrange(emb, "(b d) d2 -> b (d d2)", b=b, d=dims, d2=self.outdim)
1043
+ return emb
1044
+
1045
+
1046
+ class GaussianEncoder(Encoder, AbstractEmbModel):
1047
+ def __init__(
1048
+ self, weight: float = 1.0, flatten_output: bool = True, *args, **kwargs
1049
+ ):
1050
+ super().__init__(*args, **kwargs)
1051
+ self.posterior = DiagonalGaussianRegularizer()
1052
+ self.weight = weight
1053
+ self.flatten_output = flatten_output
1054
+
1055
+ def forward(self, x) -> Tuple[Dict, torch.Tensor]:
1056
+ z = super().forward(x)
1057
+ z, log = self.posterior(z)
1058
+ log["loss"] = log["kl_loss"]
1059
+ log["weight"] = self.weight
1060
+ if self.flatten_output:
1061
+ z = rearrange(z, "b c h w -> b (h w ) c")
1062
+ return log, z