File size: 14,630 Bytes
2a50f45
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
import torch
import torch.nn as nn
from torch.utils.checkpoint import checkpoint
import kornia
import open_clip
from transformers import T5Tokenizer, T5EncoderModel, CLIPTokenizer, CLIPTextModel
from lvdm.common import autocast
from utils.utils import count_params

class AbstractEncoder(nn.Module):
    def __init__(self):
        super().__init__()

    def encode(self, *args, **kwargs):
        raise NotImplementedError


class IdentityEncoder(AbstractEncoder):

    def encode(self, x):
        return x


class ClassEmbedder(nn.Module):
    def __init__(self, embed_dim, n_classes=1000, key='class', ucg_rate=0.1):
        super().__init__()
        self.key = key
        self.embedding = nn.Embedding(n_classes, embed_dim)
        self.n_classes = n_classes
        self.ucg_rate = ucg_rate

    def forward(self, batch, key=None, disable_dropout=False):
        if key is None:
            key = self.key
        # this is for use in crossattn
        c = batch[key][:, None]
        if self.ucg_rate > 0. and not disable_dropout:
            mask = 1. - torch.bernoulli(torch.ones_like(c) * self.ucg_rate)
            c = mask * c + (1 - mask) * torch.ones_like(c) * (self.n_classes - 1)
            c = c.long()
        c = self.embedding(c)
        return c

    def get_unconditional_conditioning(self, bs, device="cuda"):
        uc_class = self.n_classes - 1  # 1000 classes --> 0 ... 999, one extra class for ucg (class 1000)
        uc = torch.ones((bs,), device=device) * uc_class
        uc = {self.key: uc}
        return uc


def disabled_train(self, mode=True):
    """Overwrite model.train with this function to make sure train/eval mode
    does not change anymore."""
    return self


class FrozenT5Embedder(AbstractEncoder):
    """Uses the T5 transformer encoder for text"""

    def __init__(self, version="google/t5-v1_1-large", device="cuda", max_length=77,
                 freeze=True):  # others are google/t5-v1_1-xl and google/t5-v1_1-xxl
        super().__init__()
        self.tokenizer = T5Tokenizer.from_pretrained(version)
        self.transformer = T5EncoderModel.from_pretrained(version)
        self.device = device
        self.max_length = max_length  # TODO: typical value?
        if freeze:
            self.freeze()

    def freeze(self):
        self.transformer = self.transformer.eval()
        # self.train = disabled_train
        for param in self.parameters():
            param.requires_grad = False

    def forward(self, text):
        batch_encoding = self.tokenizer(text, truncation=True, max_length=self.max_length, return_length=True,
                                        return_overflowing_tokens=False, padding="max_length", return_tensors="pt")
        tokens = batch_encoding["input_ids"].to(self.device)
        outputs = self.transformer(input_ids=tokens)

        z = outputs.last_hidden_state
        return z

    def encode(self, text):
        return self(text)


class FrozenCLIPEmbedder(AbstractEncoder):
    """Uses the CLIP transformer encoder for text (from huggingface)"""
    LAYERS = [
        "last",
        "pooled",
        "hidden"
    ]

    def __init__(self, version="openai/clip-vit-large-patch14", device="cuda", max_length=77,
                 freeze=True, layer="last", layer_idx=None):  # clip-vit-base-patch32
        super().__init__()
        assert layer in self.LAYERS
        self.tokenizer = CLIPTokenizer.from_pretrained(version)
        self.transformer = CLIPTextModel.from_pretrained(version)
        self.device = device
        self.max_length = max_length
        if freeze:
            self.freeze()
        self.layer = layer
        self.layer_idx = layer_idx
        if layer == "hidden":
            assert layer_idx is not None
            assert 0 <= abs(layer_idx) <= 12

    def freeze(self):
        self.transformer = self.transformer.eval()
        # self.train = disabled_train
        for param in self.parameters():
            param.requires_grad = False

    def forward(self, text):
        batch_encoding = self.tokenizer(text, truncation=True, max_length=self.max_length, return_length=True,
                                        return_overflowing_tokens=False, padding="max_length", return_tensors="pt")
        tokens = batch_encoding["input_ids"].to(self.device)
        outputs = self.transformer(input_ids=tokens, output_hidden_states=self.layer == "hidden")
        if self.layer == "last":
            z = outputs.last_hidden_state
        elif self.layer == "pooled":
            z = outputs.pooler_output[:, None, :]
        else:
            z = outputs.hidden_states[self.layer_idx]
        return z

    def encode(self, text):
        return self(text)


class ClipImageEmbedder(nn.Module):
    def __init__(
            self,
            model,
            jit=False,
            device='cuda' if torch.cuda.is_available() else 'cpu',
            antialias=True,
            ucg_rate=0.
    ):
        super().__init__()
        from clip import load as load_clip
        self.model, _ = load_clip(name=model, device=device, jit=jit)

        self.antialias = antialias

        self.register_buffer('mean', torch.Tensor([0.48145466, 0.4578275, 0.40821073]), persistent=False)
        self.register_buffer('std', torch.Tensor([0.26862954, 0.26130258, 0.27577711]), persistent=False)
        self.ucg_rate = ucg_rate

    def preprocess(self, x):
        # normalize to [0,1]
        x = kornia.geometry.resize(x, (224, 224),
                                   interpolation='bicubic', align_corners=True,
                                   antialias=self.antialias)
        x = (x + 1.) / 2.
        # re-normalize according to clip
        x = kornia.enhance.normalize(x, self.mean, self.std)
        return x

    def forward(self, x, no_dropout=False):
        # x is assumed to be in range [-1,1]
        out = self.model.encode_image(self.preprocess(x))
        out = out.to(x.dtype)
        if self.ucg_rate > 0. and not no_dropout:
            out = torch.bernoulli((1. - self.ucg_rate) * torch.ones(out.shape[0], device=out.device))[:, None] * out
        return out


class FrozenOpenCLIPEmbedder(AbstractEncoder):
    """
    Uses the OpenCLIP transformer encoder for text
    """
    LAYERS = [
        # "pooled",
        "last",
        "penultimate"
    ]

    def __init__(self, arch="ViT-H-14", version="laion2b_s32b_b79k", device="cuda", max_length=77,
                 freeze=True, layer="last"):
        super().__init__()
        assert layer in self.LAYERS
        model, _, _ = open_clip.create_model_and_transforms(arch, device=torch.device('cpu'))
        del model.visual
        self.model = model

        self.device = device
        self.max_length = max_length
        if freeze:
            self.freeze()
        self.layer = layer
        if self.layer == "last":
            self.layer_idx = 0
        elif self.layer == "penultimate":
            self.layer_idx = 1
        else:
            raise NotImplementedError()

    def freeze(self):
        self.model = self.model.eval()
        for param in self.parameters():
            param.requires_grad = False

    def forward(self, text):
        self.device = self.model.positional_embedding.device
        tokens = open_clip.tokenize(text)
        z = self.encode_with_transformer(tokens.to(self.device))
        return z

    def encode_with_transformer(self, text):
        x = self.model.token_embedding(text)  # [batch_size, n_ctx, d_model]
        x = x + self.model.positional_embedding
        x = x.permute(1, 0, 2)  # NLD -> LND
        x = self.text_transformer_forward(x, attn_mask=self.model.attn_mask)
        x = x.permute(1, 0, 2)  # LND -> NLD
        x = self.model.ln_final(x)
        return x

    def text_transformer_forward(self, x: torch.Tensor, attn_mask=None):
        for i, r in enumerate(self.model.transformer.resblocks):
            if i == len(self.model.transformer.resblocks) - self.layer_idx:
                break
            if self.model.transformer.grad_checkpointing and not torch.jit.is_scripting():
                x = checkpoint(r, x, attn_mask)
            else:
                x = r(x, attn_mask=attn_mask)
        return x

    def encode(self, text):
        return self(text)


class FrozenOpenCLIPImageEmbedder(AbstractEncoder):
    """
    Uses the OpenCLIP vision transformer encoder for images
    """

    def __init__(self, arch="ViT-H-14", version="laion2b_s32b_b79k", device="cuda", max_length=77,
                 freeze=True, layer="pooled", antialias=True, ucg_rate=0.):
        super().__init__()
        model, _, _ = open_clip.create_model_and_transforms(arch, device=torch.device('cpu'),
                                                            pretrained=version, )
        del model.transformer
        self.model = model

        self.device = device
        self.max_length = max_length
        if freeze:
            self.freeze()
        self.layer = layer
        if self.layer == "penultimate":
            raise NotImplementedError()
            self.layer_idx = 1

        self.antialias = antialias

        self.register_buffer('mean', torch.Tensor([0.48145466, 0.4578275, 0.40821073]), persistent=False)
        self.register_buffer('std', torch.Tensor([0.26862954, 0.26130258, 0.27577711]), persistent=False)
        self.ucg_rate = ucg_rate

    def preprocess(self, x):
        # normalize to [0,1]
        x = kornia.geometry.resize(x, (224, 224),
                                   interpolation='bicubic', align_corners=True,
                                   antialias=self.antialias)
        x = (x + 1.) / 2.
        # renormalize according to clip
        x = kornia.enhance.normalize(x, self.mean, self.std)
        return x

    def freeze(self):
        self.model = self.model.eval()
        for param in self.parameters():
            param.requires_grad = False

    @autocast
    def forward(self, image, no_dropout=False):
        z = self.encode_with_vision_transformer(image)
        if self.ucg_rate > 0. and not no_dropout:
            z = torch.bernoulli((1. - self.ucg_rate) * torch.ones(z.shape[0], device=z.device))[:, None] * z
        return z

    def encode_with_vision_transformer(self, img):
        img = self.preprocess(img)
        x = self.model.visual(img)
        return x

    def encode(self, text):
        return self(text)



class FrozenOpenCLIPImageEmbedderV2(AbstractEncoder):
    """
    Uses the OpenCLIP vision transformer encoder for images
    """

    def __init__(self, arch="ViT-H-14", version="laion2b_s32b_b79k", device="cuda",
                 freeze=True, layer="pooled", antialias=True):
        super().__init__()
        model, _, _ = open_clip.create_model_and_transforms(arch, device=torch.device('cpu'),
                                                            pretrained=version, )
        del model.transformer
        self.model = model
        self.device = device

        if freeze:
            self.freeze()
        self.layer = layer
        if self.layer == "penultimate":
            raise NotImplementedError()
            self.layer_idx = 1

        self.antialias = antialias
        self.register_buffer('mean', torch.Tensor([0.48145466, 0.4578275, 0.40821073]), persistent=False)
        self.register_buffer('std', torch.Tensor([0.26862954, 0.26130258, 0.27577711]), persistent=False)


    def preprocess(self, x):
        # normalize to [0,1]
        x = kornia.geometry.resize(x, (224, 224),
                                   interpolation='bicubic', align_corners=True,
                                   antialias=self.antialias)
        x = (x + 1.) / 2.
        # renormalize according to clip
        x = kornia.enhance.normalize(x, self.mean, self.std)
        return x

    def freeze(self):
        self.model = self.model.eval()
        for param in self.model.parameters():
            param.requires_grad = False

    def forward(self, image, no_dropout=False):
        ## image: b c h w
        z = self.encode_with_vision_transformer(image)
        return z

    def encode_with_vision_transformer(self, x):
        x = self.preprocess(x)

        # to patches - whether to use dual patchnorm - https://arxiv.org/abs/2302.01327v1
        if self.model.visual.input_patchnorm:
            # einops - rearrange(x, 'b c (h p1) (w p2) -> b (h w) (c p1 p2)')
            x = x.reshape(x.shape[0], x.shape[1], self.model.visual.grid_size[0], self.model.visual.patch_size[0], self.model.visual.grid_size[1], self.model.visual.patch_size[1])
            x = x.permute(0, 2, 4, 1, 3, 5)
            x = x.reshape(x.shape[0], self.model.visual.grid_size[0] * self.model.visual.grid_size[1], -1)
            x = self.model.visual.patchnorm_pre_ln(x)
            x = self.model.visual.conv1(x)
        else:
            x = self.model.visual.conv1(x)  # shape = [*, width, grid, grid]
            x = x.reshape(x.shape[0], x.shape[1], -1)  # shape = [*, width, grid ** 2]
            x = x.permute(0, 2, 1)  # shape = [*, grid ** 2, width]

        # class embeddings and positional embeddings
        x = torch.cat(
            [self.model.visual.class_embedding.to(x.dtype) + torch.zeros(x.shape[0], 1, x.shape[-1], dtype=x.dtype, device=x.device),
             x], dim=1)  # shape = [*, grid ** 2 + 1, width]
        x = x + self.model.visual.positional_embedding.to(x.dtype)

        # a patch_dropout of 0. would mean it is disabled and this function would do nothing but return what was passed in
        x = self.model.visual.patch_dropout(x)
        x = self.model.visual.ln_pre(x)

        x = x.permute(1, 0, 2)  # NLD -> LND
        x = self.model.visual.transformer(x)
        x = x.permute(1, 0, 2)  # LND -> NLD

        return x


class FrozenCLIPT5Encoder(AbstractEncoder):
    def __init__(self, clip_version="openai/clip-vit-large-patch14", t5_version="google/t5-v1_1-xl", device="cuda",
                 clip_max_length=77, t5_max_length=77):
        super().__init__()
        self.clip_encoder = FrozenCLIPEmbedder(clip_version, device, max_length=clip_max_length)
        self.t5_encoder = FrozenT5Embedder(t5_version, device, max_length=t5_max_length)
        print(f"{self.clip_encoder.__class__.__name__} has {count_params(self.clip_encoder) * 1.e-6:.2f} M parameters, "
              f"{self.t5_encoder.__class__.__name__} comes with {count_params(self.t5_encoder) * 1.e-6:.2f} M params.")

    def encode(self, text):
        return self(text)

    def forward(self, text):
        clip_z = self.clip_encoder.encode(text)
        t5_z = self.t5_encoder.encode(text)
        return [clip_z, t5_z]