File size: 15,863 Bytes
a25563f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
""" CLIP Model

Adapted from https://github.com/openai/CLIP. Originally MIT License, Copyright (c) 2021 OpenAI.
"""
from dataclasses import dataclass
import logging
import math
from typing import Optional, Tuple, Union

import numpy as np
import torch
import torch.nn.functional as F
from torch import nn
from .transformer import LayerNormFp32, LayerNorm, QuickGELU, Attention, VisionTransformer, TextTransformer
from .utils import to_2tuple

@dataclass
class CLIPVisionCfg:
    layers: Union[Tuple[int, int, int, int], int] = 12
    width: int = 768
    head_width: int = 64
    mlp_ratio: float = 4.0
    patch_size: int = 16
    image_size: Union[Tuple[int, int], int] = 224
    ls_init_value: Optional[float] = None  # layer scale initial value
    patch_dropout: float = 0.  # what fraction of patches to dropout during training (0 would mean disabled and no patches dropped) - 0.5 to 0.75 recommended in the paper for optimal results
    input_patchnorm: bool = False # whether to use dual patchnorm - would only apply the input layernorm on each patch, as post-layernorm already exist in original clip vit design
    global_average_pool: bool = False  # whether to global average pool the last embedding layer, instead of using CLS token (https://arxiv.org/abs/2205.01580)
    attentional_pool: bool = False # whether to use attentional pooler in the last embedding layer
    n_queries: int = 256 # n_queries for attentional pooler
    attn_pooler_heads: int = 8 # n heads for attentional_pooling
    timm_model_name: str = None  # a valid model name overrides layers, width, patch_size
    timm_model_pretrained: bool = False  # use (imagenet) pretrained weights for named model
    timm_pool: str = 'avg'  # feature pooling for timm model ('abs_attn', 'rot_attn', 'avg', '')
    timm_proj: str = 'linear'  # linear projection for timm model output ('linear', 'mlp', '')
    timm_proj_bias: bool = False  # enable bias final projection
    timm_drop: float = 0.  # head dropout
    timm_drop_path: Optional[float] = None  # backbone stochastic depth
    output_tokens: bool = False


@dataclass
class CLIPTextCfg:
    context_length: int = 77
    vocab_size: int = 49408
    width: int = 512
    heads: int = 8
    layers: int = 12
    ls_init_value: Optional[float] = None  # layer scale initial value
    hf_model_name: str = None
    hf_tokenizer_name: str = None
    hf_model_pretrained: bool = True
    proj: str = 'mlp'
    pooler_type: str = 'mean_pooler'
    embed_cls: bool = False
    pad_id: int = 0
    output_tokens: bool = False


def get_cast_dtype(precision: str):
    cast_dtype = None
    if precision == 'bf16':
        cast_dtype = torch.bfloat16
    elif precision == 'fp16':
        cast_dtype = torch.float16
    return cast_dtype


def _build_vision_tower(
        embed_dim: int,
        vision_cfg: CLIPVisionCfg,
        quick_gelu: bool = False,
        cast_dtype: Optional[torch.dtype] = None,
):
    if isinstance(vision_cfg, dict):
        vision_cfg = CLIPVisionCfg(**vision_cfg)

    # OpenAI models are pretrained w/ QuickGELU but native nn.GELU is both faster and more
    # memory efficient in recent PyTorch releases (>= 1.10).
    # NOTE: timm models always use native GELU regardless of quick_gelu flag.
    act_layer = QuickGELU if quick_gelu else nn.GELU

    vision_heads = vision_cfg.width // vision_cfg.head_width
    norm_layer = LayerNormFp32 if cast_dtype in (torch.float16, torch.bfloat16) else LayerNorm
    visual = VisionTransformer(
        image_size=vision_cfg.image_size,
        patch_size=vision_cfg.patch_size,
        width=vision_cfg.width,
        layers=vision_cfg.layers,
        heads=vision_heads,
        mlp_ratio=vision_cfg.mlp_ratio,
        ls_init_value=vision_cfg.ls_init_value,
        patch_dropout=vision_cfg.patch_dropout,
        input_patchnorm=vision_cfg.input_patchnorm,
        global_average_pool=vision_cfg.global_average_pool,
        attentional_pool=vision_cfg.attentional_pool,
        n_queries=vision_cfg.n_queries,
        attn_pooler_heads=vision_cfg.attn_pooler_heads,
        output_tokens=vision_cfg.output_tokens,
        output_dim=embed_dim,
        act_layer=act_layer,
        norm_layer=norm_layer
    )

    return visual


def _build_text_tower(
        embed_dim: int,
        text_cfg: CLIPTextCfg,
        quick_gelu: bool = False,
        cast_dtype: Optional[torch.dtype] = None,
):
    if isinstance(text_cfg, dict):
        text_cfg = CLIPTextCfg(**text_cfg)

    act_layer = QuickGELU if quick_gelu else nn.GELU
    norm_layer = LayerNormFp32 if cast_dtype in (torch.float16, torch.bfloat16) else LayerNorm

    text = TextTransformer(
        context_length=text_cfg.context_length,
        vocab_size=text_cfg.vocab_size,
        width=text_cfg.width,
        heads=text_cfg.heads,
        layers=text_cfg.layers,
        ls_init_value=text_cfg.ls_init_value,
        output_dim=embed_dim,
        embed_cls=text_cfg.embed_cls,
        output_tokens=text_cfg.output_tokens,
        pad_id=text_cfg.pad_id,
        act_layer=act_layer,
        norm_layer=norm_layer
    )

    return text


class CLIP(nn.Module):
    output_dict: torch.jit.Final[bool]

    def __init__(
            self,
            embed_dim: int,
            vision_cfg: CLIPVisionCfg,
            text_cfg: CLIPTextCfg,
            quick_gelu: bool = False,
            cast_dtype: Optional[torch.dtype] = None,
            output_dict: bool = False,
    ):
        super().__init__()
        self.output_dict = output_dict
        self.visual = _build_vision_tower(embed_dim, vision_cfg, quick_gelu, cast_dtype)
        text = _build_text_tower(embed_dim, text_cfg, quick_gelu, cast_dtype)
        self.transformer = text.transformer
        self.vocab_size = text.vocab_size
        self.token_embedding = text.token_embedding
        self.positional_embedding = text.positional_embedding
        self.ln_final = text.ln_final
        self.text_projection = text.text_projection
        self.register_buffer('attn_mask', text.attn_mask, persistent=False)

        self.logit_scale = nn.Parameter(torch.ones([]) * np.log(1 / 0.07))


    def lock_image_tower(self, unlocked_groups=0, freeze_bn_stats=False):
        # lock image tower as per LiT - https://arxiv.org/abs/2111.07991
        self.visual.lock(unlocked_groups=unlocked_groups, freeze_bn_stats=freeze_bn_stats)

    @torch.jit.ignore
    def set_grad_checkpointing(self, enable=True):
        self.visual.set_grad_checkpointing(enable)
        self.transformer.grad_checkpointing = enable

    def encode_image(self, image, out_layers):

        x = image
        # to patches - whether to use dual patchnorm - https://arxiv.org/abs/2302.01327v1
        if self.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.visual.grid_size[0],
                          self.visual.patch_size[0],
                          self.visual.grid_size[1],
                          self.visual.patch_size[1])
            x = x.permute(0, 2, 4, 1, 3, 5)
            x = x.reshape(x.shape[0], self.visual.grid_size[0] * self.visual.grid_size[1], -1)
            x = self.visual.patchnorm_pre_ln(x)
            x = self.visual.conv1(x)
        else:
            x = self.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.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.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.visual.patch_dropout(x)
        x = self.visual.ln_pre(x)

        x = x.permute(1, 0, 2)  # NLD -> LND

        patch_tokens = []

        idx = 0
        for r in self.visual.transformer.resblocks:
            idx += 1
            # add prompt here
            x, attn_tmp = r(x, attn_mask=None)
            if idx in out_layers:
                patch_tokens.append(x)

        x = x.permute(1, 0, 2)  # LND -> NLD
        patch_tokens = [patch_tokens[t].permute(1, 0, 2) for t in range(len(patch_tokens))]  # LND -> NLD

        if self.visual.attn_pool is not None:
            x = self.visual.attn_pool(x)
            x = self.visual.ln_post(x)
            pooled, tokens = self.visual._global_pool(x)
        else:
            pooled, tokens = self.visual._global_pool(x)
            pooled = self.visual.ln_post(pooled)

        if self.visual.proj is not None:
            pooled = pooled @ self.visual.proj

        if self.visual.output_tokens:
            return pooled, patch_tokens

        return pooled, patch_tokens

    def encode_text(self, text):
        cast_dtype = self.transformer.get_cast_dtype()

        x = self.token_embedding(text).to(cast_dtype)  # [batch_size, n_ctx, d_model]

        x = x + self.positional_embedding.to(cast_dtype)
        x = x.permute(1, 0, 2)  # NLD -> LND

        for r in self.visual.transformer.resblocks:
            # add prompt here

            x, attn_tmp = r(x, attn_mask=self.attn_mask)

        x = x.permute(1, 0, 2)  # LND -> NLD
        x = self.ln_final(x)  # [batch_size, n_ctx, transformer.width]

        # take features from the eot embedding (eot_token is the highest number in each sequence)
        x = x[torch.arange(x.shape[0]), text.argmax(dim=-1)] @ self.text_projection
        return x



def convert_weights_to_lp(model: nn.Module, dtype=torch.float16):
    """Convert applicable model parameters to low-precision (bf16 or fp16)"""

    def _convert_weights(l):
        if isinstance(l, (nn.Conv1d, nn.Conv2d, nn.Linear)):
            l.weight.data = l.weight.data.to(dtype)
            if l.bias is not None:
                l.bias.data = l.bias.data.to(dtype)

        if isinstance(l, (nn.MultiheadAttention, Attention)):
            for attr in [*[f"{s}_proj_weight" for s in ["in", "q", "k", "v"]], "in_proj_bias", "bias_k", "bias_v"]:
                tensor = getattr(l, attr)
                if tensor is not None:
                    tensor.data = tensor.data.to(dtype)

        for name in ["text_projection", "proj"]:
            if hasattr(l, name):
                attr = getattr(l, name)
                if attr is not None:
                    attr.data = attr.data.to(dtype)

    model.apply(_convert_weights)


convert_weights_to_fp16 = convert_weights_to_lp  # backwards compat


# used to maintain checkpoint compatibility
def convert_to_custom_text_state_dict(state_dict: dict):
    if 'text_projection' in state_dict:
        # old format state_dict, move text tower -> .text
        new_state_dict = {}
        for k, v in state_dict.items():
            if any(k.startswith(p) for p in (
                'text_projection',
                'positional_embedding',
                'token_embedding',
                'transformer',
                'ln_final',
            )):
                k = 'text.' + k
            new_state_dict[k] = v
        return new_state_dict
    return state_dict


def build_model_from_openai_state_dict(
        state_dict: dict,
        quick_gelu=True,
        cast_dtype=torch.float16,
):
    vit = "visual.proj" in state_dict

    if vit:
        vision_width = state_dict["visual.conv1.weight"].shape[0]
        vision_layers = len(
            [k for k in state_dict.keys() if k.startswith("visual.") and k.endswith(".attn.in_proj_weight")])
        vision_patch_size = state_dict["visual.conv1.weight"].shape[-1]
        grid_size = round((state_dict["visual.positional_embedding"].shape[0] - 1) ** 0.5)
        image_size = vision_patch_size * grid_size
    else:
        counts: list = [
            len(set(k.split(".")[2] for k in state_dict if k.startswith(f"visual.layer{b}"))) for b in [1, 2, 3, 4]]
        vision_layers = tuple(counts)
        vision_width = state_dict["visual.layer1.0.conv1.weight"].shape[0]
        output_width = round((state_dict["visual.attnpool.positional_embedding"].shape[0] - 1) ** 0.5)
        vision_patch_size = None
        assert output_width ** 2 + 1 == state_dict["visual.attnpool.positional_embedding"].shape[0]
        image_size = output_width * 32

    embed_dim = state_dict["text_projection"].shape[1]
    context_length = state_dict["positional_embedding"].shape[0]
    vocab_size = state_dict["token_embedding.weight"].shape[0]
    transformer_width = state_dict["ln_final.weight"].shape[0]
    transformer_heads = transformer_width // 64
    transformer_layers = len(set(k.split(".")[2] for k in state_dict if k.startswith(f"transformer.resblocks")))

    vision_cfg = CLIPVisionCfg(
        layers=vision_layers,
        width=vision_width,
        patch_size=vision_patch_size,
        image_size=image_size,
    )
    text_cfg = CLIPTextCfg(
        context_length=context_length,
        vocab_size=vocab_size,
        width=transformer_width,
        heads=transformer_heads,
        layers=transformer_layers,
    )
    model = CLIP(
        embed_dim,
        vision_cfg=vision_cfg,
        text_cfg=text_cfg,
        quick_gelu=quick_gelu,  # OpenAI models were trained with QuickGELU
        cast_dtype=cast_dtype,
    )

    for key in ["input_resolution", "context_length", "vocab_size"]:
        state_dict.pop(key, None)

    convert_weights_to_fp16(model)  # OpenAI state dicts are partially converted to float16
    model.load_state_dict(state_dict)
    return model.eval()


def trace_model(model, batch_size=256, device=torch.device('cpu')):
    model.eval()
    image_size = model.visual.image_size
    example_images = torch.ones((batch_size, 3, image_size, image_size), device=device)
    example_text = torch.zeros((batch_size, model.context_length), dtype=torch.int, device=device)
    model = torch.jit.trace_module(
        model,
        inputs=dict(
            forward=(example_images, example_text),
            encode_text=(example_text,),
            encode_image=(example_images,)
        ))
    model.visual.image_size = image_size
    return model


def resize_pos_embed(state_dict, model, interpolation: str = 'bicubic'):
    # Rescale the grid of position embeddings when loading from state_dict
    old_pos_embed = state_dict.get('visual.positional_embedding', None)
    if old_pos_embed is None or not hasattr(model.visual, 'grid_size'):
        return
    grid_size = to_2tuple(model.visual.grid_size)
    extra_tokens = 1  # FIXME detect different token configs (ie no class token, or more)
    new_seq_len = grid_size[0] * grid_size[1] + extra_tokens
    if new_seq_len == old_pos_embed.shape[0]:
        return

    if extra_tokens:
        pos_emb_tok, pos_emb_img = old_pos_embed[:extra_tokens], old_pos_embed[extra_tokens:]
    else:
        pos_emb_tok, pos_emb_img = None, old_pos_embed
    old_grid_size = to_2tuple(int(math.sqrt(len(pos_emb_img))))

    logging.info('Resizing position embedding grid-size from %s to %s', old_grid_size, grid_size)
    pos_emb_img = pos_emb_img.reshape(1, old_grid_size[0], old_grid_size[1], -1).permute(0, 3, 1, 2)
    pos_emb_img = F.interpolate(
        pos_emb_img,
        size=grid_size,
        mode=interpolation,
        align_corners=False,
    )
    pos_emb_img = pos_emb_img.permute(0, 2, 3, 1).reshape(1, grid_size[0] * grid_size[1], -1)[0]
    if pos_emb_tok is not None:
        new_pos_embed = torch.cat([pos_emb_tok, pos_emb_img], dim=0)
    else:
        new_pos_embed = pos_emb_img
    state_dict['visual.positional_embedding'] = new_pos_embed