File size: 11,580 Bytes
74e4bcd |
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 |
# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.
# Modified from github.com/openai/CLIP
from collections import OrderedDict
import numpy as np
import timm
import torch
from torch import nn
import losses
class LayerNorm(nn.LayerNorm):
"""Subclass torch's LayerNorm to handle fp16."""
def forward(self, x: torch.Tensor):
orig_type = x.dtype
ret = super().forward(x.type(torch.float32))
return ret.type(orig_type)
class QuickGELU(nn.Module):
def forward(self, x: torch.Tensor):
return x * torch.sigmoid(1.702 * x)
class ResidualAttentionBlock(nn.Module):
def __init__(self, d_model: int, n_head: int, attn_mask: torch.Tensor = None):
super().__init__()
self.attn = nn.MultiheadAttention(d_model, n_head)
self.ln_1 = LayerNorm(d_model)
self.mlp = nn.Sequential(OrderedDict([
("c_fc", nn.Linear(d_model, d_model * 4)),
("gelu", QuickGELU()),
("c_proj", nn.Linear(d_model * 4, d_model))
]))
self.ln_2 = LayerNorm(d_model)
self.attn_mask = attn_mask
def attention(self, x: torch.Tensor):
self.attn_mask = self.attn_mask.to(dtype=x.dtype, device=x.device) if self.attn_mask is not None else None
return self.attn(x, x, x, need_weights=False, attn_mask=self.attn_mask)[0]
def forward(self, x: torch.Tensor):
x = x + self.attention(self.ln_1(x))
x = x + self.mlp(self.ln_2(x))
return x
class Transformer(nn.Module):
def __init__(self, width: int, layers: int, heads: int, attn_mask: torch.Tensor = None):
super().__init__()
self.width = width
self.layers = layers
self.resblocks = nn.Sequential(*[ResidualAttentionBlock(width, heads, attn_mask) for _ in range(layers)])
def forward(self, x: torch.Tensor):
return self.resblocks(x)
class CLIP(nn.Module):
def __init__(self,
embed_dim: int,
# vision
vision_width: int,
vision_model: nn.Module,
# text
context_length: int,
vocab_size: int,
transformer_width: int,
transformer_heads: int,
transformer_layers: int,
**kwargs,
):
super().__init__()
self.context_length = context_length
self.vision_width = vision_width
self.visual = vision_model
self.transformer = Transformer(
width=transformer_width,
layers=transformer_layers,
heads=transformer_heads,
attn_mask=self.build_attention_mask(),
)
self.vocab_size = vocab_size
self.token_embedding = nn.Embedding(vocab_size, transformer_width)
self.positional_embedding = nn.Parameter(torch.empty(self.context_length, transformer_width))
self.ln_final = LayerNorm(transformer_width)
self.image_projection = nn.Parameter(torch.empty(vision_width, embed_dim))
self.text_projection = nn.Parameter(torch.empty(transformer_width, embed_dim))
self.logit_scale = nn.Parameter(torch.ones([]) * np.log(1 / 0.07))
self.initialize_parameters()
def initialize_parameters(self):
nn.init.normal_(self.token_embedding.weight, std=0.02)
nn.init.normal_(self.positional_embedding, std=0.01)
proj_std = (self.transformer.width ** -0.5) * ((2 * self.transformer.layers) ** -0.5)
attn_std = self.transformer.width ** -0.5
fc_std = (2 * self.transformer.width) ** -0.5
for block in self.transformer.resblocks:
nn.init.normal_(block.attn.in_proj_weight, std=attn_std)
nn.init.normal_(block.attn.out_proj.weight, std=proj_std)
nn.init.normal_(block.mlp.c_fc.weight, std=fc_std)
nn.init.normal_(block.mlp.c_proj.weight, std=proj_std)
nn.init.normal_(self.image_projection, std=self.vision_width ** -0.5)
nn.init.normal_(self.text_projection, std=self.transformer.width ** -0.5)
def build_attention_mask(self):
# lazily create causal attention mask, with full attention between the vision tokens
# pytorch uses additive attention mask; fill with -inf
mask = torch.empty(self.context_length, self.context_length)
mask.fill_(float("-inf"))
mask.triu_(1) # zero out the lower diagonal
return mask
def encode_image(self, image):
x = self.visual(image)
x = x @ self.image_projection
return x
def encode_text(self, text):
x = self.token_embedding(text) # [batch_size, n_ctx, d_model]
x = x + self.positional_embedding
x = x.permute(1, 0, 2) # NLD -> LND
x = self.transformer(x)
x = x.permute(1, 0, 2) # LND -> NLD
x = self.ln_final(x)
# x.shape = [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 forward(self, image, text):
image_embed = self.encode_image(image)
text_embed = self.encode_text(text)
return {'image_embed': image_embed,
'text_embed': text_embed,
'logit_scale': self.logit_scale.exp()}
class SIMCLR(nn.Module):
def __init__(self,
# vision
vision_width: int,
vision_model: nn.Module,
# ssl
ssl_mlp_dim: int,
ssl_emb_dim: int,
**kwargs,
):
super().__init__()
self.vision_width = vision_width
self.visual = vision_model
self.image_mlp = self._build_mlp(in_dim=vision_width, mlp_dim=ssl_mlp_dim, out_dim=ssl_emb_dim)
def _build_mlp(self, in_dim, mlp_dim, out_dim):
return nn.Sequential(OrderedDict([
("layer1", nn.Linear(in_dim, mlp_dim)),
("bn1", nn.SyncBatchNorm(mlp_dim)),
("relu1", nn.ReLU(inplace=True)),
("layer2", nn.Linear(mlp_dim, mlp_dim)),
("bn2", nn.SyncBatchNorm(mlp_dim)),
("relu2", nn.ReLU(inplace=True)),
("layer3", nn.Linear(mlp_dim, out_dim)),
]))
def encode_image(self, image):
x = self.visual(image)
return x
def forward(self, aug1, aug2):
h1 = self.visual(aug1)
h2 = self.visual(aug2)
aug1_embed = self.image_mlp(h1)
aug2_embed = self.image_mlp(h2)
return {'aug1_embed': aug1_embed,
'aug2_embed': aug2_embed}
class SLIP(CLIP):
def __init__(self,
ssl_mlp_dim: int,
ssl_emb_dim: int,
**kwargs,
):
super().__init__(**kwargs)
self.image_mlp = self._build_mlp(in_dim=self.vision_width, mlp_dim=ssl_mlp_dim, out_dim=ssl_emb_dim)
def _build_mlp(self, in_dim, mlp_dim, out_dim):
return nn.Sequential(OrderedDict([
("layer1", nn.Linear(in_dim, mlp_dim)),
("bn1", nn.SyncBatchNorm(mlp_dim)),
("relu1", nn.ReLU(inplace=True)),
("layer2", nn.Linear(mlp_dim, mlp_dim)),
("bn2", nn.SyncBatchNorm(mlp_dim)),
("relu2", nn.ReLU(inplace=True)),
("layer3", nn.Linear(mlp_dim, out_dim)),
]))
def forward(self, image, text, aug1, aug2):
aug1_embed = self.image_mlp(self.visual(aug1))
aug2_embed = self.image_mlp(self.visual(aug2))
image_embed = self.encode_image(image)
text_embed = self.encode_text(text)
return {'image_embed': image_embed,
'text_embed': text_embed,
'logit_scale': self.logit_scale.exp(),
'aug1_embed': aug1_embed,
'aug2_embed': aug2_embed}
def get_loss(model, ssl_temp, ssl_scale):
if model.startswith('SLIP'):
ssl_loss = losses.SIMCLRLoss(temperature=ssl_temp)
return losses.SLIPLoss(ssl_loss, ssl_scale)
if model.startswith('CLIP'):
return losses.CLIPLoss()
if model.startswith('SIMCLR'):
return losses.SIMCLRLoss(temperature=ssl_temp)
def get_metric_names(model):
if model.startswith('SLIP'):
return ['loss', 'clip_loss', 'ssl_loss', 'clip_acc', 'ssl_acc']
elif model.startswith('CLIP'):
return ['loss', 'clip_loss', 'clip_acc']
else:
return ['loss', 'ssl_loss', 'ssl_acc']
@timm.models.registry.register_model
def vit_small_mocov3_patch16_224(**kwargs):
model_kwargs = dict(patch_size=16, embed_dim=384, depth=12, num_heads=12, **kwargs)
model = timm.models.vision_transformer._create_vision_transformer('vit_small_patch16_224', **model_kwargs)
return model
def CLIP_VITS16(**kwargs):
vision_model = timm.create_model('vit_small_mocov3_patch16_224', num_classes=0)
model = CLIP(embed_dim=512, vision_width=384, vision_model=vision_model, context_length=77, vocab_size=49408,
transformer_width=512, transformer_heads=8, transformer_layers=12, **kwargs)
return model
def SIMCLR_VITS16(**kwargs):
vision_model = timm.create_model('vit_small_mocov3_patch16_224', num_classes=0)
model = SIMCLR(vision_width=384, vision_model=vision_model, **kwargs)
return model
def SLIP_VITS16(**kwargs):
vision_model = timm.create_model('vit_small_mocov3_patch16_224', num_classes=0)
model = SLIP(embed_dim=512, vision_width=384, vision_model=vision_model, context_length=77, vocab_size=49408,
transformer_width=512, transformer_heads=8, transformer_layers=12, **kwargs)
return model
def CLIP_VITB16(**kwargs):
vision_model = timm.create_model('vit_base_patch16_224', num_classes=0)
model = CLIP(embed_dim=512, vision_width=768, vision_model=vision_model, context_length=77, vocab_size=49408,
transformer_width=512, transformer_heads=8, transformer_layers=12, **kwargs)
return model
def SIMCLR_VITB16(**kwargs):
vision_model = timm.create_model('vit_base_patch16_224', num_classes=0)
model = SIMCLR(vision_width=768, vision_model=vision_model, **kwargs)
return model
def SLIP_VITB16(**kwargs):
vision_model = timm.create_model('vit_base_patch16_224', num_classes=0)
model = SLIP(embed_dim=512, vision_width=768, vision_model=vision_model, context_length=77, vocab_size=49408,
transformer_width=512, transformer_heads=8, transformer_layers=12, **kwargs)
return model
def CLIP_VITL16(**kwargs):
vision_model = timm.create_model('vit_large_patch16_224', num_classes=0)
model = CLIP(embed_dim=512, vision_width=1024, vision_model=vision_model, context_length=77, vocab_size=49408,
transformer_width=512, transformer_heads=8, transformer_layers=12, **kwargs)
return model
def SIMCLR_VITL16(**kwargs):
vision_model = timm.create_model('vit_large_patch16_224', num_classes=0)
model = SIMCLR(vision_width=1024, vision_model=vision_model, **kwargs)
return model
def SLIP_VITL16(**kwargs):
vision_model = timm.create_model('vit_large_patch16_224', num_classes=0)
model = SLIP(embed_dim=512, vision_width=1024, vision_model=vision_model, context_length=77, vocab_size=49408,
transformer_width=512, transformer_heads=8, transformer_layers=12, **kwargs)
return model
|