File size: 20,689 Bytes
8dd41a8 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 |
import math
from os.path import basename, dirname, join, isfile
import torch
from torch import nn
from torch.nn import functional as nnf
from torch.nn.modules.activation import ReLU
def get_prompt_list(prompt):
if prompt == 'plain':
return ['{}']
elif prompt == 'fixed':
return ['a photo of a {}.']
elif prompt == 'shuffle':
return ['a photo of a {}.', 'a photograph of a {}.', 'an image of a {}.', '{}.']
elif prompt == 'shuffle+':
return ['a photo of a {}.', 'a photograph of a {}.', 'an image of a {}.', '{}.',
'a cropped photo of a {}.', 'a good photo of a {}.', 'a photo of one {}.',
'a bad photo of a {}.', 'a photo of the {}.']
else:
raise ValueError('Invalid value for prompt')
def forward_multihead_attention(x, b, with_aff=False, attn_mask=None):
"""
Simplified version of multihead attention (taken from torch source code but without tons of if clauses).
The mlp and layer norm come from CLIP.
x: input.
b: multihead attention module.
"""
x_ = b.ln_1(x)
q, k, v = nnf.linear(x_, b.attn.in_proj_weight, b.attn.in_proj_bias).chunk(3, dim=-1)
tgt_len, bsz, embed_dim = q.size()
head_dim = embed_dim // b.attn.num_heads
scaling = float(head_dim) ** -0.5
q = q.contiguous().view(tgt_len, bsz * b.attn.num_heads, b.attn.head_dim).transpose(0, 1)
k = k.contiguous().view(-1, bsz * b.attn.num_heads, b.attn.head_dim).transpose(0, 1)
v = v.contiguous().view(-1, bsz * b.attn.num_heads, b.attn.head_dim).transpose(0, 1)
q = q * scaling
attn_output_weights = torch.bmm(q, k.transpose(1, 2)) # n_heads * batch_size, tokens^2, tokens^2
if attn_mask is not None:
attn_mask_type, attn_mask = attn_mask
n_heads = attn_output_weights.size(0) // attn_mask.size(0)
attn_mask = attn_mask.repeat(n_heads, 1)
if attn_mask_type == 'cls_token':
# the mask only affects similarities compared to the readout-token.
attn_output_weights[:, 0, 1:] = attn_output_weights[:, 0, 1:] * attn_mask[None,...]
# attn_output_weights[:, 0, 0] = 0*attn_output_weights[:, 0, 0]
if attn_mask_type == 'all':
# print(attn_output_weights.shape, attn_mask[:, None].shape)
attn_output_weights[:, 1:, 1:] = attn_output_weights[:, 1:, 1:] * attn_mask[:, None]
attn_output_weights = torch.softmax(attn_output_weights, dim=-1)
attn_output = torch.bmm(attn_output_weights, v)
attn_output = attn_output.transpose(0, 1).contiguous().view(tgt_len, bsz, embed_dim)
attn_output = b.attn.out_proj(attn_output)
x = x + attn_output
x = x + b.mlp(b.ln_2(x))
if with_aff:
return x, attn_output_weights
else:
return x
class CLIPDenseBase(nn.Module):
def __init__(self, version, reduce_cond, reduce_dim, prompt, n_tokens):
super().__init__()
import clip
# prec = torch.FloatTensor
self.clip_model, _ = clip.load(version, device='cpu', jit=False)
self.model = self.clip_model.visual
# if not None, scale conv weights such that we obtain n_tokens.
self.n_tokens = n_tokens
for p in self.clip_model.parameters():
p.requires_grad_(False)
# conditional
if reduce_cond is not None:
self.reduce_cond = nn.Linear(512, reduce_cond)
for p in self.reduce_cond.parameters():
p.requires_grad_(False)
else:
self.reduce_cond = None
self.film_mul = nn.Linear(512 if reduce_cond is None else reduce_cond, reduce_dim)
self.film_add = nn.Linear(512 if reduce_cond is None else reduce_cond, reduce_dim)
self.reduce = nn.Linear(768, reduce_dim)
self.prompt_list = get_prompt_list(prompt)
# precomputed prompts
import pickle
if isfile('precomputed_prompt_vectors.pickle'):
precomp = pickle.load(open('precomputed_prompt_vectors.pickle', 'rb'))
self.precomputed_prompts = {k: torch.from_numpy(v) for k, v in precomp.items()}
else:
self.precomputed_prompts = dict()
def rescaled_pos_emb(self, new_size):
assert len(new_size) == 2
a = self.model.positional_embedding[1:].T.view(1, 768, *self.token_shape)
b = nnf.interpolate(a, new_size, mode='bicubic', align_corners=False).squeeze(0).view(768, new_size[0]*new_size[1]).T
return torch.cat([self.model.positional_embedding[:1], b])
def visual_forward(self, x_inp, extract_layers=(), skip=False, mask=None):
with torch.no_grad():
inp_size = x_inp.shape[2:]
if self.n_tokens is not None:
stride2 = x_inp.shape[2] // self.n_tokens
conv_weight2 = nnf.interpolate(self.model.conv1.weight, (stride2, stride2), mode='bilinear', align_corners=True)
x = nnf.conv2d(x_inp, conv_weight2, bias=self.model.conv1.bias, stride=stride2, dilation=self.model.conv1.dilation)
else:
x = self.model.conv1(x_inp) # 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]
x = torch.cat([self.model.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]
standard_n_tokens = 50 if self.model.conv1.kernel_size[0] == 32 else 197
if x.shape[1] != standard_n_tokens:
new_shape = int(math.sqrt(x.shape[1]-1))
x = x + self.rescaled_pos_emb((new_shape, new_shape)).to(x.dtype)[None,:,:]
else:
x = x + self.model.positional_embedding.to(x.dtype)
x = self.model.ln_pre(x)
x = x.permute(1, 0, 2) # NLD -> LND
activations, affinities = [], []
for i, res_block in enumerate(self.model.transformer.resblocks):
if mask is not None:
mask_layer, mask_type, mask_tensor = mask
if mask_layer == i or mask_layer == 'all':
# import ipdb; ipdb.set_trace()
size = int(math.sqrt(x.shape[0] - 1))
attn_mask = (mask_type, nnf.interpolate(mask_tensor.unsqueeze(1).float(), (size, size)).view(mask_tensor.shape[0], size * size))
else:
attn_mask = None
else:
attn_mask = None
x, aff_per_head = forward_multihead_attention(x, res_block, with_aff=True, attn_mask=attn_mask)
if i in extract_layers:
affinities += [aff_per_head]
#if self.n_tokens is not None:
# activations += [nnf.interpolate(x, inp_size, mode='bilinear', align_corners=True)]
#else:
activations += [x]
if len(extract_layers) > 0 and i == max(extract_layers) and skip:
print('early skip')
break
x = x.permute(1, 0, 2) # LND -> NLD
x = self.model.ln_post(x[:, 0, :])
if self.model.proj is not None:
x = x @ self.model.proj
return x, activations, affinities
def sample_prompts(self, words, prompt_list=None):
prompt_list = prompt_list if prompt_list is not None else self.prompt_list
prompt_indices = torch.multinomial(torch.ones(len(prompt_list)), len(words), replacement=True)
prompts = [prompt_list[i] for i in prompt_indices]
return [promt.format(w) for promt, w in zip(prompts, words)]
def get_cond_vec(self, conditional, batch_size):
# compute conditional from a single string
if conditional is not None and type(conditional) == str:
cond = self.compute_conditional(conditional)
cond = cond.repeat(batch_size, 1)
# compute conditional from string list/tuple
elif conditional is not None and type(conditional) in {list, tuple} and type(conditional[0]) == str:
assert len(conditional) == batch_size
cond = self.compute_conditional(conditional)
# use conditional directly
elif conditional is not None and type(conditional) == torch.Tensor and conditional.ndim == 2:
cond = conditional
# compute conditional from image
elif conditional is not None and type(conditional) == torch.Tensor:
with torch.no_grad():
cond, _, _ = self.visual_forward(conditional)
else:
raise ValueError('invalid conditional')
return cond
def compute_conditional(self, conditional):
import clip
dev = next(self.parameters()).device
if type(conditional) in {list, tuple}:
text_tokens = clip.tokenize(conditional).to(dev)
cond = self.clip_model.encode_text(text_tokens)
else:
if conditional in self.precomputed_prompts:
cond = self.precomputed_prompts[conditional].float().to(dev)
else:
text_tokens = clip.tokenize([conditional]).to(dev)
cond = self.clip_model.encode_text(text_tokens)[0]
if self.shift_vector is not None:
return cond + self.shift_vector
else:
return cond
def clip_load_untrained(version):
assert version == 'ViT-B/16'
from clip.model import CLIP
from clip.clip import _MODELS, _download
model = torch.jit.load(_download(_MODELS['ViT-B/16'])).eval()
state_dict = model.state_dict()
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_resolution = vision_patch_size * grid_size
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")))
return CLIP(embed_dim, image_resolution, vision_layers, vision_width, vision_patch_size,
context_length, vocab_size, transformer_width, transformer_heads, transformer_layers)
class CLIPDensePredT(CLIPDenseBase):
def __init__(self, version='ViT-B/32', extract_layers=(3, 6, 9), cond_layer=0, reduce_dim=128, n_heads=4, prompt='fixed',
extra_blocks=0, reduce_cond=None, fix_shift=False,
learn_trans_conv_only=False, limit_to_clip_only=False, upsample=False,
add_calibration=False, rev_activations=False, trans_conv=None, n_tokens=None, complex_trans_conv=False):
super().__init__(version, reduce_cond, reduce_dim, prompt, n_tokens)
# device = 'cpu'
self.extract_layers = extract_layers
self.cond_layer = cond_layer
self.limit_to_clip_only = limit_to_clip_only
self.process_cond = None
self.rev_activations = rev_activations
depth = len(extract_layers)
if add_calibration:
self.calibration_conds = 1
self.upsample_proj = nn.Conv2d(reduce_dim, 1, kernel_size=1) if upsample else None
self.add_activation1 = True
self.version = version
self.token_shape = {'ViT-B/32': (7, 7), 'ViT-B/16': (14, 14)}[version]
if fix_shift:
# self.shift_vector = nn.Parameter(torch.load(join(dirname(basename(__file__)), 'clip_text_shift_vector.pth')), requires_grad=False)
self.shift_vector = nn.Parameter(torch.load(join(dirname(basename(__file__)), 'shift_text_to_vis.pth')), requires_grad=False)
# self.shift_vector = nn.Parameter(-1*torch.load(join(dirname(basename(__file__)), 'shift2.pth')), requires_grad=False)
else:
self.shift_vector = None
if trans_conv is None:
trans_conv_ks = {'ViT-B/32': (32, 32), 'ViT-B/16': (16, 16)}[version]
else:
# explicitly define transposed conv kernel size
trans_conv_ks = (trans_conv, trans_conv)
if not complex_trans_conv:
self.trans_conv = nn.ConvTranspose2d(reduce_dim, 1, trans_conv_ks, stride=trans_conv_ks)
else:
assert trans_conv_ks[0] == trans_conv_ks[1]
tp_kernels = (trans_conv_ks[0] // 4, trans_conv_ks[0] // 4)
self.trans_conv = nn.Sequential(
nn.Conv2d(reduce_dim, reduce_dim, kernel_size=3, padding=1),
nn.ReLU(),
nn.ConvTranspose2d(reduce_dim, reduce_dim // 2, kernel_size=tp_kernels[0], stride=tp_kernels[0]),
nn.ReLU(),
nn.ConvTranspose2d(reduce_dim // 2, 1, kernel_size=tp_kernels[1], stride=tp_kernels[1]),
)
# self.trans_conv = nn.ConvTranspose2d(reduce_dim, 1, trans_conv_ks, stride=trans_conv_ks)
assert len(self.extract_layers) == depth
self.reduces = nn.ModuleList([nn.Linear(768, reduce_dim) for _ in range(depth)])
self.blocks = nn.ModuleList([nn.TransformerEncoderLayer(d_model=reduce_dim, nhead=n_heads) for _ in range(len(self.extract_layers))])
self.extra_blocks = nn.ModuleList([nn.TransformerEncoderLayer(d_model=reduce_dim, nhead=n_heads) for _ in range(extra_blocks)])
# refinement and trans conv
if learn_trans_conv_only:
for p in self.parameters():
p.requires_grad_(False)
for p in self.trans_conv.parameters():
p.requires_grad_(True)
self.prompt_list = get_prompt_list(prompt)
def forward(self, inp_image, conditional=None, return_features=False, mask=None):
assert type(return_features) == bool
inp_image = inp_image.to(self.model.positional_embedding.device)
if mask is not None:
raise ValueError('mask not supported')
# x_inp = normalize(inp_image)
x_inp = inp_image
bs, dev = inp_image.shape[0], x_inp.device
cond = self.get_cond_vec(conditional, bs)
visual_q, activations, _ = self.visual_forward(x_inp, extract_layers=[0] + list(self.extract_layers))
activation1 = activations[0]
activations = activations[1:]
_activations = activations[::-1] if not self.rev_activations else activations
a = None
for i, (activation, block, reduce) in enumerate(zip(_activations, self.blocks, self.reduces)):
if a is not None:
a = reduce(activation) + a
else:
a = reduce(activation)
if i == self.cond_layer:
if self.reduce_cond is not None:
cond = self.reduce_cond(cond)
a = self.film_mul(cond) * a + self.film_add(cond)
a = block(a)
for block in self.extra_blocks:
a = a + block(a)
a = a[1:].permute(1, 2, 0) # rm cls token and -> BS, Feats, Tokens
size = int(math.sqrt(a.shape[2]))
a = a.view(bs, a.shape[1], size, size)
a = self.trans_conv(a)
if self.n_tokens is not None:
a = nnf.interpolate(a, x_inp.shape[2:], mode='bilinear', align_corners=True)
if self.upsample_proj is not None:
a = self.upsample_proj(a)
a = nnf.interpolate(a, x_inp.shape[2:], mode='bilinear')
if return_features:
return a, visual_q, cond, [activation1] + activations
else:
return a,
class CLIPDensePredTMasked(CLIPDensePredT):
def __init__(self, version='ViT-B/32', extract_layers=(3, 6, 9), cond_layer=0, reduce_dim=128, n_heads=4,
prompt='fixed', extra_blocks=0, reduce_cond=None, fix_shift=False, learn_trans_conv_only=False,
refine=None, limit_to_clip_only=False, upsample=False, add_calibration=False, n_tokens=None):
super().__init__(version=version, extract_layers=extract_layers, cond_layer=cond_layer, reduce_dim=reduce_dim,
n_heads=n_heads, prompt=prompt, extra_blocks=extra_blocks, reduce_cond=reduce_cond,
fix_shift=fix_shift, learn_trans_conv_only=learn_trans_conv_only,
limit_to_clip_only=limit_to_clip_only, upsample=upsample, add_calibration=add_calibration,
n_tokens=n_tokens)
def visual_forward_masked(self, img_s, seg_s):
return super().visual_forward(img_s, mask=('all', 'cls_token', seg_s))
def forward(self, img_q, cond_or_img_s, seg_s=None, return_features=False):
if seg_s is None:
cond = cond_or_img_s
else:
img_s = cond_or_img_s
with torch.no_grad():
cond, _, _ = self.visual_forward_masked(img_s, seg_s)
return super().forward(img_q, cond, return_features=return_features)
class CLIPDenseBaseline(CLIPDenseBase):
def __init__(self, version='ViT-B/32', cond_layer=0,
extract_layer=9, reduce_dim=128, reduce2_dim=None, prompt='fixed',
reduce_cond=None, limit_to_clip_only=False, n_tokens=None):
super().__init__(version, reduce_cond, reduce_dim, prompt, n_tokens)
device = 'cpu'
# self.cond_layer = cond_layer
self.extract_layer = extract_layer
self.limit_to_clip_only = limit_to_clip_only
self.shift_vector = None
self.token_shape = {'ViT-B/32': (7, 7), 'ViT-B/16': (14, 14)}[version]
assert reduce2_dim is not None
self.reduce2 = nn.Sequential(
nn.Linear(reduce_dim, reduce2_dim),
nn.ReLU(),
nn.Linear(reduce2_dim, reduce_dim)
)
trans_conv_ks = {'ViT-B/32': (32, 32), 'ViT-B/16': (16, 16)}[version]
self.trans_conv = nn.ConvTranspose2d(reduce_dim, 1, trans_conv_ks, stride=trans_conv_ks)
def forward(self, inp_image, conditional=None, return_features=False):
inp_image = inp_image.to(self.model.positional_embedding.device)
# x_inp = normalize(inp_image)
x_inp = inp_image
bs, dev = inp_image.shape[0], x_inp.device
cond = self.get_cond_vec(conditional, bs)
visual_q, activations, affinities = self.visual_forward(x_inp, extract_layers=[self.extract_layer])
a = activations[0]
a = self.reduce(a)
a = self.film_mul(cond) * a + self.film_add(cond)
if self.reduce2 is not None:
a = self.reduce2(a)
# the original model would execute a transformer block here
a = a[1:].permute(1, 2, 0) # rm cls token and -> BS, Feats, Tokens
size = int(math.sqrt(a.shape[2]))
a = a.view(bs, a.shape[1], size, size)
a = self.trans_conv(a)
if return_features:
return a, visual_q, cond, activations
else:
return a,
class CLIPSegMultiLabel(nn.Module):
def __init__(self, model) -> None:
super().__init__()
from third_party.JoEm.data_loader import get_seen_idx, get_unseen_idx, VOC
self.pascal_classes = VOC
from clip.clipseg import CLIPDensePredT
from general_utils import load_model
# self.clipseg = load_model('rd64-vit16-neg0.2-phrasecut', strict=False)
self.clipseg = load_model(model, strict=False)
self.clipseg.eval()
def forward(self, x):
bs = x.shape[0]
out = torch.ones(21, bs, 352, 352).to(x.device) * -10
for class_id, class_name in enumerate(self.pascal_classes):
fac = 3 if class_name == 'background' else 1
with torch.no_grad():
pred = torch.sigmoid(self.clipseg(x, class_name)[0][:,0]) * fac
out[class_id] += pred
out = out.permute(1, 0, 2, 3)
return out
# construct output tensor
|