Spaces:
Runtime error
Runtime error
File size: 11,644 Bytes
8fc2b4e |
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 |
import torch
import torch.nn as nn
import torch.nn.functional as F
import numpy as np
class DotAttn(nn.Module):
""" Dot-Attention """
def forward(self, inp, h):
score = self.softmax(inp, h)
return score.expand_as(inp).mul(inp).sum(1), score
def softmax(self, inp, h):
raw_score = inp.bmm(h.unsqueeze(2))
score = F.softmax(raw_score, dim=1)
return score
class ScaledDotAttn(nn.Module):
""" Scaled Dot-Attention """
def forward(self, inp, h):
score = self.softmax(inp, h)
return score.expand_as(inp).mul(inp).sum(1), score
def softmax(self, inp, h):
raw_score = inp.bmm(h.unsqueeze(2)) / np.sqrt(h.shape[-1])
score = F.softmax(raw_score, dim=1)
return score
class Fusion(nn.Module):
""" Base Fusion Class"""
def __init__(self, input_dim=3):
super().__init__()
self.input_dim = input_dim
def tile_x2(self, x1, x2, x2_proj=None):
if x2_proj:
x2 = x2_proj(x2)
x2 = x2.unsqueeze(-1).unsqueeze(-1)
x2 = x2.repeat(x1.shape[0], 1, x1.shape[-2], x1.shape[-1])
return x2
def batch_tile_x2(self, x1, x2, x2_proj=None):
if x2_proj:
x2 = x2_proj(x2)
x2 = x2.unsqueeze(-1).unsqueeze(-1)
x2 = x2.repeat(1, 1, x1.shape[-2], x1.shape[-1])
return x2
def forward(self, x1, x2, x2_mask=None, x2_proj=None):
raise NotImplementedError()
class FusionAdd(Fusion):
""" x1 + x2 """
def __init__(self, input_dim=3):
super(FusionAdd, self).__init__(input_dim=input_dim)
def forward(self, x1, x2, x2_mask=None, x2_proj=None):
if x1.shape != x2.shape and len(x1.shape) != len(x2.shape):
x2 = self.tile_x2(x1, x2, x2_proj)
return x1 + x2
class FusionMult(Fusion):
""" x1 * x2 """
def __init__(self, input_dim=3):
super(FusionMult, self).__init__(input_dim=input_dim)
def forward(self, x1, x2, x2_mask=None, x2_proj=None):
if x1.shape != x2.shape and len(x1.shape) != len(x2.shape):
x2 = self.batch_tile_x2(x1, x2, x2_proj) # self.batch_tile_x2(x1, x2, x2_proj)
return x1 * x2
class FusionMax(Fusion):
""" max(x1, x2) """
def __init__(self, input_dim=3):
super(FusionMax, self).__init__(input_dim=input_dim)
def forward(self, x1, x2, x2_mask=None, x2_proj=None):
if x1.shape != x2.shape and len(x1.shape) != len(x2.shape):
x2 = self.tile_x2(x1, x2, x2_proj)
return torch.max(x1, x2)
class FusionConcat(Fusion):
""" [x1; x2] """
def __init__(self, input_dim=3):
super(FusionConcat, self).__init__(input_dim=input_dim)
def forward(self, x1, x2, x2_mask=None, x2_proj=None):
if x1.shape != x2.shape and len(x1.shape) != len(x2.shape):
x2 = self.tile_x2(x1, x2, x2_proj)
return torch.cat([x1, x2], dim=1)
class FusionConv(Fusion):
""" 1x1 convs after [x1; x2] """
def __init__(self, input_dim=3):
super(FusionConv, self).__init__(input_dim=input_dim)
self.conv = nn.Sequential(
nn.ReLU(True),
nn.Conv2d(input_dim * 2, input_dim, kernel_size=1, bias=False)
)
def forward(self, x1, x2, x2_mask=None, x2_proj=None):
if x1.shape != x2.shape and len(x1.shape) != len(x2.shape):
x2 = self.tile_x2(x1, x2, x2_proj)
x = torch.cat([x1, x2], dim=1) # [B, 2C, H, W]
x = self.conv(x) # [B, C, H, W]
return x
class FusionConvLat(Fusion):
""" 1x1 convs after [x1; x2] for lateral fusion """
def __init__(self, input_dim=3, output_dim=3):
super(FusionConvLat, self).__init__(input_dim=input_dim)
self.conv = nn.Sequential(
nn.ReLU(True),
nn.Conv2d(input_dim, output_dim, kernel_size=1, bias=False)
)
def forward(self, x1, x2, x2_mask=None, x2_proj=None):
if x1.shape != x2.shape and len(x1.shape) != len(x2.shape):
x2 = self.tile_x2(x1, x2, x2_proj)
x = torch.cat([x1, x2], dim=1) # [B, input_dim, H, W]
x = self.conv(x) # [B, output_dim, H, W]
return x
## ------------- NOTE ----------------
## The following are various fusion types I experimented with.
## Most of them didn't work well ¯\_(ツ)_/¯
## But it doesn't mean there isn't a better way of
## doing lateral and multi-modal (language+vision) fusion.
class FusionFiLM(Fusion):
""" FiLM (Perez et. al, https://arxiv.org/abs/1709.07871).
Note: This is not used inside a Residual block before ReLU.
I had a version this in UpBlock with FiLM, which didn't seem to work at all.
"""
def __init__(self, input_dim=3, output_dim=3):
super(FusionFiLM, self).__init__(input_dim=input_dim)
def forward(self, x1, x2, gamma, beta):
g = self.tile_x2(x1, x2, gamma)
b = self.tile_x2(x1, x2, beta)
return x1 * g + b
class FusionDeepConv(Fusion):
""" Multi-Layer 1x1 convs after [x1; x2] """
def __init__(self, input_dim=3):
super(FusionDeepConv, self).__init__(input_dim=input_dim)
self.conv = nn.Sequential(
nn.ReLU(True),
nn.Conv2d(input_dim * 2, input_dim, kernel_size=1, bias=False),
nn.ReLU(True),
nn.Conv2d(input_dim, input_dim, kernel_size=1, bias=False),
nn.ReLU(True),
nn.Conv2d(input_dim, input_dim, kernel_size=1, bias=False),
)
def forward(self, x1, x2, x2_mask=None, x2_proj=None):
if x1.shape != x2.shape and len(x1.shape) != len(x2.shape):
x2 = self.tile_x2(x1, x2, x2_proj)
x = torch.cat([x1, x2], dim=1) # [B, 2C, H, W]
x = self.conv(x) # [B, C, H, W]
return x
class FusionMultWord(nn.Module):
""" Product with weighted-sum of words """
def __init__(self, input_dim=3):
super().__init__()
self.input_dim = input_dim
def forward(self, x1, x2, x2_mask=None, x2_proj=None):
B, D, H, W = x1.shape
x2_len = int(x2_mask.count_nonzero())
weighted_x1 = torch.zeros_like(x1)
for t in range(x2_len):
x2_t = x2_proj(x2[:,t]) if x2_proj else x2[:,t]
x2_t = x2_t.unsqueeze(-1).unsqueeze(-1).repeat(B, 1, H, W)
weighted_x1 += x1 * x2_t
weighted_x1 /= x2_len
return weighted_x1
class FusionWordAttention(nn.Module):
""" Word Attention """
def __init__(self, input_dim=3):
super().__init__()
self.input_dim = input_dim
self.dot_attn = DotAttn()
def forward(self, x1, x2, x2_mask=None, x2_proj=None):
B, D, H, W = x1.shape
x1_flat = x1.reshape(B, D, H*W)
x2_len = int(x2_mask.count_nonzero())
# TODO: batch this unrolling?
weight_sum_x1_flat = torch.zeros_like(x1_flat)
for t in range(x2_len):
x2_t = x2_proj(x2[:,t]) if x2_proj else x2[:,t]
x2_t = x2_t.repeat(B, 1)
_, attn_x1 = self.dot_attn(x1_flat.transpose(1, 2), x2_t)
weight_sum_x1_flat += x1_flat * attn_x1.transpose(1, 2)
weight_sum_x1_flat /= x2_len
x2 = weight_sum_x1_flat.reshape(B, D, H, W)
return x2
class FusionSentenceAttention(nn.Module):
""" Sentence Attention """
def __init__(self, input_dim=3):
super().__init__()
self.input_dim = input_dim
self.dot_attn = ScaledDotAttn()
def forward(self, x1, x2, x2_mask=None, x2_proj=None):
B, D, H, W = x1.shape
x1_flat = x1.reshape(B, D, H*W)
x2_t = x2_proj(x2) if x2_proj else x2
x2_t = x2_t.repeat(B, 1)
_, attn_x1 = self.dot_attn(x1_flat.transpose(1, 2), x2_t)
weight_sum_x1_flat = x1_flat * attn_x1.transpose(1, 2)
x2 = weight_sum_x1_flat.reshape(B, D, H, W)
return x2
class CrossModalAttention2d(nn.Module):
""" Cross-Modal Attention. Adapted from: https://github.com/openai/CLIP/blob/main/clip/model.py#L56 """
def __init__(self, spacial_dim=7, embed_dim=1024, num_heads=32,
output_dim=1024, lang_dim=512, lang_max_tokens=77):
super().__init__()
self.embed_dim = embed_dim
self.lang_dim = lang_dim
self.lang_max_tokens = lang_max_tokens
self.num_heads = num_heads
self.lang_proj = nn.Linear(self.lang_dim, embed_dim)
self.vision_positional_embedding = nn.Parameter(torch.randn(spacial_dim ** 2, embed_dim) / embed_dim ** 0.5)
self.lang_positional_embedding = nn.Parameter(torch.randn(lang_max_tokens, embed_dim) / embed_dim ** 0.5)
self.k_proj = nn.Linear(embed_dim, embed_dim)
self.q_proj = nn.Linear(embed_dim, embed_dim)
self.v_proj = nn.Linear(embed_dim, embed_dim)
self.c_proj = nn.Linear(embed_dim, output_dim or embed_dim)
def forward(self, x, l, l_mask):
# reshape vision features
x_shape = x.shape
x = x.reshape(x.shape[0], x.shape[1], x.shape[2] * x.shape[3]).permute(2, 0, 1) # NCHW -> (HW)NC
x = x + self.vision_positional_embedding[:x.shape[0], None, :].to(x.dtype) # (HW)NC
# project language
l = l.permute(1, 0, 2)
l_shape = l.shape
l = l.reshape(-1, self.lang_dim)
l = self.lang_proj(l)
l = l.reshape(l_shape[0], l_shape[1], self.embed_dim)
l = l + self.lang_positional_embedding[:, None, :].to(l.dtype)
# hard language mask
l_len = int(l_mask.count_nonzero())
l = l[:l_len]
l = l.repeat(1, x.shape[1], 1)
x, _ = F.multi_head_attention_forward(
query=x, key=l, value=l,
embed_dim_to_check=x.shape[-1],
num_heads=self.num_heads,
q_proj_weight=self.q_proj.weight,
k_proj_weight=self.k_proj.weight,
v_proj_weight=self.v_proj.weight,
in_proj_weight=None,
in_proj_bias=torch.cat([self.q_proj.bias, self.k_proj.bias, self.v_proj.bias]),
bias_k=None,
bias_v=None,
add_zero_attn=False,
dropout_p=0,
out_proj_weight=self.c_proj.weight,
out_proj_bias=self.c_proj.bias,
use_separate_proj_weight=True,
training=self.training,
need_weights=False
)
x = x.permute(1, 2, 0)
x = x.reshape(x_shape)
return x
class FusionMultiHeadedWordAttention(nn.Module):
""" Multi-Headed Word Attention that uses Cross Modal Attention at different scales """
def __init__(self, input_dim=3):
super().__init__()
self.input_dim = input_dim
self.attn1 = CrossModalAttention2d(spacial_dim=7, embed_dim=1024, output_dim=1024)
self.attn2 = CrossModalAttention2d(spacial_dim=14, embed_dim=512, output_dim=512)
self.attn3 = CrossModalAttention2d(spacial_dim=28, embed_dim=256, output_dim=256)
self.multi_headed_attns = {
1024: self.attn1,
512: self.attn2,
256: self.attn3,
}
def forward(self, x1, x2, x2_mask=None, x2_proj=None):
emb_dim = x1.shape[1]
x = self.multi_headed_attns[emb_dim](x1, x2, x2_mask)
return x
names = {
'add': FusionAdd,
'mult': FusionMult,
'mult_word': FusionMultWord,
'film': FusionFiLM,
'max': FusionMax,
'concat': FusionConcat,
'conv': FusionConv,
'deep_conv': FusionDeepConv,
'word_attn': FusionWordAttention,
'sent_attn': FusionSentenceAttention,
'multi_headed_word_attn': FusionMultiHeadedWordAttention,
}
|