File size: 9,747 Bytes
9bc4638 |
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 |
# 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.
import math
import torch
import torch.nn as nn
import torch.nn.functional as F
def FeedForward(dim, mult=4):
inner_dim = int(dim * mult)
return nn.Sequential(
nn.LayerNorm(dim),
nn.Linear(dim, inner_dim, bias=False),
nn.GELU(),
nn.Linear(inner_dim, dim, bias=False),
)
class PerceiverAttention(nn.Module):
def __init__(
self, *, dim, dim_head=64, heads=8, dropout_p=0.05, concat_kv_latents=True
):
super().__init__()
self.scale = dim_head**-0.5
self.heads = heads
inner_dim = dim_head * heads
self.norm_x = nn.LayerNorm(dim)
self.norm_latents = nn.LayerNorm(dim)
self.to_q = nn.Linear(dim, inner_dim, bias=False)
self.to_kv = nn.Linear(dim, inner_dim * 2, bias=False)
self.to_out = nn.Linear(inner_dim, dim, bias=False)
self.dropout_p = dropout_p
self.concat_kv_latents = concat_kv_latents
def _separate_heads(self, x: torch.Tensor, num_heads: int) -> torch.Tensor:
b, n, c = x.shape
x = x.reshape(b, n, num_heads, c // num_heads)
return x.transpose(1, 2) # B x N_heads x N_tokens x C_per_head
def _recombine_heads(self, x: torch.Tensor) -> torch.Tensor:
b, n_heads, n_tokens, c_per_head = x.shape
x = x.transpose(1, 2)
return x.reshape(b, n_tokens, n_heads * c_per_head) # B x N_tokens x C
def forward(self, latents, x, pos=None):
latents = self.norm_latents(latents)
x = self.norm_x(x)
q = self.to_q(latents)
# the paper differs from Perceiver in which they also concat the key / values derived from the latents to be attended to
if self.concat_kv_latents:
kv_input = torch.cat((x, latents), dim=-2)
else:
kv_input = x
k, v = self.to_kv(kv_input).chunk(2, dim=-1)
q = self._separate_heads(q, self.heads)
k = self._separate_heads(k, self.heads)
v = self._separate_heads(v, self.heads)
if pos is not None:
assert not self.concat_kv_latents
pos = self._separate_heads(pos, self.heads)
k, v = k + pos, v + pos
out = F.scaled_dot_product_attention(
q,
k,
v,
attn_mask=None,
dropout_p=self.dropout_p if self.training else 0.0,
)
out = self._recombine_heads(out)
return self.to_out(out)
class Attention(nn.Module):
def __init__(self, *, dim, dim_head=64, heads=8, dropout_p=0.05):
super().__init__()
self.scale = dim_head**-0.5
self.heads = heads
inner_dim = dim_head * heads
self.norm = nn.LayerNorm(dim)
self.to_q = nn.Linear(dim, inner_dim, bias=False)
self.to_kv = nn.Linear(dim, inner_dim * 2, bias=False)
self.to_out = nn.Linear(inner_dim, dim, bias=False)
self.dropout_p = dropout_p
def _separate_heads(self, x: torch.Tensor, num_heads: int) -> torch.Tensor:
b, n, c = x.shape
x = x.reshape(b, n, num_heads, c // num_heads)
return x.transpose(1, 2) # B x N_heads x N_tokens x C_per_head
def _recombine_heads(self, x: torch.Tensor) -> torch.Tensor:
b, n_heads, n_tokens, c_per_head = x.shape
x = x.transpose(1, 2)
return x.reshape(b, n_tokens, n_heads * c_per_head) # B x N_tokens x C
def forward(self, x):
x = self.norm(x)
q = self.to_q(x)
k, v = self.to_kv(x).chunk(2, dim=-1)
q = self._separate_heads(q, self.heads)
k = self._separate_heads(k, self.heads)
v = self._separate_heads(v, self.heads)
out = F.scaled_dot_product_attention(
q,
k,
v,
attn_mask=None,
dropout_p=self.dropout_p if self.training else 0.0,
)
out = self._recombine_heads(out)
return self.to_out(out)
class PerceiverEncoderLayer(nn.Module):
def __init__(
self,
dim,
dim_head=64,
heads=8,
ff_mult=4,
hidden_dropout_p=0.0,
attention_dropout_p=0.0,
concat_kv_latents=False,
use_self_attn=False,
):
super().__init__()
self.attn = PerceiverAttention(
dim=dim,
dim_head=dim_head,
heads=heads,
dropout_p=attention_dropout_p,
concat_kv_latents=concat_kv_latents,
)
self.ff = FeedForward(dim=dim, mult=ff_mult)
self.dropout = nn.Dropout(hidden_dropout_p)
self.use_self_attn = use_self_attn
if use_self_attn:
self.self_attn = Attention(
dim=dim,
dim_head=dim_head,
heads=heads,
dropout_p=attention_dropout_p,
)
self.self_ff = FeedForward(dim=dim, mult=ff_mult)
def forward(self, latents, x, pos=None):
latents = self.attn(latents, x, pos) + latents
latents = self.dropout(latents)
latents = self.ff(latents) + latents
if self.use_self_attn:
latents = self.self_attn(latents) + latents
latents = self.self_ff(latents) + latents
return latents
def window_partition(x, window_size):
"""
Args:
x: (B, H, W, C)
window_size (int): window size
Returns:
windows: (num_windows*B, window_size, window_size, C)
"""
B, H, W, C = x.shape
x = x.view(B, H // window_size, window_size, W // window_size, window_size, C)
windows = (
x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size, window_size, C)
)
return windows
def window_reverse(windows, window_size, H, W):
"""
Args:
windows: (num_windows*B, window_size, window_size, C)
window_size (int): Window size
H (int): Height of image
W (int): Width of image
Returns:
x: (B, H, W, C)
"""
B = int(windows.shape[0] / (H * W / window_size / window_size))
x = windows.view(
B, H // window_size, W // window_size, window_size, window_size, -1
)
x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(B, H, W, -1)
return x
class PerceiverResampler(nn.Module):
def __init__(
self,
*,
dim,
depth,
dim_head=64,
heads=1,
num_latents=-1,
num_latents_2d=-1,
ff_mult=4,
hidden_dropout_p=0.1,
attention_dropout_p=0.05,
pos_enc_at_key_value=False,
concat_kv_latents=False,
position_encoding=None,
use_self_attn=False,
**kwargs,
):
super().__init__()
self.num_latents = num_latents
self.num_latents_2d = num_latents_2d
if num_latents > 0:
self.latents = nn.Parameter(torch.randn(num_latents, dim))
if num_latents_2d > 0:
self.latents_2d = nn.Parameter(torch.randn(num_latents_2d, dim))
self.position_encoding = position_encoding
self.layers = nn.ModuleList([])
for _ in range(depth):
self.layers.append(
PerceiverEncoderLayer(
dim=dim,
dim_head=dim_head,
heads=heads,
ff_mult=ff_mult,
hidden_dropout_p=hidden_dropout_p,
attention_dropout_p=attention_dropout_p,
concat_kv_latents=concat_kv_latents,
use_self_attn=use_self_attn,
),
)
self.norm = nn.LayerNorm(dim)
self.pos_enc_at_key_value = pos_enc_at_key_value
def forward(self, x, pos=None):
out_latents = []
out_pos = []
if self.num_latents > 0:
latents_1d, pos_1d = self.forward_1d(x, pos)
out_latents.append(latents_1d)
out_pos.append(pos_1d)
if self.num_latents_2d > 0:
latents_2d, pos_2d = self.forward_2d(x)
out_latents.append(latents_2d)
out_pos.append(pos_2d)
latents = torch.concat(out_latents, dim=1)
if pos is not None:
pos = torch.concat(out_pos, dim=1)
return latents, pos
def forward_1d(self, x, pos):
latents = self.latents.unsqueeze(0).expand(x.shape[0], -1, -1)
x = x.permute(0, 2, 3, 1).flatten(1, 2)
if not self.pos_enc_at_key_value:
_pos = None
if pos is not None:
_pos = pos.permute(0, 2, 3, 1).flatten(1, 2)
else:
_pos = None
for layer in self.layers:
latents = layer(latents, x, _pos)
if pos is not None:
pos = torch.zeros_like(latents)
latents = self.norm(latents)
return latents, pos
def forward_2d(self, x):
B, C, H, W = x.shape
latents_2d = self.latents_2d.unsqueeze(0).expand(B, -1, -1).view(-1, 1, C)
num_window = int(math.sqrt(self.num_latents_2d))
window_size = H // num_window
x = x.permute(0, 2, 3, 1)
x = window_partition(x, window_size)
x = x.flatten(1, 2)
for layer in self.layers:
latents_2d = layer(latents_2d, x)
latents_2d = latents_2d.view(B, num_window, num_window, C).permute(0, 3, 1, 2)
pos_2d = self.position_encoding(latents_2d)
pos_2d = pos_2d.permute(0, 2, 3, 1).flatten(1, 2)
latents_2d = latents_2d.permute(0, 2, 3, 1).flatten(1, 2)
latents_2d = self.norm(latents_2d)
return latents_2d, pos_2d
|