File size: 8,743 Bytes
e1aaaac |
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 |
"""
Based on: https://github.com/lucidrains/flamingo-pytorch
"""
import torch
from einops import rearrange, repeat
from einops_exts import rearrange_many
from torch import einsum, nn
def exists(val):
return val is not None
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):
super().__init__()
self.scale = dim_head**-0.5
self.heads = heads
inner_dim = dim_head * heads
self.norm_media = 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)
def forward(self, x, latents):
"""
Args:
x (torch.Tensor): image features
shape (b, T, n1, D)
latent (torch.Tensor): latent features
shape (b, T, n2, D)
"""
x = self.norm_media(x)
latents = self.norm_latents(latents)
h = self.heads
q = self.to_q(latents)
kv_input = torch.cat((x, latents), dim=-2)
k, v = self.to_kv(kv_input).chunk(2, dim=-1)
q, k, v = rearrange_many((q, k, v), "b t n (h d) -> b h t n d", h=h)
q = q * self.scale
# attention
sim = einsum("... i d, ... j d -> ... i j", q, k)
sim = sim - sim.amax(dim=-1, keepdim=True).detach()
attn = sim.softmax(dim=-1)
out = einsum("... i j, ... j d -> ... i d", attn, v)
out = rearrange(out, "b h t n d -> b t n (h d)", h=h)
return self.to_out(out)
class PerceiverResampler(nn.Module):
def __init__(
self,
*,
dim,
depth=6,
dim_head=64,
heads=8,
num_latents=64,
max_num_media=None,
max_num_frames=None,
ff_mult=4,
):
super().__init__()
self.latents = nn.Parameter(torch.randn(num_latents, dim))
self.frame_embs = (
nn.Parameter(torch.randn(max_num_frames, dim))
if exists(max_num_frames)
else None
)
self.media_time_embs = (
nn.Parameter(torch.randn(max_num_media, 1, dim))
if exists(max_num_media)
else None
)
self.layers = nn.ModuleList([])
for _ in range(depth):
self.layers.append(
nn.ModuleList(
[
PerceiverAttention(dim=dim, dim_head=dim_head, heads=heads),
FeedForward(dim=dim, mult=ff_mult),
]
)
)
self.norm = nn.LayerNorm(dim)
def forward(self, x):
"""
Args:
x (torch.Tensor): image features
shape (b, T, F, v, D)
Returns:
shape (b, T, n, D) where n is self.num_latents
"""
b, T, F, v = x.shape[:4]
# frame and media time embeddings
if exists(self.frame_embs):
frame_embs = repeat(self.frame_embs[:F], "F d -> b T F v d", b=b, T=T, v=v)
x = x + frame_embs
x = rearrange(
x, "b T F v d -> b T (F v) d"
) # flatten the frame and spatial dimensions
if exists(self.media_time_embs):
x = x + self.media_time_embs[:T]
# blocks
latents = repeat(self.latents, "n d -> b T n d", b=b, T=T)
for attn, ff in self.layers:
latents = attn(x, latents) + latents
latents = ff(latents) + latents
return self.norm(latents)
# gated cross attention
class MaskedCrossAttention(nn.Module):
def __init__(
self,
*,
dim,
dim_visual,
dim_head=64,
heads=8,
only_attend_immediate_media=True,
):
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_visual, inner_dim * 2, bias=False)
self.to_out = nn.Linear(inner_dim, dim, bias=False)
# whether for text to only attend to immediate preceding image, or all previous images
self.only_attend_immediate_media = only_attend_immediate_media
def forward(self, x, media, media_locations=None, use_cached_media=False):
"""
Args:
x (torch.Tensor): text features
shape (B, T_txt, D_txt)
media (torch.Tensor): image features
shape (B, T_img, n, D_img) where n is the dim of the latents
media_locations: boolean mask identifying the media tokens in x
shape (B, T_txt)
use_cached_media: bool
If true, treat all of x as if they occur after the last media
registered in media_locations. T_txt does not need to exactly
equal media_locations.shape[1] in this case
"""
if not use_cached_media:
assert (
media_locations.shape[1] == x.shape[1]
), f"media_location.shape is {media_locations.shape} but x.shape is {x.shape}"
T_txt = x.shape[1]
_, T_img, n = media.shape[:3]
h = self.heads
x = self.norm(x)
q = self.to_q(x)
media = rearrange(media, "b t n d -> b (t n) d")
k, v = self.to_kv(media).chunk(2, dim=-1)
q, k, v = rearrange_many((q, k, v), "b n (h d) -> b h n d", h=h)
q = q * self.scale
sim = einsum("... i d, ... j d -> ... i j", q, k)
if exists(media_locations):
media_time = torch.arange(T_img, device=x.device) + 1
if use_cached_media:
# text time is set to the last cached media location
text_time = repeat(
torch.count_nonzero(media_locations, dim=1),
"b -> b i",
i=T_txt,
)
else:
# at each boolean of True, increment the time counter (relative to media time)
text_time = media_locations.cumsum(dim=-1)
# text time must equal media time if only attending to most immediate image
# otherwise, as long as text time is greater than media time (if attending to all previous images / media)
mask_op = torch.eq if self.only_attend_immediate_media else torch.ge
text_to_media_mask = mask_op(
rearrange(text_time, "b i -> b 1 i 1"),
repeat(media_time, "j -> 1 1 1 (j n)", n=n),
)
sim = sim.masked_fill(~text_to_media_mask, -torch.finfo(sim.dtype).max)
sim = sim - sim.amax(dim=-1, keepdim=True).detach()
attn = sim.softmax(dim=-1)
if exists(media_locations) and self.only_attend_immediate_media:
# any text without a preceding media needs to have attention zeroed out
text_without_media_mask = text_time == 0
text_without_media_mask = rearrange(
text_without_media_mask, "b i -> b 1 i 1"
)
attn = attn.masked_fill(text_without_media_mask, 0.0)
out = einsum("... i j, ... j d -> ... i d", attn, v)
out = rearrange(out, "b h n d -> b n (h d)")
return self.to_out(out)
class GatedCrossAttentionBlock(nn.Module):
def __init__(
self,
*,
dim,
dim_visual,
dim_head=64,
heads=8,
ff_mult=4,
only_attend_immediate_media=True,
):
super().__init__()
self.attn = MaskedCrossAttention(
dim=dim,
dim_visual=dim_visual,
dim_head=dim_head,
heads=heads,
only_attend_immediate_media=only_attend_immediate_media,
)
self.attn_gate = nn.Parameter(torch.tensor([0.0]))
self.ff = FeedForward(dim, mult=ff_mult)
self.ff_gate = nn.Parameter(torch.tensor([0.0]))
def forward(
self,
x,
media,
media_locations=None,
use_cached_media=False,
):
x = (
self.attn(
x,
media,
media_locations=media_locations,
use_cached_media=use_cached_media,
)
* self.attn_gate.tanh()
+ x
)
x = self.ff(x) * self.ff_gate.tanh() + x
return x
|