Spaces:
Running
on
Zero
Running
on
Zero
File size: 5,065 Bytes
c62903f |
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 |
"""
Taken from https://github.com/lucidrains/flamingo-pytorch
"""
import torch
from einops import rearrange, repeat
try:
from einops_exts import rearrange_many
except:
pass
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 PerceiverResamplerModule(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)
if ff_mult > 0
else nn.Identity(),
]
)
)
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)
class PerceiverResampler(nn.Module):
def __init__(self, model_args, vision_tower):
super().__init__()
self.depth = model_args.mm_perceiver_depth
self.num_latents = model_args.mm_perceiver_latents
self.ff_mult = model_args.mm_perceiver_ff_mult
self.pretrained = model_args.mm_perceiver_pretrained
self.perceiver = PerceiverResamplerModule(
dim=vision_tower.hidden_size,
depth=self.depth,
num_latents=self.num_latents,
ff_mult=self.ff_mult,
)
if self.pretrained is not None:
self.load_state_dict(torch.load(self.pretrained))
def forward(self, image_features, *args, **kwargs):
return self.perceiver(image_features[:, None, None]).squeeze(1)
@property
def config(self):
return {
"mm_resampler_type": "perceiver",
"mm_perceiver_depth": self.depth,
"mm_perceiver_latents": self.num_latents,
"mm_perceiver_ff_mult": self.ff_mult,
"mm_perceiver_pretrained": self.pretrained,
}
|