jadechoghari
commited on
Commit
•
c56ba76
1
Parent(s):
c577b6d
Create transformer.py
Browse files- transformer.py +135 -0
transformer.py
ADDED
@@ -0,0 +1,135 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Copyright (c) Meta Platforms, Inc. and affiliates.
|
2 |
+
# All rights reserved.
|
3 |
+
#
|
4 |
+
# This source code is licensed under the license found in the
|
5 |
+
# LICENSE file in the root directory of this source tree.
|
6 |
+
|
7 |
+
|
8 |
+
import torch
|
9 |
+
import torch.nn as nn
|
10 |
+
|
11 |
+
|
12 |
+
class ModLN(nn.Module):
|
13 |
+
"""
|
14 |
+
Modulation with adaLN.
|
15 |
+
|
16 |
+
References:
|
17 |
+
DiT: https://github.com/facebookresearch/DiT/blob/main/models.py#L101
|
18 |
+
"""
|
19 |
+
def __init__(self, inner_dim: int, mod_dim: int, eps: float):
|
20 |
+
super().__init__()
|
21 |
+
self.norm = nn.LayerNorm(inner_dim, eps=eps)
|
22 |
+
self.mlp = nn.Sequential(
|
23 |
+
nn.SiLU(),
|
24 |
+
nn.Linear(mod_dim, inner_dim * 2),
|
25 |
+
)
|
26 |
+
|
27 |
+
@staticmethod
|
28 |
+
def modulate(x, shift, scale):
|
29 |
+
# x: [N, L, D]
|
30 |
+
# shift, scale: [N, D]
|
31 |
+
return x * (1 + scale.unsqueeze(1)) + shift.unsqueeze(1)
|
32 |
+
|
33 |
+
def forward(self, x, cond):
|
34 |
+
shift, scale = self.mlp(cond).chunk(2, dim=-1) # [N, D]
|
35 |
+
return self.modulate(self.norm(x), shift, scale) # [N, L, D]
|
36 |
+
|
37 |
+
|
38 |
+
class ConditionModulationBlock(nn.Module):
|
39 |
+
"""
|
40 |
+
Transformer block that takes in a cross-attention condition and another modulation vector applied to sub-blocks.
|
41 |
+
"""
|
42 |
+
# use attention from torch.nn.MultiHeadAttention
|
43 |
+
# Block contains a cross-attention layer, a self-attention layer, and a MLP
|
44 |
+
def __init__(self, inner_dim: int, cond_dim: int, mod_dim: int, num_heads: int, eps: float,
|
45 |
+
attn_drop: float = 0., attn_bias: bool = False,
|
46 |
+
mlp_ratio: float = 4., mlp_drop: float = 0.):
|
47 |
+
super().__init__()
|
48 |
+
self.norm1 = ModLN(inner_dim, mod_dim, eps)
|
49 |
+
self.cross_attn = nn.MultiheadAttention(
|
50 |
+
embed_dim=inner_dim, num_heads=num_heads, kdim=cond_dim, vdim=cond_dim,
|
51 |
+
dropout=attn_drop, bias=attn_bias, batch_first=True)
|
52 |
+
self.norm2 = ModLN(inner_dim, mod_dim, eps)
|
53 |
+
self.self_attn = nn.MultiheadAttention(
|
54 |
+
embed_dim=inner_dim, num_heads=num_heads,
|
55 |
+
dropout=attn_drop, bias=attn_bias, batch_first=True)
|
56 |
+
self.norm3 = ModLN(inner_dim, mod_dim, eps)
|
57 |
+
self.mlp = nn.Sequential(
|
58 |
+
nn.Linear(inner_dim, int(inner_dim * mlp_ratio)),
|
59 |
+
nn.GELU(),
|
60 |
+
nn.Dropout(mlp_drop),
|
61 |
+
nn.Linear(int(inner_dim * mlp_ratio), inner_dim),
|
62 |
+
nn.Dropout(mlp_drop),
|
63 |
+
)
|
64 |
+
|
65 |
+
def forward(self, x, cond, mod):
|
66 |
+
# x: [N, L, D]
|
67 |
+
# cond: [N, L_cond, D_cond]
|
68 |
+
# mod: [N, D_mod]
|
69 |
+
x = x + self.cross_attn(self.norm1(x, mod), cond, cond, need_weights=False)[0]
|
70 |
+
before_sa = self.norm2(x, mod)
|
71 |
+
x = x + self.self_attn(before_sa, before_sa, before_sa, need_weights=False)[0]
|
72 |
+
x = x + self.mlp(self.norm3(x, mod))
|
73 |
+
return x
|
74 |
+
|
75 |
+
|
76 |
+
class TriplaneTransformer(nn.Module):
|
77 |
+
"""
|
78 |
+
Transformer with condition and modulation that generates a triplane representation.
|
79 |
+
|
80 |
+
Reference:
|
81 |
+
Timm: https://github.com/huggingface/pytorch-image-models/blob/main/timm/models/vision_transformer.py#L486
|
82 |
+
"""
|
83 |
+
def __init__(self, inner_dim: int, image_feat_dim: int, camera_embed_dim: int,
|
84 |
+
triplane_low_res: int, triplane_high_res: int, triplane_dim: int,
|
85 |
+
num_layers: int, num_heads: int,
|
86 |
+
eps: float = 1e-6):
|
87 |
+
super().__init__()
|
88 |
+
|
89 |
+
# attributes
|
90 |
+
self.triplane_low_res = triplane_low_res
|
91 |
+
self.triplane_high_res = triplane_high_res
|
92 |
+
self.triplane_dim = triplane_dim
|
93 |
+
|
94 |
+
# modules
|
95 |
+
# initialize pos_embed with 1/sqrt(dim) * N(0, 1)
|
96 |
+
self.pos_embed = nn.Parameter(torch.randn(1, 3*triplane_low_res**2, inner_dim) * (1. / inner_dim) ** 0.5)
|
97 |
+
self.layers = nn.ModuleList([
|
98 |
+
ConditionModulationBlock(
|
99 |
+
inner_dim=inner_dim, cond_dim=image_feat_dim, mod_dim=camera_embed_dim, num_heads=num_heads, eps=eps)
|
100 |
+
for _ in range(num_layers)
|
101 |
+
])
|
102 |
+
self.norm = nn.LayerNorm(inner_dim, eps=eps)
|
103 |
+
self.deconv = nn.ConvTranspose2d(inner_dim, triplane_dim, kernel_size=2, stride=2, padding=0)
|
104 |
+
|
105 |
+
def forward(self, image_feats, camera_embeddings):
|
106 |
+
# image_feats: [N, L_cond, D_cond]
|
107 |
+
# camera_embeddings: [N, D_mod]
|
108 |
+
|
109 |
+
assert image_feats.shape[0] == camera_embeddings.shape[0], \
|
110 |
+
f"Mismatched batch size: {image_feats.shape[0]} vs {camera_embeddings.shape[0]}"
|
111 |
+
|
112 |
+
N = image_feats.shape[0]
|
113 |
+
H = W = self.triplane_low_res
|
114 |
+
L = 3 * H * W
|
115 |
+
|
116 |
+
x = self.pos_embed.repeat(N, 1, 1) # [N, L, D]
|
117 |
+
for layer in self.layers:
|
118 |
+
x = layer(x, image_feats, camera_embeddings)
|
119 |
+
x = self.norm(x)
|
120 |
+
|
121 |
+
# separate each plane and apply deconv
|
122 |
+
x = x.view(N, 3, H, W, -1)
|
123 |
+
x = torch.einsum('nihwd->indhw', x) # [3, N, D, H, W]
|
124 |
+
x = x.contiguous().view(3*N, -1, H, W) # [3*N, D, H, W]
|
125 |
+
x = self.deconv(x) # [3*N, D', H', W']
|
126 |
+
x = x.view(3, N, *x.shape[-3:]) # [3, N, D', H', W']
|
127 |
+
x = torch.einsum('indhw->nidhw', x) # [N, 3, D', H', W']
|
128 |
+
x = x.contiguous()
|
129 |
+
|
130 |
+
assert self.triplane_high_res == x.shape[-2], \
|
131 |
+
f"Output triplane resolution does not match with expected: {x.shape[-2]} vs {self.triplane_high_res}"
|
132 |
+
assert self.triplane_dim == x.shape[-3], \
|
133 |
+
f"Output triplane dimension does not match with expected: {x.shape[-3]} vs {self.triplane_dim}"
|
134 |
+
|
135 |
+
return x
|