-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathmodel.py
243 lines (199 loc) · 8.49 KB
/
model.py
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
import torch
from torch import nn
import torch.nn.functional as F
from einops import rearrange, repeat
from einops.layers.torch import Rearrange
from utils import pair
def posemb_sincos_3d(f, h, w, dim, device='cuda', temperature = 10000, dtype = torch.float32):
z, y, x = torch.meshgrid(
torch.arange(f, device = device),
torch.arange(h, device = device),
torch.arange(w, device = device),
indexing = 'ij')
fourier_dim = dim // 6
omega = torch.arange(fourier_dim, device = device) / (fourier_dim - 1)
omega = 1. / (temperature ** omega)
z = z.flatten()[:, None] * omega[None, :]
y = y.flatten()[:, None] * omega[None, :]
x = x.flatten()[:, None] * omega[None, :]
pe = torch.cat((x.sin(), x.cos(), y.sin(), y.cos(), z.sin(), z.cos()), dim = 1)
pe = F.pad(pe, (0, dim - (fourier_dim * 6))) # pad if feature dimension not cleanly divisible by 6
return pe.type(dtype)
class PreNorm(nn.Module):
def __init__(self, dim, fn):
super().__init__()
self.norm = nn.LayerNorm(dim)
self.fn = fn
def forward(self, x, **kwargs):
return self.fn(self.norm(x), **kwargs)
class XPreNorm(nn.Module):
def __init__(self, dim, fn):
super().__init__()
self.norm = nn.LayerNorm(dim)
self.fn = fn
def forward(self, x, target, **kwargs):
return self.fn(self.norm(x), self.norm(target), **kwargs)
class FeedForward(nn.Module):
def __init__(self, dim, hidden_dim, dropout = 0.):
super().__init__()
self.net = nn.Sequential(
nn.Linear(dim, hidden_dim),
nn.GELU(),
nn.Dropout(dropout),
nn.Linear(hidden_dim, dim),
nn.Dropout(dropout)
)
def forward(self, x):
return self.net(x)
class Attention(nn.Module):
def __init__(self, dim, heads = 8, dim_head = 64, dropout = 0.):
super().__init__()
inner_dim = dim_head * heads
project_out = not (heads == 1 and dim_head == dim)
self.heads = heads
self.scale = dim_head ** -0.5
self.attend = nn.Softmax(dim = -1)
self.dropout = nn.Dropout(dropout)
self.to_qkv = nn.Linear(dim, inner_dim * 3, bias = False)
self.to_out = nn.Sequential(
nn.Linear(inner_dim, dim),
nn.Dropout(dropout)
) if project_out else nn.Identity()
def forward(self, x):
qkv = self.to_qkv(x).chunk(3, dim = -1)
q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> b h n d', h = self.heads), qkv)
dots = torch.matmul(q, k.transpose(-1, -2)) * self.scale
attn = self.attend(dots)
attn = self.dropout(attn)
out = torch.matmul(attn, v)
out = rearrange(out, 'b h n d -> b n (h d)')
return self.to_out(out)
class CrossAttention(nn.Module):
def __init__(self, dim, heads = 8, dim_head = 64, dropout = 0.):
super().__init__()
inner_dim = dim_head * heads
project_out = not (heads == 1 and dim_head == dim)
self.heads = heads
self.scale = dim_head ** -0.5
self.attend = nn.Softmax(dim = -1)
self.dropout = nn.Dropout(dropout)
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.Sequential(
nn.Linear(inner_dim, dim),
nn.Dropout(dropout)
) if project_out else nn.Identity()
def forward(self, x, target):
q = self.to_q(x)#.chunk(3, dim = -1)
k, v = self.to_kv(target).chunk(2, dim=-1)
qkv = q, k, v
q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> b h n d', h = self.heads), qkv)
dots = torch.matmul(q, k.transpose(-1, -2)) * self.scale
attn = self.attend(dots)
attn = self.dropout(attn)
out = torch.matmul(attn, v)
out = rearrange(out, 'b h n d -> b n (h d)')
return self.to_out(out)
class Transformer(nn.Module):
def __init__(self, dim, depth, heads, dim_head, mlp_dim, dropout = 0.):
super().__init__()
self.layers = nn.ModuleList([])
for _ in range(depth):
self.layers.append(nn.ModuleList([
PreNorm(dim, Attention(dim, heads = heads, dim_head = dim_head, dropout = dropout)),
PreNorm(dim, FeedForward(dim, mlp_dim, dropout = dropout))
]))
def forward(self, x):
for attn, ff in self.layers:
x = attn(x) + x
x = ff(x) + x
return x
class TransformerDecoder(nn.Module):
def __init__(self, dim, depth, heads, dim_head, mlp_dim, dropout = 0.):
super().__init__()
self.layers = nn.ModuleList([])
for _ in range(depth):
self.layers.append(nn.ModuleList([
XPreNorm(dim, CrossAttention(dim, heads = heads, dim_head = dim_head, dropout = dropout)),
PreNorm(dim, FeedForward(dim, mlp_dim, dropout = dropout))
]))
def forward(self, x, target):
for xattn, ff in self.layers:
x = xattn(x,target) + x
x = ff(x) + x
return x
class Made2Order(nn.Module):
def __init__(
self,
image_size,
image_patch_size,
frames,
channels = 3,
frame_patch_size = 1,
decoder_dim = 64,
dim = 256,
heads = 4,
mlp_dim = 512,
dim_head = 64,
dropout = 0.,
):
super().__init__()
image_height, image_width = pair(image_size)
self.patch_size = image_patch_size
patch_height, patch_width = pair(image_patch_size)
assert image_height % patch_height == 0 and image_width % patch_width == 0, 'Image dimensions must be divisible by the patch size.'
assert frames % frame_patch_size == 0, 'Frames must be divisible by frame patch size'
num_image_patches = (image_height // patch_height) * (image_width // patch_width)
num_frame_patches = (frames // frame_patch_size)
patch_dim = channels * patch_height * patch_width * frame_patch_size
self.to_patch_embedding = nn.Sequential(
Rearrange('b c (f pf) (h p1) (w p2) -> b f h w (p1 p2 pf c)', p1 = patch_height, p2 = patch_width, pf = frame_patch_size),
nn.LayerNorm(patch_dim),
nn.Linear(patch_dim, dim),
nn.LayerNorm(dim)
)
self.pos_embedding = posemb_sincos_3d(frames, image_height // patch_height,image_width // patch_width,dim)
self.pos_embedding = rearrange(self.pos_embedding, '(f hw) d -> 1 f hw d', f=frames)
self.T1 = Transformer(dim, 1, heads, dim_head, mlp_dim, dropout)
self.T2 = Transformer(dim, 1, heads, dim_head, mlp_dim, dropout)
self.T3 = Transformer(dim, 1, heads, dim_head, mlp_dim, dropout)
self.T4 = Transformer(dim, 1, heads, dim_head, mlp_dim, dropout)
self.T5 = Transformer(dim, 1, heads, dim_head, mlp_dim, dropout)
self.T6 = Transformer(dim, 1, heads, dim_head, mlp_dim, dropout)
self.projection_head = nn.Sequential(
nn.LayerNorm(dim),
nn.Linear(dim, decoder_dim)
)
self.query = torch.nn.Parameter(torch.randn(1, frames, decoder_dim).cuda())
self.decoder = TransformerDecoder(decoder_dim, 3, heads, dim_head, mlp_dim, dropout)
def forward(self, video):
video = rearrange(video, 'b f c h w -> b c f h w')
#patch+pe
x = self.to_patch_embedding(video)
b, f, h, w, d = x.shape
x = rearrange(x, 'b f h w d -> b f (h w) d') + self.pos_embedding[:, :f]
#TE
x = rearrange(x, 'b f n d -> (b f) n d')
x = self.T1(x)
x = rearrange(x, '(b f) n d -> (b n) f d', b = b)
x = self.T2(x)
x = rearrange(x, '(b f) n d -> (b n) f d', b = b)
x = self.T3(x)
x = rearrange(x, '(b f) n d -> (b n) f d', b = b)
x = self.T4(x)
x = rearrange(x, '(b f) n d -> (b n) f d', b = b)
x = self.T5(x)
x = rearrange(x, '(b f) n d -> (b n) f d', b = b)
x = self.T6(x)
x = rearrange(x, '(b n) f d -> b (f n) d', b = b)
te_output = self.projection_head(x)
#TD
query = self.query.repeat(b,1,1)
td_output = self.decoder(query, te_output)
#similarity
te_output = F.normalize(te_output, dim=2)
td_output = F.normalize(td_output, dim=2)
output = torch.einsum('bik,bjk->bij', te_output, td_output)
attn = rearrange(output, 'b (f n) d -> b f n d', f=f)
output = torch.max(attn, 2)[0] /0.2
return output, torch.clip(attn,0,1)