-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathmodel.py
230 lines (197 loc) · 8.54 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
import torch
import torch.nn as nn
import torch.nn.functional as F
import math
import matplotlib.pyplot as plt
from utils import imshow
# Implementation of HiP-16 and HiP-256
class SelfAttentionLayer(nn.Module):
def __init__(self, embed_dim, num_heads, expanding_factor=4):
super().__init__()
self.mha = nn.MultiheadAttention(embed_dim, num_heads)
self.linear1 = nn.Linear(embed_dim, embed_dim * expanding_factor)
self.linear2 = nn.Linear(embed_dim * expanding_factor, embed_dim)
self.norm1 = nn.LayerNorm(embed_dim)
self.norm2 = nn.LayerNorm(embed_dim)
def forward(self, x):
x = self.norm1(self.mha(x, x, x)[0] + x)
x = self.norm2(F.gelu(self.linear2(F.gelu(self.linear1(x)))) + x)
return x
class HiPLayer(nn.Module):
def __init__(self, num_groups, num_attn_layers, num_latents, num_channels,
num_heads, input_channels=None):
super().__init__()
self.num_groups = num_groups
self.num_attn_layers = num_attn_layers
self.num_latents = num_latents
self.num_channels = num_channels
self.num_heads = num_heads
self.dot_scale = 1 / math.sqrt(num_channels)
input_channels = num_channels if input_channels == None else input_channels
self.linear = nn.Linear(input_channels, num_channels)
self.learned_embedding = nn.Parameter(
torch.randn(num_groups, num_latents, num_channels))
self.attention = nn.MultiheadAttention(input_channels, num_heads)
self.stem = nn.ModuleList(
[SelfAttentionLayer(num_channels, num_heads)
for _ in range(num_attn_layers)]
)
def forward(self, x):
"""
Input
-----
x : torch.tensor (batch_size x num_tokens x num_channels)
"""
B, N, C = x.shape
x = x.view(-1, self.num_groups, N//self.num_groups, C)
x = self.linear(x)
# print(x.shape)
attn = torch.einsum(
"gkd,bghd->bgkh", self.learned_embedding, x) * self.dot_scale
attn = torch.softmax(attn, dim=-2)
x = torch.einsum("bgkh,bghd->bgkd", attn, x)
# print(x.shape)
x = x.view(-1, self.num_latents, self.num_channels)
for _, layer in enumerate(self.stem):
x = layer(x)
# print(x.shape)
x = x.view(B, self.num_groups * self.num_latents, self.num_channels)
# print(x.shape)
return x, attn
class HiP16_small(nn.Module):
def __init__(self):
super().__init__()
self.encoder = nn.ModuleList(
[
HiPLayer(num_groups=16, num_attn_layers=2, num_latents=32,
num_channels=64, num_heads=4, input_channels=32),
HiPLayer(num_groups=4, num_attn_layers=2, num_latents=64,
num_channels=128, num_heads=8, input_channels=64),
HiPLayer(num_groups=1, num_attn_layers=18, num_latents=128,
num_channels=256, num_heads=16, input_channels=128),
HiPLayer(num_groups=1, num_attn_layers=2, num_latents=64,
num_channels=512, num_heads=32, input_channels=256),
]
)
self.decoder = nn.ModuleList(
[
HiPLayer(num_groups=1, num_attn_layers=1, num_latents=128,
num_channels=256, num_heads=16, input_channels=512),
HiPLayer(num_groups=4, num_attn_layers=1, num_latents=64,
num_channels=128, num_heads=8, input_channels=256),
HiPLayer(num_groups=16, num_attn_layers=1, num_latents=32,
num_channels=64, num_heads=4, input_channels=128)
]
)
def forward(self, x):
skips = []
for _, layer in enumerate(self.encoder):
skips.append(x)
x = layer(x)[0]
for _, layer in enumerate(self.decoder):
x = layer(x)[0] + skips.pop()
return x
class HiP16(nn.Module):
def __init__(self):
super().__init__()
self.encoder = nn.ModuleList(
[
HiPLayer(num_groups=16, num_attn_layers=2, num_latents=128,
num_channels=128, num_heads=4, input_channels=32),
HiPLayer(num_groups=4, num_attn_layers=2, num_latents=256,
num_channels=256, num_heads=8, input_channels=128),
HiPLayer(num_groups=1, num_attn_layers=18, num_latents=256,
num_channels=512, num_heads=16, input_channels=256),
HiPLayer(num_groups=1, num_attn_layers=2, num_latents=64,
num_channels=1024, num_heads=32, input_channels=512),
]
)
self.decoder = nn.ModuleList(
[
HiPLayer(num_groups=1, num_attn_layers=1, num_latents=256,
num_channels=512, num_heads=16, input_channels=1024),
HiPLayer(num_groups=4, num_attn_layers=1, num_latents=256,
num_channels=256, num_heads=8, input_channels=512),
HiPLayer(num_groups=16, num_attn_layers=1, num_latents=128,
num_channels=128, num_heads=4, input_channels=256)
]
)
def forward(self, x):
skips = []
for _, layer in enumerate(self.encoder):
skips.append(x)
x = layer(x)[0]
for _, layer in enumerate(self.decoder):
x = layer(x)[0] + skips.pop()
return x
class HiP256(nn.Module):
def __init__(self):
super().__init__()
self.encoder = nn.ModuleList(
[
HiPLayer(num_groups=256, num_attn_layers=1, num_latents=32,
num_channels=64, num_heads=1, input_channels=16),
HiPLayer(num_groups=64, num_attn_layers=1, num_latents=64,
num_channels=96, num_heads=2, input_channels=64),
HiPLayer(num_groups=16, num_attn_layers=2, num_latents=128,
num_channels=128, num_heads=4, input_channels=96),
HiPLayer(num_groups=4, num_attn_layers=2, num_latents=256,
num_channels=256, num_heads=8, input_channels=128),
HiPLayer(num_groups=1, num_attn_layers=18, num_latents=256,
num_channels=512, num_heads=16, input_channels=256),
HiPLayer(num_groups=1, num_attn_layers=2, num_latents=64,
num_channels=1024, num_heads=32, input_channels=512),
]
)
self.decoder = nn.ModuleList(
[
HiPLayer(num_groups=1, num_attn_layers=1, num_latents=256,
num_channels=256, num_heads=16, input_channels=1024),
HiPLayer(num_groups=4, num_attn_layers=1, num_latents=256,
num_channels=128, num_heads=8, input_channels=256),
HiPLayer(num_groups=16, num_attn_layers=1, num_latents=128,
num_channels=64, num_heads=4, input_channels=128),
HiPLayer(num_groups=64, num_attn_layers=1, num_latents=64,
num_channels=32, num_heads=2, input_channels=64),
HiPLayer(num_groups=256, num_attn_layers=1, num_latents=32,
num_channels=16, num_heads=1, input_channels=32)
]
)
def forward(self, x):
# Skip connection is not yet implemented for HiP-256 because the shape
# doesn't match according to paper setup.
for _, layer in enumerate(self.encoder):
x = layer(x)[0]
for _, layer in enumerate(self.decoder):
x = layer(x)[0]
return x
class HiPIO(nn.Module):
def __init__(self, preprocess_layer, backbone, postprocess_layer):
super().__init__()
self.preprocess_layer = preprocess_layer
self.backbone = backbone
self.postprocess_layer = postprocess_layer
def forward(self, x):
x = self.preprocess_layer(x)
x = self.backbone(x)
x = self.postprocess_layer(x)
return x
if __name__ == "__main__":
hip16s = HiP16_small()
hip16 = HiP16()
hip256 = HiP256()
x = torch.randn(1, 1024, 32)
y = hip16s(x)
print(y.shape)
x = torch.randn(1, 1024, 32)
y = hip16(x)
print(y.shape)
x = torch.randn(1, 1024, 16)
y = hip256(x)
print(y.shape)
model = HiPIO(preprocess_layer, hip16s, postprocess_layer)
# x = torch.randn(2, 3, 32, 32)
x = torch.randn(2, 3, 16, 16)
y = model(x)
print(y.shape)
imshow(y[0], batched=False)