-
Notifications
You must be signed in to change notification settings - Fork 9
/
MSAF.py
156 lines (135 loc) · 7.64 KB
/
MSAF.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
# Copyright (c) 2020 Anita Hu and Kevin Su
#
# Permission is hereby granted, free of charge, to any person obtaining a copy
# of this software and associated documentation files (the "Software"), to deal
# in the Software without restriction, including without limitation the rights
# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
# copies of the Software, and to permit persons to whom the Software is
# furnished to do so, subject to the following conditions:
#
# The above copyright notice and this permission notice shall be included in all
# copies or substantial portions of the Software.
#
# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
# SOFTWARE.
import torch
import torch.nn as nn
import torch.nn.functional as F
import math
# The probability of dropping a block
class BlockDropout(nn.Module):
def __init__(self, p: float = 0.5):
super(BlockDropout, self).__init__()
if p < 0 or p > 1:
raise ValueError(
"dropout probability has to be between 0 and 1, " "but got {}".format(p)
)
self.p: float = p
def forward(self, X):
if self.training:
blocks_per_mod = [x.shape[1] for x in X]
mask_size = torch.Size([X[0].shape[0], sum(blocks_per_mod)])
binomial = torch.distributions.binomial.Binomial(probs=1 - self.p)
mask = binomial.sample(mask_size).to(X[0].device) * (1.0 / (1 - self.p))
mask_shapes = [list(x.shape[:2]) + [1] * (x.dim() - 2) for x in X]
grouped_masks = torch.split(mask, blocks_per_mod, dim=1)
grouped_masks = [m.reshape(s) for m, s in zip(grouped_masks, mask_shapes)]
X = [x * m for x, m in zip(X, grouped_masks)]
return X, grouped_masks
return X, None
# squeeze dim default 1: i.e. channel in (bs, channel, height, width, ...)
# Parameters:
# in_channels: a list of channel numbers for modalities
# block_channel: the channel number of each equal-sized block
# reduction_factor: c' = c / reduction_factor, where c is block_channel
# lowest_atten: float number between 0 and 1. Attention value will be mapped to:
# lowest_atten + attention_value * (1 - lowest_atten)
class MSAFBlock(nn.Module):
def __init__(self, in_channels, block_channel, block_dropout=0., lowest_atten=0., reduction_factor=4):
super(MSAFBlock, self).__init__()
self.block_channel = block_channel
self.in_channels = in_channels
self.lowest_atten = lowest_atten
self.num_modality = len(in_channels)
self.reduced_channel = self.block_channel // reduction_factor
self.block_dropout = BlockDropout(p=block_dropout) if 0 < block_dropout < 1 else None
self.joint_features = nn.Sequential(
nn.Linear(self.block_channel, self.reduced_channel),
nn.BatchNorm1d(self.reduced_channel),
nn.ReLU(inplace=True)
)
self.num_blocks = [math.ceil(ic / self.block_channel) for ic in
in_channels] # number of blocks for each modality
self.last_block_padding = [ic % self.block_channel for ic in in_channels]
self.dense_group = nn.ModuleList([nn.Linear(self.reduced_channel, self.block_channel)
for i in range(sum(self.num_blocks))])
self.soft_attention = nn.Softmax(dim=0)
self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
# X: a list of features from different modalities
def forward(self, X):
bs_ch = [x.size()[:2] for x in X]
for bc, ic in zip(bs_ch, self.in_channels):
assert bc[1] == ic, "X shape and in_channels are different. X channel {} but got {}".format(str(bc[1]),
str(ic))
# pad channel if block_channel non divisible
padded_X = [F.pad(x, (0, pad_size)) for pad_size, x in zip(self.last_block_padding, X)]
# reshape each feature map: [batch size, N channels, ...] -> [batch size, N blocks, block channel, ...]
desired_shape = [[x.shape[0], nb, self.block_channel] + list(x.shape[2:]) for x, nb in zip(padded_X, self.num_blocks)]
reshaped_X = [torch.reshape(x, ds) for x, ds in zip(padded_X, desired_shape)]
if self.block_dropout:
reshaped_X, masks = self.block_dropout(reshaped_X)
# element wise sum of blocks then global ave pooling on channel
elem_sum_X = [torch.sum(x, dim=1) for x in reshaped_X]
gap = [F.adaptive_avg_pool1d(sx.view(list(sx.size()[:2]) + [-1]), 1) for sx in elem_sum_X]
# combine GAP over modalities and generate attention values
gap = torch.stack(gap).sum(dim=0) # / (self.num_modality - 1)
gap = torch.squeeze(gap, -1)
gap = self.joint_features(gap)
atten = self.soft_attention(torch.stack([dg(gap) for dg in self.dense_group])).permute(1, 0, 2)
atten = self.lowest_atten + atten * (1 - self.lowest_atten)
# apply attention values to features
atten_shapes = [list(x.shape[:3]) + [1] * (x.dim() - 3) for x in reshaped_X]
grouped_atten = torch.split(atten, self.num_blocks, dim=1)
grouped_atten = [a.reshape(s) for a, s in zip(grouped_atten, atten_shapes)]
if self.block_dropout and self.training:
reshaped_X = [x * m * a for x, m, a in zip(reshaped_X, masks, grouped_atten)]
else:
reshaped_X = [x * a for x, a in zip(reshaped_X, grouped_atten)]
X = [x.reshape(org_x.shape) for x, org_x in zip(reshaped_X, X)]
return X
class MSAF(nn.Module):
def __init__(self, in_channels, block_channel, block_dropout, lowest_atten=0., reduction_factor=4,
split_block=1):
super(MSAF, self).__init__()
self.num_modality = len(in_channels)
self.split_block = split_block
self.blocks = nn.ModuleList([MSAFBlock(in_channels, block_channel, block_dropout, lowest_atten,
reduction_factor) for i in range(split_block)])
# X: a list of features from different modalities
def forward(self, X):
if self.split_block == 1:
ret = self.blocks[0](X) # only 1 MSAF block
else:
# split into multiple time segments, assumes at dim=2
segment_shapes = [[x.shape[2] // self.split_block] * self.split_block for x in X]
for x, seg_shape in zip(X, segment_shapes):
seg_shape[-1] += x.shape[2] % self.split_block
segmented_x = [torch.split(x, seg_shape, dim=2) for x, seg_shape in zip(X, segment_shapes)]
# process segments using MSAF blocks
ret_segments = [self.blocks[i]([x[i] for x in segmented_x]) for i in range(self.split_block)]
# put segments back together
ret = [torch.cat([r[m] for r in ret_segments], dim=2) for m in range(self.num_modality)]
return ret
if __name__ == "__main__":
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
m1 = torch.rand(4, 32, 64, 64, 50).to(device)
m2 = torch.rand(4, 16, 32, 32, 53).to(device)
x = [m1, m2]
net = MSAF([32, 16], 8, block_dropout=0.2, reduction_factor=4, split_block=5).to(device)
y = net(x)
print(y[0].shape, y[1].shape)