-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathlayer.py
147 lines (116 loc) · 6.28 KB
/
layer.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
import torch
import torch.nn as nn
import torch.nn.functional as F
from torchvision import models
import math
import numpy as np
import random
from torch.autograd import Variable
class Pretrained3DResNet(nn.Module):
def __init__(self, out_dim):
super(Pretrained3DResNet, self).__init__()
# ResNet (2+1)D
self.model = models.video.r2plus1d_18(pretrained=True)
#Add one fully connected layer
self.model.fc = nn.Sequential(nn.Dropout(0.0),
nn.Linear(in_features=self.model.fc.in_features, out_features=out_dim))
def forward(self, video):
return self.model(video)
class PretrainedResNet(nn.Module):
def __init__(self, out_dim, pretrained=True):
super(PretrainedResNet, self).__init__()
self.resnet = models.resnet18(pretrained=pretrained)
self.resnet.fc = nn.Sequential(nn.Dropout(0.0),
nn.Linear(in_features=self.resnet.fc.in_features, out_features=out_dim))
self.out_dim = out_dim
def forward(self, audio):
out = self.resnet(audio)
if self.out_dim == 1:
out = torch.sigmoid(out)
return out
class AddNorm(nn.Module):
def __init__(self, d_model, dropout=0.1, norm_method='layer'):
super(AddNorm, self).__init__()
self.dropout = nn.Dropout(dropout)
if norm_method == 'layer':
self.norm = nn.LayerNorm(d_model)
else:
self.norm = nn.BatchNorm1d(d_model)
def forward(self, x_old, x_new):
return self.norm(self.dropout(x_new) + x_old)
def attention_matmul(q,k,v):
d_k = q.size(-1)
scores = torch.matmul(q, k.transpose(-2, -1)/math.sqrt(d_k))
#print('QK dimension', scores.shape)
attention = F.softmax(scores, dim=-1)
return torch.matmul(attention, v)
class MultiHeadAttention(nn.Module):
def __init__(self, num_of_heads, d_model):
super(MultiHeadAttention, self).__init__()
assert d_model % num_of_heads == 0
self.d_k = int(d_model/num_of_heads)
self.num_of_heads = num_of_heads
self.linears = nn.ModuleList([nn.Linear(d_model, d_model) for _ in range(3)])
def forward(self, query, key, value):
num_of_batches = query.size(0)
query, key, value = [linear(x).view(num_of_batches, -1, self.num_of_heads, self.d_k).transpose(1,2)\
for linear, x in zip(self.linears, (query, key, value))]
x = attention_matmul(query, key, value)
x = x.transpose(1, 2).contiguous().view(num_of_batches, self.num_of_heads*self.d_k)
return x
class Sec2SecCoattention(nn.Module):
def __init__(self, d, num_of_heads, lstm_hidden_dim, num_of_lstm_layers, num_of_dense_layers,\
dense_hidden_dim, device, norm_method, dropout=0.1):
super(Sec2SecCoattention, self).__init__()
self.lstm_hidden_dim = lstm_hidden_dim
self.num_of_lstm_layers = num_of_lstm_layers
self.dense_hidden_dim = dense_hidden_dim
self.device = device
self.d = d
self.num_of_heads = num_of_heads
self.visual = PretrainedResNet(d)
self.audio = PretrainedResNet(d)
self.norm_visual = AddNorm(d, dropout, norm_method)
self.norm_audio = AddNorm(d, dropout, norm_method)
self.linear_visual = nn.Linear(d, d)
self.linear_audio = nn.Linear(d, d)
self.visual_layer = MultiHeadAttention(num_of_heads, d)
self.audio_layer = MultiHeadAttention(num_of_heads, d)
self.coattention_visual = MultiHeadAttention(num_of_heads, d)
self.coattention_audio = MultiHeadAttention(num_of_heads, d)
self.coattention_norm_visual = AddNorm(d, dropout, norm_method)
self.coattention_norm_audio = AddNorm(d, dropout, norm_method)
self.lstm = nn.LSTM(d+d, lstm_hidden_dim, num_of_lstm_layers, bias=True, batch_first=True)
fc_layers = []
prev_dense_dim = lstm_hidden_dim
for i in range(num_of_dense_layers):
linear = nn.Linear(prev_dense_dim, dense_hidden_dim)
linear.weight.data.normal_(mean=0.0, std=1)
linear.bias.data.zero_()
fc_layers.append(linear)
prev_hidden_dim = dense_hidden_dim
fc_layers.append(nn.Linear(prev_hidden_dim, 1))
self.fc = nn.Sequential(*fc_layers)
def get_init_hidden_state(self, batch_size):
h0 = Variable(torch.zeros(self.num_of_lstm_layers, batch_size, self.lstm_hidden_dim), requires_grad=False).to(self.device)
c0 = Variable(torch.zeros(self.num_of_lstm_layers, batch_size, self.lstm_hidden_dim), requires_grad=False).to(self.device)
return h0, c0
def forward(self, visual_batch, audio_batch):
visual_shape = visual_batch.shape
audio_shape = audio_batch.shape
#visual_shape: batch_size, audio_length, 3_channel, feat_dim1, feat_dim2
#audio_shape: batch_size, audio_length, 3_channel, feat_dim1, feat_dim2
x_visual = visual_batch.view(visual_shape[0] * visual_shape[1], visual_shape[2], visual_shape[3], visual_shape[4])
x_audio = audio_batch.view(audio_shape[0] * audio_shape[1], audio_shape[2], audio_shape[3], audio_shape[4])
x_visual = self.visual(x_visual)
x_audio = self.audio(x_audio)
x_visual = self.linear_visual(self.norm_visual(x_visual,self.visual_layer(x_visual, x_visual, x_visual)))
x_audio = self.linear_audio(self.norm_audio(x_audio, self.audio_layer(x_audio, x_audio, x_audio)))
x_coattention_visual = self.coattention_norm_visual(x_visual, self.coattention_visual(x_visual, x_audio, x_audio))
x_coattention_audio = self.coattention_norm_audio(x_audio, self.coattention_audio(x_audio, x_visual, x_visual))
x_coattention = torch.cat((x_coattention_visual, x_coattention_audio), dim=1)
x_coattention = x_coattention.view(visual_shape[0], visual_shape[1], self.d + self.d)
lstm_output, _ = self.lstm(x_coattention, self.get_init_hidden_state(visual_shape[0]))
lstm_last_step_output = lstm_output[: , visual_shape[1] - 1, :]
out = self.fc(lstm_last_step_output)
return torch.sigmoid(out)