-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathmodel.py
158 lines (132 loc) · 6.9 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
import torch
import torch.nn as nn
import torch.nn.functional as F
from utils import generate_candidates
class DynamicGRU(nn.Module):
def __init__(self, input_size, hidden_size, num_layers=1, bias=True,
batch_first=False, dropout=0, bidirectional=False):
super().__init__()
self.batch_first = batch_first
self.gru = nn.GRU(input_size, hidden_size, num_layers=num_layers, bias=bias,
batch_first=batch_first, dropout=dropout, bidirectional=bidirectional)
self.gru.flatten_parameters()
def forward(self, x, seq_len, max_num_frames):
sorted_seq_len, sorted_idx = torch.sort(seq_len, dim=0, descending=True)
_, original_idx = torch.sort(sorted_idx, dim=0, descending=False)
if self.batch_first:
sorted_x = x.index_select(0, sorted_idx)
else:
sorted_x = x.index_select(1, sorted_idx)
packed_x = nn.utils.rnn.pack_padded_sequence(
sorted_x, sorted_seq_len.cpu().data.numpy(), batch_first=self.batch_first)
out, state = self.gru(packed_x)
unpacked_x, unpacked_len = nn.utils.rnn.pad_packed_sequence(out, batch_first=self.batch_first)
if self.batch_first:
out = unpacked_x.index_select(0, original_idx)
if out.shape[1] < max_num_frames:
out = F.pad(out, [0, 0, 0, max_num_frames - out.shape[1]])
else:
out = unpacked_x.index_select(1, original_idx)
if out.shape[0] < max_num_frames:
out = F.pad(out, [0, 0, 0, 0, 0, max_num_frames - out.shape[0]])
return out
class NodeInitializer(nn.Module):
def __init__(self, node_num, input_dim, node_dim, dropout):
super().__init__()
self.node_num = node_num
self.dropout = dropout
self.rnn = DynamicGRU(input_dim, node_dim >> 1, bidirectional=True, batch_first=True)
self.fc = nn.Linear(node_dim, node_dim)
def forward(self, x, mask):
length = mask.sum(dim=-1)
x = self.rnn(x, length, self.node_num)
x = F.leaky_relu(self.fc(x))
x = F.dropout(x, self.dropout, self.training)
return x
class GraphConvolution(nn.Module):
def __init__(self, node_dim):
super().__init__()
self.node_dim = node_dim
self.wvv = nn.Linear(self.node_dim, self.node_dim, bias=False)
self.wss = nn.Linear(self.node_dim, self.node_dim, bias=False)
self.wvs = nn.Linear(self.node_dim, self.node_dim, bias=False)
self.wsv = nn.Linear(self.node_dim, self.node_dim, bias=False)
self.wgatev = nn.Linear(self.node_dim << 1, self.node_dim)
self.wgates = nn.Linear(self.node_dim << 1, self.node_dim)
def forward(self, v, avv, s, ass):
vs = torch.matmul(v, s.transpose(2, 1))
avs = torch.softmax(vs, -1)
asv = torch.softmax(vs.transpose(2, 1), -1)
v = F.leaky_relu(self.wvv(torch.matmul(avv, v)))
s = F.leaky_relu(self.wss(torch.matmul(ass, s)))
hv = self.wsv(torch.matmul(avs, s)) # batch_size x T x node_dim
zv = torch.sigmoid(self.wgatev(torch.cat([v, hv], dim=-1))) # batch_size x T x node_dim
v = zv * v + (1 - zv) * hv # batch_size x T x node_dim
v = F.leaky_relu(v)
hs = self.wvs(torch.matmul(asv, v))
zs = torch.sigmoid(self.wgates(torch.cat([s, hs], dim=-1)))
s = zs * s + (1 - zs) * hs
s = F.leaky_relu(s)
return v, s
class Model(nn.Module):
def __init__(self, opt):
super().__init__()
self.alpha = opt.alpha
self.beta = opt.beta
self.dropout = opt.dropout
self.vnode_initializer = NodeInitializer(node_num=opt.max_frames_num,
input_dim=opt.frame_feature_dim,
node_dim=opt.node_dim, dropout=self.dropout)
self.wnode_initializer = NodeInitializer(node_num=opt.max_words_num,
input_dim=opt.word_feature_dim,
node_dim=opt.node_dim, dropout=self.dropout)
self.s_embed = nn.Linear(opt.max_words_num, 1)
self.gcn_layers = nn.ModuleList([
GraphConvolution(opt.node_dim)
for _ in range(opt.gcn_layers_num)
])
self.candidates, self.window_widths = generate_candidates(opt.max_frames_num,
opt.window_widths, opt.window_stride)
self.candidates = torch.from_numpy(self.candidates).float().cuda()
self.conv_cls = nn.ModuleList([
nn.Conv1d(opt.node_dim << 1, 1, w * 2, padding=w // 2, stride=opt.window_stride)
for w in self.window_widths
])
self.conv_reg = nn.ModuleList([
nn.Conv1d(opt.node_dim << 1, 2, w * 2, padding=w // 2, stride=opt.window_stride)
for w in self.window_widths
])
self.criterion_BCE = nn.BCEWithLogitsLoss()
self.criterion_CE = nn.CrossEntropyLoss()
self.criterion_reg = nn.SmoothL1Loss()
def forward(self, vfeats, frame_mask, frame_mat, wfeats, word_mask, word_mats, label, scores):
v = self.vnode_initializer(vfeats, frame_mask) # batch_size x T x node_dim
s = self.wnode_initializer(wfeats, word_mask) # batch_size x L x node_dim
for g in self.gcn_layers:
v, s = g(v, frame_mat, s, word_mats)
v = F.dropout(v, self.dropout, self.training)
s = F.dropout(s, self.dropout, self.training)
s = F.leaky_relu(self.s_embed(s.permute(0, 2, 1)).permute(0, 2, 1))
s = s.expand(s.shape[0], v.shape[1], s.shape[2]) # batch_size x window_num x node_dim
v = torch.cat((v, s), 2) # batch_size x window_num x 2node_dim
predict_scores = torch.cat([
self.conv_cls[i](v.permute(0, 2, 1)).permute(0, 2, 1)
for i in range(len(self.conv_cls))
], dim=1).squeeze(2)
offset = torch.cat([
self.conv_reg[i](v.permute(0, 2, 1)).permute(0, 2, 1)
for i in range(len(self.conv_reg))
], dim=1)
if self.training:
indices = scores.max(dim=1)[1]
else:
indices = predict_scores.max(dim=1)[1]
predict_box = self.candidates[indices] # batch_size x 2
predict_reg = offset[range(offset.shape[0]), indices] # batch_size x 2
refined_box = predict_box + predict_reg # batch_size x 2
bce_loss = self.criterion_BCE(predict_scores, scores)
indices_label = scores.max(-1)[-1] # batch_size x 1
ce_loss = self.criterion_CE(predict_scores, indices_label)
reg_loss = self.criterion_reg(refined_box, label.float())
loss = bce_loss + self.alpha * ce_loss + self.beta * reg_loss
return refined_box, loss