-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathmodel.py
171 lines (153 loc) · 8.41 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
from transformer import *
class CNN_Trans_LID(nn.Module):
def __init__(self, input_dim, feat_dim,
d_k, d_v, d_ff, n_heads=8,
dropout=0.1, n_lang=3, max_seq_len=10000):
super(CNN_Trans_LID, self).__init__()
self.input_dim = input_dim
self.feat_dim = feat_dim
self.dropout = nn.Dropout(p=dropout)
self.shared_TDNN = nn.Sequential(nn.Dropout(p=dropout),
nn.Conv1d(in_channels=input_dim, out_channels=512, kernel_size=1),
nn.ReLU(),
nn.BatchNorm1d(512, momentum=0.1, affine=True),
nn.Conv1d(in_channels=512, out_channels=512, kernel_size=1),
nn.ReLU(),
nn.BatchNorm1d(512, momentum=0.1, affine=True),
nn.Conv1d(in_channels=512, out_channels=512, kernel_size=1),
nn.ReLU(),
nn.BatchNorm1d(512, momentum=0.1, affine=True),
)
self.fc_xv = nn.Linear(1024, feat_dim)
self.layernorm1 = LayerNorm(feat_dim)
self.pos_encoding = PositionalEncoding(max_seq_len=max_seq_len, features_dim=feat_dim)
self.layernorm2 = LayerNorm(feat_dim)
self.d_model = feat_dim * n_heads
self.n_heads = n_heads
self.attention_block1 = EncoderBlock(self.d_model, d_k, d_v, d_ff, n_heads, dropout=dropout)
self.attention_block2 = EncoderBlock(self.d_model, d_k, d_v, d_ff, n_heads, dropout=dropout)
self.fc1 = nn.Linear(self.d_model * 2, self.d_model)
self.fc2 = nn.Linear(self.d_model, self.d_model)
self.fc3 = nn.Linear(self.d_model, n_lang)
def mean_std_pooling(self, x, batchsize, seq_lens, mask_mean, weight_mean, mask_std, weight_unb):
max_len = seq_lens[0]
feat_dim = x.size(-1)
if mask_mean is not None:
assert mask_mean.size() == x.size()
x.masked_fill_(mask_mean, 0)
correct_mean = x.mean(dim=1).transpose(0, 1) * weight_mean
correct_mean = correct_mean.transpose(0, 1)
center_seq = x - correct_mean.repeat(1, 1, max_len).view(batchsize, -1, feat_dim)
variance = torch.mean(torch.mul(torch.abs(center_seq) ** 2, mask_std), dim=1).transpose(0,1) \
* weight_unb * weight_mean
std = torch.sqrt(variance.transpose(0, 1))
return torch.cat((correct_mean, std), dim=1)
def forward(self, x, seq_len, mean_mask_=None, weight_mean=None, std_mask_=None, weight_unbaised=None,
atten_mask=None, eps=1e-5):
batch_size = x.size(0)
T_len = x.size(1)
x = self.dropout(x)
x = x.view(batch_size * T_len, -1, self.input_dim).transpose(-1, -2)
x = self.shared_TDNN(x)
if self.training:
shape = x.size()
noise = torch.Tensor(shape)
noise = noise.type_as(x)
torch.randn(shape, out=noise)
x += noise * eps
stats = torch.cat((x.mean(dim=2), x.std(dim=2)), dim=1)
embedding = self.fc_xv(stats)
embedding = embedding.view(batch_size, T_len, self.feat_dim)
output = self.layernorm1(embedding)
output = self.pos_encoding(output, seq_len)
output = self.layernorm2(output)
output = output.unsqueeze(1).repeat(1, self.n_heads, 1, 1)
output = output.transpose(1, 2).contiguous().view(batch_size, -1, self.d_model)
output, _ = self.attention_block1(output, atten_mask)
output, _ = self.attention_block2(output, atten_mask)
if std_mask_ is not None:
stats = self.mean_std_pooling(output, batch_size, seq_len, mean_mask_, weight_mean,
std_mask_, weight_unbaised)
else:
stats = torch.cat((output.mean(dim=1), output.std(dim=1)), dim=1)
output = F.relu(self.fc1(stats))
output = F.relu(self.fc2(output))
output = self.fc3(output)
return output
class PHOLID(nn.Module):
def __init__(self,input_dim, feat_dim,
d_k, d_v, d_ff, n_heads=8,
dropout=0.1, n_lang=3, max_seq_len=10000):
super(PHOLID, self).__init__()
self.input_dim = input_dim
self.d_model = feat_dim * n_heads
self.n_heads = n_heads
self.feat_dim = feat_dim
self.shared_TDNN = nn.Sequential(nn.Dropout(p=dropout),
nn.Conv1d(in_channels=input_dim, out_channels=512, kernel_size=1),
nn.ReLU(),
nn.BatchNorm1d(512, momentum=0.1),
nn.Conv1d(in_channels=512, out_channels=512, kernel_size=1),
nn.ReLU(),
nn.BatchNorm1d(512, momentum=0.1),
nn.Conv1d(in_channels=512, out_channels=512, kernel_size=1),
nn.ReLU(),
nn.BatchNorm1d(512, momentum=0.1))
self.phoneme_proj = nn.Linear(512, 64)
self.pos_encoding = PositionalEncoding(max_seq_len=max_seq_len, features_dim=feat_dim)
self.layernorm2 = LayerNorm(feat_dim)
self.fc_xv = nn.Linear(1024, feat_dim)
self.layernorm1 = LayerNorm(feat_dim)
self.pos_encoding = PositionalEncoding(max_seq_len=max_seq_len, features_dim=feat_dim)
self.layernorm2 = LayerNorm(feat_dim)
self.attention_block1 = EncoderBlock(self.d_model, d_k, d_v, d_ff, n_heads, dropout=dropout)
self.attention_block2 = EncoderBlock(self.d_model, d_k, d_v, d_ff, n_heads, dropout=dropout)
self.lid_clf = nn.Sequential(nn.Linear(self.d_model * 2, self.d_model),
nn.ReLU(),
nn.Linear(self.d_model, self.d_model),
nn.ReLU(),
nn.Linear(self.d_model, n_lang))
def mean_std_pooling(self, x, batchsize, seq_lens, mask_mean, weight_mean, mask_std, weight_unb):
max_len = seq_lens[0]
feat_dim = x.size(-1)
if mask_mean is not None:
assert mask_mean.size() == x.size()
x.masked_fill_(mask_mean, 0)
correct_mean = x.mean(dim=1).transpose(0, 1) * weight_mean
correct_mean = correct_mean.transpose(0, 1)
center_seq = x - correct_mean.repeat(1, 1, max_len).view(batchsize, -1, feat_dim)
variance = torch.mean(torch.mul(torch.abs(center_seq) ** 2, mask_std), dim=1).transpose(0, 1) \
* weight_unb * weight_mean
std = torch.sqrt(variance.transpose(0, 1))
return torch.cat((correct_mean, std), dim=1)
def forward(self, x, seq_len, mean_mask_=None, weight_mean=None, std_mask_=None, weight_unbaised=None,
atten_mask=None, eps=1e-5):
batch_size = x.size(0)
T_len = x.size(1)
x = x.view(batch_size * T_len, -1, self.input_dim).transpose(-1, -2)
x = self.shared_TDNN(x)
pho_x = x.transpose(-1, -2)
pho_out = self.phoneme_proj(pho_x)
if self.training:
shape = x.size()
noise = torch.Tensor(shape)
noise = noise.type_as(x)
torch.randn(shape, out=noise)
x += noise * eps
seg_stats = torch.cat((x.mean(dim=2), x.std(dim=2)), dim=1)
embedding = self.fc_xv(seg_stats)
embedding = embedding.view(batch_size, T_len, self.feat_dim)
output = self.layernorm1(embedding)
output = self.pos_encoding(output, seq_len)
output = self.layernorm2(output)
output = output.unsqueeze(1).repeat(1, self.n_heads, 1, 1)
output = output.transpose(1, 2).contiguous().view(batch_size, -1, self.d_model)
output, _ = self.attention_block1(output, atten_mask)
output, _ = self.attention_block2(output, atten_mask)
if std_mask_ is not None:
stats = self.mean_std_pooling(output, batch_size, seq_len, mean_mask_, weight_mean,
std_mask_, weight_unbaised)
else:
stats = torch.cat((output.mean(dim=1), output.std(dim=1)), dim=1)
output = self.lid_clf(stats)
return output, pho_out.reshape(batch_size, T_len, -1, 64)