forked from hutchresearch/fine-grained-cad
-
Notifications
You must be signed in to change notification settings - Fork 0
/
models.py
176 lines (142 loc) · 5.78 KB
/
models.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
import json
import warnings
from pase.models.frontend import wf_builder
import torch
import torch.nn as nn
warnings.filterwarnings('ignore')
class TransposeLayerNorm(nn.Module):
def __init__(self, features, *args, **kwargs):
super(TransposeLayerNorm, self).__init__()
self.lnorm = nn.LayerNorm(features, *args, **kwargs)
def forward(self, x):
# MBS x EMB x SEQ
x = x.transpose(1, 2)
x = self.lnorm(x)
x = x.transpose(1, 2)
return x
def build_smooth_out(hid, cla, smooth=1):
return nn.Sequential(
nn.Conv1d(hid, hid, smooth, padding=smooth//2, padding_mode='replicate'),
nn.LeakyReLU(),
nn.Conv1d(hid, cla, 1)
)
def build_norm(norm, hid, affine=True):
if norm == 'bnorm':
return nn.BatchNorm1d(hid, affine=affine)
elif norm == 'lnorm':
return TransposeLayerNorm(hid, elementwise_affine=affine)
elif norm == 'inorm':
return nn.InstanceNorm1d(hid, affine=False)
elif norm == 'affinorm':
return nn.InstanceNorm1d(hid, affine=True)
elif norm is None:
return None
else:
raise TypeError('Unrecognized norm type: ', norm)
class PASEEncodedModel(nn.Module):
def __init__(self, head, cfg, checkpoint, drop_inp=0.0, drop_emb=0.0, freeze_bn=False, tune=False):
super(PASEEncodedModel, self).__init__()
with open(cfg, 'r') as f:
cfg = json.load(f)
# Note: Modify cfg dict here if desired
# This seems to overall make things worse though
# cfg['norm_type'] = norm
encoder = wf_builder(cfg)
encoder.load_pretrained(checkpoint, load_last=True, verbose=False)
if not tune:
encoder.requires_grad_(False)
self.tune = tune
self.freeze_bn = freeze_bn
self.drop_inp = nn.Dropout(p=drop_inp)
self.encoder = encoder
self.drop_emb = nn.Dropout(p=drop_emb)
self.head = head
def forward(self, signal=None, precomputed=None):
if signal is not None:
x = self.drop_inp(signal)
x = self.encoder(x)
if not self.tune:
x = x.detach()
if precomputed is not None:
if signal is None:
x = precomputed
else:
x = torch.cat([x, precomputed], dim=1)
x = self.drop_emb(x)
y = self.head(x)
return y
def train(self, mode=True):
self.training = mode
for module in self.children():
module.train(mode)
if self.freeze_bn:
self.encoder.eval()
return self
class CNNHead(nn.Module):
def __init__(self, input_size, num_classes, dilation_base, hidden_channels,
smooth, context_size, kernel_size, norm, dropout):
super(CNNHead, self).__init__()
# Convs
self.dilated_convs = nn.ModuleList()
hidden_channels.insert(0, input_size)
for i in range(1, len(hidden_channels)):
ks = context_size if i == 1 else kernel_size
dilation = int(round(dilation_base**(i-1)))
block = [
nn.Conv1d(hidden_channels[i-1], hidden_channels[i], ks, dilation=dilation),
nn.LeakyReLU(),
]
if norm is not None:
block.append(build_norm(norm, hidden_channels[i], affine=True))
self.dilated_convs.append(nn.Sequential(*block))
# Hidden to output
self.dropout = nn.Dropout(p=dropout)
self.hid2out = build_smooth_out(hidden_channels[-1], num_classes, smooth=smooth)
def forward(self, x):
# Forward through CNN
for dc in self.dilated_convs:
x = dc(x) # x: [mbs x DC[i] x seq]
# Get output and smooth
x = self.dropout(x)
y = self.hid2out(x) # x: [mbs x cla x seq]
return y
class LSTMHead(nn.Module):
def __init__(self, input_size, num_classes, hidden_size, hidden_layers,
smooth, bidirectional, dropout):
super(LSTMHead, self).__init__()
# LSTM
self.lstm = nn.LSTM(input_size, hidden_size, hidden_layers, dropout=dropout,
batch_first=True, bidirectional=bidirectional)
# Hidden to Output
h_size = hidden_size*2 if bidirectional else hidden_size
self.dropout = nn.Dropout(p=dropout)
self.hid2out = build_smooth_out(h_size, num_classes, smooth=smooth)
def forward(self, x):
# Forward through the LSTM
x = x.transpose(1, 2) # [mbs x emb x seq] --> [mbs x seq x emb]
x, _ = self.lstm(x) # [mbs x seq x emb] --> [mbs x seq x hid]
x = x.transpose(1, 2) # [mbs x seq x hid] --> [mbs x hid x seq]
# Get output and smooth
x = self.dropout(x)
y = self.hid2out(x) # [mbs x hid x seq] --> [mbs x cls x seq]
return y
class GRUHead(nn.Module):
def __init__(self, input_size, num_classes, hidden_size, hidden_layers,
smooth, bidirectional, dropout):
super(GRUHead, self).__init__()
# GRU
self.gru = nn.GRU(input_size, hidden_size, hidden_layers, dropout=dropout,
batch_first=True, bidirectional=bidirectional)
# Hidden to Output
h_size = hidden_size*2 if bidirectional else hidden_size
self.dropout = nn.Dropout(p=dropout)
self.hid2out = build_smooth_out(h_size, num_classes, smooth=smooth)
def forward(self, x):
# Forward through the GRU
x = x.transpose(1, 2) # [mbs x emb x seq] --> [mbs x seq x emb]
x, _ = self.gru(x) # [mbs x seq x emb] --> [mbs x seq x hid]
x = x.transpose(1, 2) # [mbs x seq x hid] --> [mbs x hid x seq]
# Get output and smooth
x = self.dropout(x)
y = self.hid2out(x) # [mbs x hid x seq] --> [mbs x cls x seq]
return y