-
Notifications
You must be signed in to change notification settings - Fork 12
/
model_speech_yolo.py
144 lines (120 loc) · 5.65 KB
/
model_speech_yolo.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
__author__ = 'YaelSegal & TzeviyaFuchs'
# VGG model was taken from Yossi Adi
import torch.nn as nn
import torch.nn.functional as F
import torch
def _make_layers(cfg, kernel=3):
layers = []
in_channels = 1
for x in cfg:
if x == 'M':
layers += [nn.MaxPool2d(kernel_size=2, stride=2)]
else:
layers += [nn.Conv2d(in_channels, x, kernel_size=kernel, padding=1),
nn.BatchNorm2d(x),
nn.ReLU(inplace=True)]
in_channels = x
layers += [nn.AvgPool2d(kernel_size=1, stride=1)]
return nn.Sequential(*layers)
cfg = {
'VGG11': [64, 'M', 128, 'M', 256, 256, 'M', 512, 512, 'M', 512, 512, 'M'],
'VGG13': [64, 64, 'M', 128, 128, 'M', 256, 256, 'M', 512, 512, 'M', 512, 512, 'M'],
'VGG16': [64, 64, 'M', 128, 128, 'M', 256, 256, 256, 'M', 512, 512, 512, 'M', 512, 512, 512, 'M'],
'VGG19': [64, 64, 'M', 128, 128, 'M', 256, 256, 256, 256, 'M', 512, 512, 512, 512, 'M', 512, 512, 512, 512, 'M'],
}
class VGG(nn.Module):
def __init__(self, vgg_name, class_num=30):
super(VGG, self).__init__()
self.features = _make_layers(cfg[vgg_name])
self.fc1 = nn.Linear(7680, 512)
self.fc2 = nn.Linear(512, class_num)
def forward(self, x):
# out = self.features(x)
for m in self.features.children():
x = m(x)
out = x
out = out.view(out.size(0), -1)
out = self.fc1(out)
out = self.fc2(out)
return F.log_softmax(out, dim=1)
class SpeechYoloVGGNet(nn.Module):
def __init__(self, classfication_model, c=7, b=2, k=10, dropout=0):
super(SpeechYoloVGGNet, self).__init__()
self.c = c
self.b = b
self.k = k
self.dropout_exists = dropout
self.dropout = nn.Dropout(p=dropout)
self.removed = list(classfication_model.children())[:-1]
last_input_size = 512
self.model = nn.Sequential(*self.removed)
self.batch_norm = nn.BatchNorm1d(last_input_size)
self.last_layer = nn.Linear(last_input_size, self.c * (self.b * 3 + self.k))
self.init_weight()
def forward(self, x):
for m in self.model.children():
classname = m.__class__.__name__
if 'Sequential' in classname:
for seq_child in m.children():
x = seq_child(x)
elif classname.find('Linear') != -1:
x = x.view(x.size(0), -1)
x = F.relu(m(x))
if not self.dropout_exists == 0.0:
x = self.dropout(x)
last_layer_output = self.last_layer(x)
reshaped_output = last_layer_output.contiguous().view(-1, self.c, self.b * 3 + self.k)
pred_coords = reshaped_output[:, :, :3 * self.b].contiguous().view(-1, self.c, self.b,
3) # get all the x,w values for all the boxes
target_xs = torch.sigmoid(pred_coords[:, :, :, 0].view(-1, self.c, self.b))
target_ws = torch.sigmoid(pred_coords[:, :, :, 1].view(-1, self.c, self.b))
target_conf = torch.sigmoid(pred_coords[:, :, :, 2].view(-1, self.c, self.b))
target_class_prob = F.softmax(reshaped_output[:, :, 3 * self.b:].contiguous().view(-1, self.c, self.k), 2)
final_output = torch.cat((target_xs, target_ws, target_conf, target_class_prob), 2)
return final_output
def init_weight(self):
torch.nn.init.xavier_normal_(self.last_layer.weight.data)
def init_mult_weights(self):
torch.nn.init.xavier_normal_(self.last_layer[0].weight.data)
torch.nn.init.xavier_normal_(self.last_layer[2].weight.data)
# [1] is batchnorm
self.last_layer[1].weight.data.normal_(1.0, 0.02)
self.last_layer[1].bias.data.normal_(1.0, 0.02)
@staticmethod
def init_pre_model_weights(m):
classname = m.__class__.__name__
if classname.find('Conv') != -1 or classname.find('Linear') != -1:
nn.init.xavier_normal_(m.weight)
elif classname.find('BatchNorm') != -1:
m.weight.data.normal_(1.0, 0.02)
m.bias.data.normal_(1.0, 0.02)
def load_model(save_dir):
checkpoint = torch.load(save_dir, map_location=lambda storage, loc: storage)
config_dict = checkpoint['config_dict']
arc_type = checkpoint['arc']
if arc_type.startswith('VGG'):
model_ = VGG(arc_type)
speech_net = SpeechYoloVGGNet(model_, config_dict["C"], config_dict["B"], config_dict["K"])
speech_net.load_state_dict(checkpoint['net'])
else:
raise Exception("No such architecture")
return speech_net, checkpoint['acc'], checkpoint['epoch']
def create_speech_model(pretrained_model, arc, config_dict, dropout):
if pretrained_model:
if arc.startswith('VGG'):
checkpoint = torch.load(pretrained_model, map_location=lambda storage,
loc: storage) # will forcefully remap everything onto CPU
class_num = checkpoint['class_num']
model_ = VGG(arc, class_num=class_num)
model_.load_state_dict(checkpoint['net'])
speech_net = SpeechYoloVGGNet(model_, config_dict["C"], config_dict["B"], config_dict["K"], dropout=dropout)
else:
raise Exception("No such architecture")
else:
if arc.startswith('VGG'):
model_ = VGG(arc)
speech_net = SpeechYoloVGGNet(model_, config_dict["C"], config_dict["B"], config_dict["K"], dropout=dropout)
speech_net.model.apply(speech_net.init_pre_model_weights)
else:
raise Exception("No such architecture")
return speech_net