forked from harvitronix/five-video-classification-methods
-
Notifications
You must be signed in to change notification settings - Fork 0
/
models.py
251 lines (222 loc) · 10 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
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
"""
A collection of models we'll use to attempt to classify videos.
"""
from keras.layers import Dense, Flatten, Dropout, ZeroPadding3D
from keras.layers.recurrent import LSTM
from keras.models import Sequential, load_model
from keras.optimizers import Adam, RMSprop
from keras.layers.wrappers import TimeDistributed
from keras.layers.convolutional import (Conv2D, MaxPooling3D, Conv3D,
MaxPooling2D)
from collections import deque
import sys
class ResearchModels():
def __init__(self, nb_classes, model, seq_length,
saved_model=None, features_length=2048):
"""
`model` = one of:
lstm
lrcn
mlp
conv_3d
c3d
`nb_classes` = the number of classes to predict
`seq_length` = the length of our video sequences
`saved_model` = the path to a saved Keras model to load
"""
# Set defaults.
self.seq_length = seq_length
self.load_model = load_model
self.saved_model = saved_model
self.nb_classes = nb_classes
self.feature_queue = deque()
# Set the metrics. Only use top k if there's a need.
metrics = ['accuracy']
if self.nb_classes >= 10:
metrics.append('top_k_categorical_accuracy')
# Get the appropriate model.
if self.saved_model is not None:
print("Loading model %s" % self.saved_model)
self.model = load_model(self.saved_model)
elif model == 'lstm':
print("Loading LSTM model.")
self.input_shape = (seq_length, features_length)
self.model = self.lstm()
elif model == 'lrcn':
print("Loading CNN-LSTM model.")
self.input_shape = (seq_length, 80, 80, 3)
self.model = self.lrcn()
elif model == 'mlp':
print("Loading simple MLP.")
self.input_shape = (seq_length, features_length)
self.model = self.mlp()
elif model == 'conv_3d':
print("Loading Conv3D")
self.input_shape = (seq_length, 80, 80, 3)
self.model = self.conv_3d()
elif model == 'c3d':
print("Loading C3D")
self.input_shape = (seq_length, 80, 80, 3)
self.model = self.c3d()
else:
print("Unknown network.")
sys.exit()
# Now compile the network.
optimizer = Adam(lr=1e-5, decay=1e-6)
self.model.compile(loss='categorical_crossentropy', optimizer=optimizer,
metrics=metrics)
print(self.model.summary())
def lstm(self):
"""Build a simple LSTM network. We pass the extracted features from
our CNN to this model predomenently."""
# Model.
model = Sequential()
model.add(LSTM(2048, return_sequences=False,
input_shape=self.input_shape,
dropout=0.5))
model.add(Dense(512, activation='relu'))
model.add(Dropout(0.5))
model.add(Dense(self.nb_classes, activation='softmax'))
return model
def lrcn(self):
"""Build a CNN into RNN.
Starting version from:
https://github.com/udacity/self-driving-car/blob/master/
steering-models/community-models/chauffeur/models.py
Heavily influenced by VGG-16:
https://arxiv.org/abs/1409.1556
Also known as an LRCN:
https://arxiv.org/pdf/1411.4389.pdf
"""
def add_default_block(model, kernel_filters, init, reg_lambda):
# conv
model.add(TimeDistributed(Conv2D(kernel_filters, (3, 3), padding='same',
kernel_initializer=init, kernel_regularizer=L2_reg(l=reg_lambda))))
model.add(TimeDistributed(BatchNormalization()))
model.add(TimeDistributed(Activation('relu')))
# conv
model.add(TimeDistributed(Conv2D(kernel_filters, (3, 3), padding='same',
kernel_initializer=init, kernel_regularizer=L2_reg(l=reg_lambda))))
model.add(TimeDistributed(BatchNormalization()))
model.add(TimeDistributed(Activation('relu')))
# max pool
model.add(TimeDistributed(MaxPooling2D((2, 2), strides=(2, 2))))
return model
initialiser = 'glorot_uniform'
reg_lambda = 0.001
model = Sequential()
# first (non-default) block
model.add(TimeDistributed(Conv2D(32, (7, 7), strides=(2, 2), padding='same',
kernel_initializer=initialiser, kernel_regularizer=L2_reg(l=reg_lambda)),
input_shape=self.input_shape))
model.add(TimeDistributed(BatchNormalization()))
model.add(TimeDistributed(Activation('relu')))
model.add(TimeDistributed(Conv2D(32, (3,3), kernel_initializer=initialiser, kernel_regularizer=L2_reg(l=reg_lambda))))
model.add(TimeDistributed(BatchNormalization()))
model.add(TimeDistributed(Activation('relu')))
model.add(TimeDistributed(MaxPooling2D((2, 2), strides=(2, 2))))
# 2nd-5th (default) blocks
model = add_default_block(model, 64, init=initialiser, reg_lambda=reg_lambda)
model = add_default_block(model, 128, init=initialiser, reg_lambda=reg_lambda)
model = add_default_block(model, 256, init=initialiser, reg_lambda=reg_lambda)
model = add_default_block(model, 512, init=initialiser, reg_lambda=reg_lambda)
# LSTM output head
model.add(TimeDistributed(Flatten()))
model.add(LSTM(256, return_sequences=False, dropout=0.5))
model.add(Dense(self.nb_classes, activation='softmax'))
return model
def mlp(self):
"""Build a simple MLP. It uses extracted features as the input
because of the otherwise too-high dimensionality."""
# Model.
model = Sequential()
model.add(Flatten(input_shape=self.input_shape))
model.add(Dense(512))
model.add(Dropout(0.5))
model.add(Dense(512))
model.add(Dropout(0.5))
model.add(Dense(self.nb_classes, activation='softmax'))
return model
def conv_3d(self):
"""
Build a 3D convolutional network, based loosely on C3D.
https://arxiv.org/pdf/1412.0767.pdf
"""
# Model.
model = Sequential()
model.add(Conv3D(
32, (3,3,3), activation='relu', input_shape=self.input_shape
))
model.add(MaxPooling3D(pool_size=(1, 2, 2), strides=(1, 2, 2)))
model.add(Conv3D(64, (3,3,3), activation='relu'))
model.add(MaxPooling3D(pool_size=(1, 2, 2), strides=(1, 2, 2)))
model.add(Conv3D(128, (3,3,3), activation='relu'))
model.add(Conv3D(128, (3,3,3), activation='relu'))
model.add(MaxPooling3D(pool_size=(1, 2, 2), strides=(1, 2, 2)))
model.add(Conv3D(256, (2,2,2), activation='relu'))
model.add(Conv3D(256, (2,2,2), activation='relu'))
model.add(MaxPooling3D(pool_size=(1, 2, 2), strides=(1, 2, 2)))
model.add(Flatten())
model.add(Dense(1024))
model.add(Dropout(0.5))
model.add(Dense(1024))
model.add(Dropout(0.5))
model.add(Dense(self.nb_classes, activation='softmax'))
return model
def c3d(self):
"""
Build a 3D convolutional network, aka C3D.
https://arxiv.org/pdf/1412.0767.pdf
With thanks:
https://gist.github.com/albertomontesg/d8b21a179c1e6cca0480ebdf292c34d2
"""
model = Sequential()
# 1st layer group
model.add(Conv3D(64, 3, 3, 3, activation='relu',
border_mode='same', name='conv1',
subsample=(1, 1, 1),
input_shape=self.input_shape))
model.add(MaxPooling3D(pool_size=(1, 2, 2), strides=(1, 2, 2),
border_mode='valid', name='pool1'))
# 2nd layer group
model.add(Conv3D(128, 3, 3, 3, activation='relu',
border_mode='same', name='conv2',
subsample=(1, 1, 1)))
model.add(MaxPooling3D(pool_size=(2, 2, 2), strides=(2, 2, 2),
border_mode='valid', name='pool2'))
# 3rd layer group
model.add(Conv3D(256, 3, 3, 3, activation='relu',
border_mode='same', name='conv3a',
subsample=(1, 1, 1)))
model.add(Conv3D(256, 3, 3, 3, activation='relu',
border_mode='same', name='conv3b',
subsample=(1, 1, 1)))
model.add(MaxPooling3D(pool_size=(2, 2, 2), strides=(2, 2, 2),
border_mode='valid', name='pool3'))
# 4th layer group
model.add(Conv3D(512, 3, 3, 3, activation='relu',
border_mode='same', name='conv4a',
subsample=(1, 1, 1)))
model.add(Conv3D(512, 3, 3, 3, activation='relu',
border_mode='same', name='conv4b',
subsample=(1, 1, 1)))
model.add(MaxPooling3D(pool_size=(2, 2, 2), strides=(2, 2, 2),
border_mode='valid', name='pool4'))
# 5th layer group
model.add(Conv3D(512, 3, 3, 3, activation='relu',
border_mode='same', name='conv5a',
subsample=(1, 1, 1)))
model.add(Conv3D(512, 3, 3, 3, activation='relu',
border_mode='same', name='conv5b',
subsample=(1, 1, 1)))
model.add(ZeroPadding3D(padding=(0, 1, 1)))
model.add(MaxPooling3D(pool_size=(2, 2, 2), strides=(2, 2, 2),
border_mode='valid', name='pool5'))
model.add(Flatten())
# FC layers group
model.add(Dense(4096, activation='relu', name='fc6'))
model.add(Dropout(0.5))
model.add(Dense(4096, activation='relu', name='fc7'))
model.add(Dropout(0.5))
model.add(Dense(self.nb_classes, activation='softmax'))
return model