forked from gogolgrind/Cascade-CNN-Face-Detection
-
Notifications
You must be signed in to change notification settings - Fork 0
/
cnn_cascade_lasagne.py
262 lines (240 loc) · 12.2 KB
/
cnn_cascade_lasagne.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
252
253
254
255
256
257
258
259
260
261
262
# -*- coding: utf-8 -*-
"""
Created on Tue May 31 20:20:22 2016
@author: Kostya
"""
from lasagne.nonlinearities import softmax, rectify as relu
from lasagne import layers
from lasagne import updates
from lasagne import regularization
from lasagne import objectives
from time import time
from six.moves import cPickle as pickle
import theano
import theano.tensor as T
import scipy as sp
import sys
sys.setrecursionlimit(10000)
class Cnn(object):
net = None
subnet = None
nn_name = ''
l_rates = []
max_epochs = 120
batch_size = 256
verbose = 0
eta = None
__train_fn__ = None
# create classifcation nets
def __build_12_net__(self):
network = layers.InputLayer((None, 3, 12, 12), input_var=self.__input_var__)
network = layers.Conv2DLayer(network,num_filters=16,filter_size=(3,3),stride=1,nonlinearity=relu)
network = layers.MaxPool2DLayer(network, pool_size = (3,3),stride = 2)
network = layers.DropoutLayer(network)
network = layers.DenseLayer(network,num_units = 16,nonlinearity = relu)
#network = layers.Conv2DLayer(network,num_filters=16,filter_size=(1,1),stride=1,nonlinearity=relu)
network = layers.DenseLayer(network,num_units = 2, nonlinearity = softmax)
return network
def __build_24_net__(self):
model12 = self.subnet
network = layers.InputLayer((None, 3, 24, 24), input_var=self.__input_var__)
network = layers.Conv2DLayer(network,num_filters=16,filter_size=(5,5),stride=1,nonlinearity=relu)
network = layers.MaxPool2DLayer(network, pool_size = (3,3),stride = 2)
network = layers.DropoutLayer(network)
network = layers.DenseLayer(network,num_units = 128,nonlinearity = relu)
#network = layers.Conv2DLayer(network,num_filters=128,filter_size=(1,1),stride=1,nonlinearity=relu)
denselayer12 = model12.net.input_layer # i.e., one layer before the output layer of model12
network = layers.ConcatLayer([network, denselayer12]) # concatenate with dense layer of this model
network = layers.DenseLayer(network,num_units = 2, nonlinearity = softmax)
return network
def __build_48_net__(self):
model24 = self.subnet
network = layers.InputLayer((None, 3, 48, 48), input_var=self.__input_var__)
network = layers.Conv2DLayer(network,num_filters=64,filter_size=(5,5),stride=1,nonlinearity=relu)
network = layers.batch_norm(layers.MaxPool2DLayer(network, pool_size = (3,3),stride = 2))
network = layers.Conv2DLayer(network,num_filters=64,filter_size=(5,5),stride=1,nonlinearity=relu)
network = layers.BatchNormLayer(network)
network = layers.MaxPool2DLayer(network, pool_size = (3,3),stride = 2)
network = layers.DenseLayer(network,num_units = 256,nonlinearity = relu)
#network = layers.Conv2DLayer(network,num_filters=256,filter_size=(1,1),stride=1,nonlinearity=relu)
denselayer24 = model24.net.input_layer
network = layers.ConcatLayer([network, denselayer24])
network = layers.DenseLayer(network,num_units = 2, nonlinearity = softmax)
return network
def __build_12_calib_net__(self):
network = layers.InputLayer((None, 3, 12, 12), input_var=self.__input_var__)
network = layers.Conv2DLayer(network,num_filters=16,filter_size=(3,3),stride=1,nonlinearity=relu)
network = layers.MaxPool2DLayer(network, pool_size = (3,3),stride = 2)
network = layers.DenseLayer(network,num_units = 128,nonlinearity = relu)
#network = layers.Conv2DLayer(network,num_filters=128,filter_size=(1,1),stride=1,nonlinearity=relu)
network = layers.DenseLayer(network,num_units = 45, nonlinearity = softmax)
return network
def __build_24_calib_net__(self):
network = layers.InputLayer((None, 3, 24, 24), input_var=self.__input_var__)
network = layers.Conv2DLayer(network,num_filters=32,filter_size=(5,5),stride=1,nonlinearity=relu)
network = layers.MaxPool2DLayer(network, pool_size = (3,3),stride = 2)
network = layers.DenseLayer(network,num_units = 64,nonlinearity = relu)
#network = layers.Conv2DLayer(network,num_filters=64,filter_size=(1,1),stride=1,nonlinearity=relu)
network = layers.DenseLayer(network,num_units = 45, nonlinearity = softmax)
return network
def __build_48_calib_net__(self):
network = layers.InputLayer((None, 3, 48, 48), input_var=self.__input_var__)
network = layers.Conv2DLayer(network,num_filters=64,filter_size=(5,5),stride=1,nonlinearity=relu)
network = layers.batch_norm(layers.MaxPool2DLayer(network, pool_size = (3,3),stride = 2))
network = layers.Conv2DLayer(network,num_filters=64,filter_size=(5,5),stride=1,nonlinearity=relu)
network = layers.batch_norm(layers.MaxPool2DLayer(network, pool_size = (3,3),stride = 2))
network = layers.DenseLayer(network,num_units = 256,nonlinearity = relu)
#network = layers.Conv2DLayer(network,num_filters=2,filter_size=(1,1),stride=1,nonlinearity=relu)
network = layers.DenseLayer(network,num_units = 45, nonlinearity = softmax)
return network
def __build_loss_train__fn__(self):
# create loss function
prediction = layers.get_output(self.net)
loss = objectives.categorical_crossentropy(prediction, self.__target_var__)
loss = loss.mean() + 1e-4 * regularization.regularize_network_params(
self.net, regularization.l2)
# create parameter update expressions
params = layers.get_all_params(self.net, trainable=True)
self.eta = theano.shared(sp.array(sp.float32(0.05), dtype=sp.float32))
update_rule = updates.nesterov_momentum(loss, params, learning_rate=self.eta,
momentum=0.9)
# compile training function that updates parameters and returns training loss
if self.nn_name == '24-net':
self.__train_fn__ = theano.function([self.__input_var__,self.subnet.__input_var__, self.__target_var__], loss, updates=update_rule)
self.__predict_fn__ = theano.function([self.__input_var__,self.subnet.__input_var__], layers.get_output(self.net,deterministic=True))
elif self.nn_name == '48-net':
self.__train_fn__ = theano.function([self.__input_var__,self.subnet.__input_var__,self.subnet.subnet.__input_var__,self.__target_var__], loss, updates=update_rule)
self.__predict_fn__ = theano.function([self.__input_var__,self.subnet.__input_var__,self.subnet.subnet.__input_var__], layers.get_output(self.net,deterministic=True))
else:
self.__train_fn__ = theano.function([self.__input_var__,self.__target_var__], loss, updates=update_rule)
self.__predict_fn__ = theano.function([self.__input_var__], layers.get_output(self.net,deterministic=True))
def __init__(self,nn_name,batch_size=256,freeze=1,l_rates = sp.float32(0.05)*sp.ones(120,dtype=sp.float32),verbose = 1,subnet= None):
self.nn_name = nn_name
self.subnet = subnet
if subnet != None and freeze:
self.subnet.__freeze__()
self.batch_size = batch_size
self.verbose = verbose
self.l_rates = l_rates
self.__input_var__ = T.tensor4('X'+self.nn_name[:2])
self.__target_var__ = T.ivector('y+'+self.nn_name[:2])
self.max_epochs = self.l_rates.shape[0]
if self.nn_name == '12-net':
self.net = self.__build_12_net__()
elif self.nn_name == '24-net':
self.net = self.__build_24_net__()
elif self.nn_name == '48-net':
self.net = self.__build_48_net__()
elif self.nn_name =='12-calib_net':
self.net = self.__build_12_calib_net__()
elif self.nn_name =='24-calib_net':
self.net = self.__build_24_calib_net__()
elif self.nn_name =='48-calib_net':
self.net = self.__build_48_calib_net__()
self.__build_loss_train__fn__()
def iterate_minibatches(self,X, y, batchsize, shuffle=False):
assert len(X) == len(y)
if shuffle:
indices = sp.arange(len(X))
sp.random.shuffle(indices)
for start_idx in range(0, len(X) - batchsize + 1, batchsize):
if shuffle:
excerpt = indices[start_idx:start_idx + batchsize]
else:
excerpt = slice(start_idx, start_idx + batchsize)
yield X[excerpt], y[excerpt]
def __freeze__(self):
for layer in layers.get_all_layers(self.net):
for param in layer.params:
layer.params[param].discard('trainable')
def fit(self,X,y,X12 = None,X24 = None):
X = X.astype(sp.float32)
y = y.astype(sp.int32)
if X12 != None:
X12 = X12.astype(sp.float32)
if X24 != None:
X24 = X24.astype(sp.float32)
if self.nn_name == '24-net' :
X = [(u,v) for u,v in zip(X,X12)]
elif self.nn_name =='48-net':
X = [(u,v,q) for u,v,q in zip(X,X24,X12)]
for epoch in range(self.max_epochs):
self.eta.set_value(self.l_rates[epoch])
loss = 0
start = time()
for input_batch, target in self.iterate_minibatches(X,y,self.batch_size):
if self.nn_name == '24-net':
x = []
sx = []
for u,v in input_batch:
x.append(u)
sx.append(v)
x = sp.array(x)
sx = sp.array(sx)
loss += self.__train_fn__(x,sx,target)
elif self.nn_name == '48-net':
x = []
sx = []
ssx = []
for u,v,q in input_batch:
x.append(u)
sx.append(v)
ssx.append(q)
x = sp.array(x)
sx = sp.array(sx)
ssx = sp.array(ssx)
loss += self.__train_fn__ (x,sx,ssx ,target)
else:
loss += self.__train_fn__ (input_batch, target)
if self.verbose:
dur = time() - start
print("epoch %d out of %d \t loss %g \t time %d s" % (epoch + 1,self.max_epochs, loss / len(X),dur))
def predict(self,X,X12 = None,X24 = None):
y_pred = sp.argmax(self.predict_proba(X=X,X24=X24,X12=X12),axis=1)
return sp.array(y_pred)
def predict_proba(self,X,X12 = None,X24 = None):
proba = []
N = max(1,self.batch_size)
if X12 != None:
X12 = X12.astype(sp.float32)
if X24 != None:
X24 = X24.astype(sp.float32)
if self.nn_name == '24-net' :
X = [(u,v) for u,v in zip(X,X12)]
elif self.nn_name =='48-net':
X = [(u,v,q) for u,v,q in zip(X,X24,X12)]
for x_chunk in [X[i:i + N] for i in range(0, len(X), N)]:
if self.nn_name == '24-net':
x = []
sx = []
for u,v in x_chunk:
x.append(u)
sx.append(v)
x = sp.array(x)
sx = sp.array(sx)
chunk_proba = self.__predict_fn__(x,sx)
elif self.nn_name == '48-net':
x = []
sx = []
ssx = []
for u,v,q in x_chunk:
x.append(u)
sx.append(v)
ssx.append(q)
x = sp.array(x)
sx = sp.array(sx)
ssx = sp.array(ssx)
chunk_proba = self.__predict_fn__(x,sx,ssx)
else:
chunk_proba = self.__predict_fn__(x_chunk)
for p in chunk_proba:
proba.append(p)
return sp.array(proba)
def save_model(self,model_name = nn_name+'.pickle'):
with open(model_name, 'wb') as f:
pickle.dump(self, f, -1)
def load_model(self,model_name = nn_name+'.pickle'):
with open(model_name, 'rb') as f:
model = pickle.load(f)
f.close()
return model