/
keras_convnet.py
213 lines (185 loc) · 8.99 KB
/
keras_convnet.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
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import os
import sys
from glob import glob
import time
import random
import numpy as np
import scipy.stats
from keras.models import Sequential
from keras.layers.core import Dense, Activation, Flatten
from keras.layers.convolutional import Convolution2D, MaxPooling2D
from keras.layers.normalization import BatchNormalization
from keras.regularizers import l1, l2, l1l2, activity_l1, activity_l2, activity_l1l2
from keras.optimizers import SGD
from keras import backend as K
def build_model(model):
regularize=0.01
## layer 1 -- input is about 600 x 800
ch01 = 16
xx01 = 12
yy01 = 12
kern_W_init = (0.06/2.0)*scipy.stats.truncnorm.rvs(-2.0, 2.0, size=(ch01, 1, xx01, yy01)).astype(np.float32)
kern_B_init = np.zeros(ch01, dtype=np.float32)
model.add(Convolution2D(ch01, xx01, yy01, border_mode='same',
weights=[kern_W_init, kern_B_init],
W_regularizer=l2(regularize),
input_shape=datareader.features_placeholder_shape()[1:]))
model.add(Activation('relu'))
model.add(MaxPooling2D(pool_size=(3,3))) #, strides=(4,4)))
model.add(BatchNormalization(epsilon=1e-06,
mode=0,
axis=1, momentum=0.9, weights=None,
beta_init='zero', gamma_init='one'))
## layer 2 input is about 200 x 270
ch02 = 16
xx02 = 10
yy02 = 10
kern_W_init = (0.06/2.0)*scipy.stats.truncnorm.rvs(-2.0, 2.0, size=(ch02, ch01, xx02, yy02)).astype(np.float32)
kern_B_init = np.zeros(ch02, dtype=np.float32)
model.add(Convolution2D(ch02, xx02, yy02,
border_mode='same',
W_regularizer=l2(regularize),
weights=[kern_W_init, kern_B_init]))
model.add(Activation('relu'))
model.add(MaxPooling2D(pool_size=(5,5)))#, strides=(4,4)))
model.add(BatchNormalization(epsilon=1e-06,
mode=0,
axis=1, momentum=0.9, weights=None,
beta_init='zero', gamma_init='one'))
## layer 3 input is about 50 x 60
ch03 = 16
xx03 = 8
yy03 = 8
kern_W_init = (0.06/2.0)*scipy.stats.truncnorm.rvs(-2.0, 2.0, size=(ch03, ch02, xx03, yy03)).astype(np.float32)
kern_B_init = np.zeros(ch03,dtype=np.float32)
model.add(Convolution2D(ch03, xx03, yy03,
border_mode='same',
W_regularizer=l2(regularize),
weights=[kern_W_init, kern_B_init]))
model.add(Activation('relu'))
model.add(MaxPooling2D(pool_size=(5,5)))#, strides=(4,4)))
model.add(BatchNormalization(epsilon=1e-06,
mode=0,
axis=1, momentum=0.9, weights=None,
beta_init='zero', gamma_init='one'))
model.add(Flatten())
## layer 4 input is about 12 x 15
hidden04 = 48
W_init = (0.06/2.0)*scipy.stats.truncnorm.rvs(-2.0, 2.0, size=(1008,hidden04)).astype(np.float32)
B_init = np.zeros(hidden04, dtype=np.float32)
model.add(Dense(hidden04,
W_regularizer=l2(regularize),
weights=[W_init, B_init]))
model.add(Activation('relu'))
model.add(BatchNormalization(epsilon=1e-06,
mode=0,
axis=1, momentum=0.9, weights=None,
beta_init='zero', gamma_init='one'))
## layer 5
hidden05 = 32
W_init = (0.06/2.0)*scipy.stats.truncnorm.rvs(-2.0, 2.0, size=(hidden04, hidden05)).astype(np.float32)
B_init = np.zeros(hidden05,dtype=np.float32)
model.add(Dense(hidden05,
W_regularizer=l2(regularize),
weights=[W_init, B_init]))
model.add(Activation('relu'))
model.add(BatchNormalization(epsilon=1e-06,
mode=0,
axis=1, momentum=0.9, weights=None,
beta_init='zero', gamma_init='one'))
## layer 6
W_init = (0.06/2.0)*scipy.stats.truncnorm.rvs(-2.0, 2.0, size=(hidden05, datareader.num_outputs())).astype(np.float32)
B_init = np.zeros(datareader.num_outputs(),dtype=np.float32)
model.add(Dense(datareader.num_outputs(),
weights=[W_init, B_init],
W_regularizer=l2(0.2*regularize)))
model.add(Activation('softmax'))
def set_learning_rate(model, step_number):
min_learning_rate = 0.001
steps_per_drop = 100
drop_by = 0.98
lr = model.optimizer.lr.get_value()
if lr <= min_learning_rate:
return lr
if step_number % steps_per_drop == (steps_per_drop-1):
lr *= drop_by
K.set_value(model.optimizer.lr, lr)
return lr
def get_confusion_matrix_one_hot(model_results, truth):
assert model_results.shape == truth.shape
num_outputs = truth.shape[1]
cmat = np.zeros((num_outputs, num_outputs), dtype=np.int32)
predictions = np.argmax(model_results,axis=1)
assert len(predictions)==truth.shape[0]
for actual_class in range(num_outputs):
idx_examples_this_class = truth[:,actual_class]==1
prediction_for_this_class = predictions[idx_examples_this_class]
for predicted_class in range(num_outputs):
count = np.sum(prediction_for_this_class==predicted_class)
cmat[actual_class, predicted_class] = count
assert np.sum(cmat)==len(truth)
return cmat
def eval_report(epoch, batch, step_times, lr, train_loss, train_accuracy, model, train_features, train_labels, validation_features, validation_labels):
t0 = time.time()
sec_per_step = np.sum(step_times)/len(step_times)
validation_predict = model.predict(validation_features)
confusion_matrix_validation = get_confusion_matrix_one_hot(validation_predict, validation_labels)
eval_accuracy = np.trace(confusion_matrix_validation)/np.sum(confusion_matrix_validation)
train_predict = model.predict(train_features)
confusion_matrix_train = get_confusion_matrix_one_hot(train_predict, train_labels)
train_accuracy_from_confusion_matrix = np.trace(confusion_matrix_train)/np.sum(confusion_matrix_train)
def fmt3(x): return '%3d' % x
cmat_tr_row0 = ' '.join(map(fmt3, confusion_matrix_train[0,:]))
cmat_ev_row0 = ' '.join(map(fmt3, confusion_matrix_validation[0,:]))
eval_time = time.time()-t0
print(" %2d:%4d |%5.1f |%8.2f |%7.2f/%6.2f |%7.2f |%8.5f | %s | %s | %5.1f" %
(epoch, batch, sec_per_step,
train_loss, train_accuracy,
train_accuracy_from_confusion_matrix,
eval_accuracy,
lr, cmat_tr_row0, cmat_ev_row0, eval_time))
for row in range(1,confusion_matrix_train.shape[0]):
print(" %s | %s | %s |" % (' '*59,
' '.join(map(fmt3, confusion_matrix_train[row,:])),
' '.join(map(fmt3, confusion_matrix_validation[row,:]))))
print("-"*99)
def run(datareader):
start_time = time.time()
batches_per_epoch = datareader.batches_per_epoch()
print("keras_convnet: datareader - %d batches per epoch" % batches_per_epoch)
validation_features, validation_labels = datareader.get_validation_set()
print("starting to build and compile keras/theano model...")
sys.stdout.flush()
t0 = time.time()
model = Sequential()
build_model(model)
model.load_weights("keras_convnet_6layer.h5")
sgd = SGD(lr=0.01, momentum=0.92)
model.compile(loss='categorical_crossentropy', optimizer=sgd)
print("building/compiling theano model took %.2f sec" % (time.time()-t0),)
sys.stdout.flush()
print(" epch:mb | tm.s | loss.tr | acc.tr/ batch | acc.ev | learn.rt| confuse mat tr | confuse mat eva | tm.e")
eval_step_interval = 100
num_steps = 8001
step_times = []
for step_number in range(1,num_steps+1):
epoch = step_number // batches_per_epoch
batch = step_number % batches_per_epoch
t0 = time.time()
train_features, train_labels = datareader.get_next_minibatch()
lr = set_learning_rate(model, step_number)
train_loss, train_accuracy = model.train_on_batch(train_features, train_labels, accuracy=True)
step_time = time.time()-t0
step_times.append(step_time)
if step_number % eval_step_interval == 0:
eval_report(epoch, batch, step_times, lr, train_loss, train_accuracy, model, train_features, train_labels, validation_features, validation_labels)
step_times = []
# print("model has %d params" % model.count_params())
model.summary()
model.save_weights("keras_convnet_6layer_B.h5", overwrite=True)
print("total time: %.2f" % (time.time()-start_time))
assert datareader is not None, 'datareader is not defined' # in globals() nor locals()'
run(datareader)