-
Notifications
You must be signed in to change notification settings - Fork 3
/
01_conv_network.py
260 lines (223 loc) · 10.9 KB
/
01_conv_network.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
from multiprocessing import freeze_support
import matplotlib.pyplot as plt
import numpy as np
import scipy.interpolate
import scipy.ndimage.filters
import dataset.cifar10_dataset
from network import activation, weight_initializer
from network.layers.conv_to_fully_connected import ConvToFullyConnected
from network.layers.convolution_im2col import Convolution
from network.layers.dropout import Dropout
from network.layers.fully_connected import FullyConnected
from network.layers.max_pool import MaxPool
from network.model import Model
from network.optimizer import GDMomentumOptimizer
if __name__ == '__main__':
"""
------------------------------------
DFA
------------------------------------
Result:
------------------------------------
loss on test set: 1.5317771647628007
accuracy on test set: 0.4463
Train statisistics:
------------------------------------
time spend during forward pass: 621.903642654419
time spend during backward pass: 90.21121907234192
time spend during update pass: 1.5559537410736084
time spend in total: 788.1902389526367
------------------------------------
BP
------------------------------------
Result:
------------------------------------
loss on test set: 1.0581702546107525
accuracy on test set: 0.6258
Train statisistics:
------------------------------------
time spend during forward pass: 575.9978523254395
time spend during backward pass: 269.75944471359253
time spend during update pass: 1.299713134765625
time spend in total: 918.6962807178497
"""
freeze_support()
num_iteration = 30
data = dataset.cifar10_dataset.load()
# layers = [
# Convolution((32, 3, 3, 3), stride=1, padding=1, dropout_rate=0, activation=activation.tanh),
# Convolution((32, 32, 3, 3), stride=1, padding=1, dropout_rate=0, activation=activation.tanh),
# MaxPool(size=2, stride=2),
# Dropout(0.2),
# Convolution((64, 32, 3, 3), stride=1, padding=1, dropout_rate=0, activation=activation.tanh),
# MaxPool(size=2, stride=2),
# Dropout(0.3),
# Convolution((128, 64, 3, 3), stride=1, padding=1, dropout_rate=0, activation=activation.tanh),
# MaxPool(size=2, stride=2),
# Dropout(0.4),
# ConvToFullyConnected(),
# FullyConnected(size=512, activation=activation.tanh),
# FullyConnected(size=10, activation=None, last_layer=True)
# ]
layers = [
MaxPool(size=2, stride=2),
Convolution((16, 3, 3, 3), stride=1, padding=1, dropout_rate=0, activation=activation.tanh),
MaxPool(size=2, stride=2),
Convolution((16, 16, 3, 3), stride=1, padding=1, dropout_rate=0, activation=activation.tanh),
MaxPool(size=2, stride=2),
Convolution((32, 16, 3, 3), stride=1, padding=1, dropout_rate=0, activation=activation.tanh),
MaxPool(size=2, stride=2),
ConvToFullyConnected(),
FullyConnected(size=64, activation=activation.tanh),
FullyConnected(size=10, activation=None, last_layer=True)
]
layers = [
# MaxPool(size=2, stride=2),
Convolution((8, 3, 4, 4), stride=2, padding=2, dropout_rate=0, activation=activation.tanh),
#MaxPool(size=2, stride=2),
Convolution((16, 8, 3, 3), stride=2, padding=1, dropout_rate=0, activation=activation.tanh),
#MaxPool(size=2, stride=2),
Convolution((32, 16, 3, 3), stride=2, padding=1, dropout_rate=0, activation=activation.tanh),
#MaxPool(size=2, stride=2),
ConvToFullyConnected(),
FullyConnected(size=64, activation=activation.tanh),
FullyConnected(size=10, activation=None, last_layer=True)
]
# -------------------------------------------------------
# Train with BP
# -------------------------------------------------------
model = Model(
layers=layers,
num_classes=10,
optimizer=GDMomentumOptimizer(lr=1e-2, mu=0.9),
lr_decay=0.5,
lr_decay_interval=7
)
print("\nRun training:\n------------------------------------")
stats_bp = model.train(data_set=data, method='bp', num_passes=num_iteration, batch_size=64)
loss, accuracy = model.cost(*data.test_set())
print("\nResult:\n------------------------------------")
print('loss on test set: {}'.format(loss))
print('accuracy on test set: {}'.format(accuracy))
print("\nTrain statisistics:\n------------------------------------")
print("time spend during forward pass: {}".format(stats_bp['forward_time']))
print("time spend during backward pass: {}".format(stats_bp['backward_time']))
print("time spend during update pass: {}".format(stats_bp['update_time']))
print("time spend in total: {}".format(stats_bp['total_time']))
# plt.title('Loss function')
# plt.xlabel('epoch')
# plt.ylabel('loss')
# plt.plot(np.arange(len(stats_bp['train_loss'])), stats_bp['train_loss'])
# plt.plot(stats_bp['valid_step'], stats_bp['valid_loss'])
# plt.legend(['train loss bp', 'validation loss bp'], loc='upper right')
# plt.grid(True)
# plt.show()
# plt.title('Accuracy')
# plt.xlabel('epoch')
# plt.ylabel('accuracy')
# plt.plot(np.arange(len(stats_bp['train_accuracy'])), stats_bp['train_accuracy'])
# plt.plot(stats_bp['valid_step'], stats_bp['valid_accuracy'])
# plt.legend(['train accuracy bp', 'validation accuracy bp'], loc='lower right')
# plt.grid(True)
# plt.show()
# exit()
# -------------------------------------------------------
# Train with DFA
# -------------------------------------------------------
model = Model(
layers=layers,
num_classes=10,
optimizer=GDMomentumOptimizer(lr=1e-2, mu=0.9),
# regularization=0.001,
# lr_decay=0.5,
# lr_decay_interval=3
# layers=layers,
# num_classes=10,
# optimizer=GDMomentumOptimizer(lr=1e-3, mu=0.9),
# regularization=0.001,
lr_decay=0.5,
lr_decay_interval=7
)
print("\nRun training:\n------------------------------------")
stats_dfa = model.train(data_set=data, method='dfa', num_passes=num_iteration, batch_size=64)
loss, accuracy = model.cost(*data.test_set())
print("\nResult:\n------------------------------------")
print('loss on test set: {}'.format(loss))
print('accuracy on test set: {}'.format(accuracy))
print("\nTrain statisistics:\n------------------------------------")
print("time spend during forward pass: {}".format(stats_dfa['forward_time']))
print("time spend during backward pass: {}".format(stats_dfa['backward_time']))
print("time spend during update pass: {}".format(stats_dfa['update_time']))
print("time spend in total: {}".format(stats_dfa['total_time']))
# plt.title('Loss function')
# plt.xlabel('epoch')
# plt.ylabel('loss')
# plt.plot(np.arange(len(stats_dfa['train_loss'])), stats_dfa['train_loss'])
# plt.plot(stats_dfa['valid_step'], stats_dfa['valid_loss'])
# plt.legend(['train loss dfa', 'validation loss dfa'], loc='upper right')
# plt.grid(True)
# plt.show()
# plt.title('Accuracy')
# plt.xlabel('epoch')
# plt.ylabel('accuracy')
# plt.plot(np.arange(len(stats_dfa['train_accuracy'])), stats_dfa['train_accuracy'])
# plt.plot(stats_dfa['valid_step'], stats_dfa['valid_accuracy'])
# plt.legend(['train accuracy dfa', 'validation accuracy dfa'], loc='lower right')
# plt.grid(True)
# plt.show()
# exit()
plt.title('Loss function')
plt.xlabel('epoch')
plt.ylabel('loss')
dfa_train_loss = scipy.ndimage.filters.gaussian_filter1d(stats_dfa['train_loss'], sigma=10)
bp_train_loss = scipy.ndimage.filters.gaussian_filter1d(stats_bp['train_loss'], sigma=10)
plt.plot(np.arange(len(stats_dfa['train_loss'])), dfa_train_loss)
plt.plot(stats_dfa['valid_step'], stats_dfa['valid_loss'])
plt.plot(np.arange(len(stats_bp['train_loss'])), bp_train_loss)
plt.plot(stats_bp['valid_step'], stats_bp['valid_loss'])
plt.legend(['train loss dfa', 'validation loss dfa', 'train loss bp', 'validation loss bp'], loc='upper right')
plt.grid(True)
plt.show()
plt.title('Accuracy')
plt.xlabel('epoch')
plt.ylabel('accuracy')
dfa_train_accuracy = scipy.ndimage.filters.gaussian_filter1d(stats_dfa['train_accuracy'], sigma=10)
bp_train_accuracy = scipy.ndimage.filters.gaussian_filter1d(stats_bp['train_accuracy'], sigma=10)
plt.plot(np.arange(len(stats_dfa['train_accuracy'])), dfa_train_accuracy)
plt.plot(stats_dfa['valid_step'], stats_dfa['valid_accuracy'])
plt.plot(np.arange(len(stats_bp['train_accuracy'])), bp_train_accuracy)
plt.plot(stats_bp['valid_step'], stats_bp['valid_accuracy'])
plt.legend(['train accuracy dfa', 'validation accuracy dfa', 'train accuracy bp', 'validation accuracy bp'], loc='lower right')
plt.grid(True)
plt.show()
# Forward, regularization, update and validation passes are excactly the same operations for dfa and bp. Therefore
# they should take euqally long. To ensure that inequalities don't affect the result, we normalize the time here.
# The reference time is the one measured for bp.
total_time_bp = stats_bp['total_time']
total_time_dfa = total_time_bp - stats_bp['backward_time'] + stats_dfa['backward_time']
step_to_time_bp = total_time_bp / len(stats_bp['train_loss'])
step_to_time_dfa = step_to_time_bp * total_time_dfa / stats_bp['total_time']
plt.title('Loss vs time')
plt.xlabel('time')
plt.ylabel('loss')
dfa_train_loss = scipy.ndimage.filters.gaussian_filter1d(stats_dfa['train_loss'], sigma=10)
bp_train_loss = scipy.ndimage.filters.gaussian_filter1d(stats_bp['train_loss'], sigma=10)
plt.plot(np.arange(len(stats_dfa['train_loss'])) * step_to_time_dfa, dfa_train_loss)
plt.plot(np.asarray(stats_dfa['valid_step']) * step_to_time_dfa, stats_dfa['valid_loss'])
plt.plot(np.arange(len(stats_bp['train_loss'])) * step_to_time_bp, bp_train_loss)
plt.plot(np.asarray(stats_bp['valid_step']) * step_to_time_bp, stats_bp['valid_loss'])
plt.legend(['train loss dfa', 'validation loss dfa', 'train loss bp', 'validation loss bp'], loc='best')
plt.grid(True)
plt.show()
plt.title('Accuracy vs time')
plt.xlabel('time')
plt.ylabel('accuracy')
dfa_train_accuracy = scipy.ndimage.filters.gaussian_filter1d(stats_dfa['train_accuracy'], sigma=10)
bp_train_accuracy = scipy.ndimage.filters.gaussian_filter1d(stats_bp['train_accuracy'], sigma=10)
plt.plot(np.arange(len(stats_dfa['train_accuracy'])) * step_to_time_dfa, dfa_train_accuracy)
plt.plot(np.asarray(stats_dfa['valid_step']) * step_to_time_dfa, stats_dfa['valid_accuracy'])
plt.plot(np.arange(len(stats_bp['train_accuracy'])) * step_to_time_bp, bp_train_accuracy)
plt.plot(np.asarray(stats_bp['valid_step']) * step_to_time_bp, stats_bp['valid_accuracy'])
plt.legend(['train accuracy dfa', 'validation accuracy dfa', 'train accuracy bp', 'validation accuracy bp'], loc='lower right')
plt.grid(True)
plt.show()