/
psoCNN.py
296 lines (218 loc) · 11.9 KB
/
psoCNN.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
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
import keras
from keras.datasets import mnist
from keras.datasets import fashion_mnist
from keras.datasets import cifar10
import keras.backend
from population import Population
import numpy as np
from copy import deepcopy
class psoCNN:
def __init__(self, dataset, n_iter, pop_size, batch_size, epochs, min_layer, max_layer, \
conv_prob, pool_prob, fc_prob, max_conv_kernel, max_out_ch, max_fc_neurons, dropout_rate):
self.pop_size = pop_size
self.n_iter = n_iter
self.epochs = epochs
self.batch_size = batch_size
self.gBest_acc = np.zeros(n_iter)
self.gBest_test_acc = np.zeros(n_iter)
if dataset == "mnist":
input_width = 28
input_height = 28
input_channels = 1
output_dim = 10
(self.x_train, self.y_train), (self.x_test, self.y_test) = mnist.load_data()
if dataset == "fashion-mnist":
input_width = 28
input_height = 28
input_channels = 1
output_dim = 10
(self.x_train, self.y_train), (self.x_test, self.y_test) = fashion_mnist.load_data()
self.x_train = self.x_train.astype('float32')
self.x_test = self.x_test.astype('float32')
self.x_train /= 255
self.x_test /= 255
if dataset == "mnist-background-images":
input_width = 28
input_height = 28
input_channels = 1
output_dim = 10
train = np.loadtxt("./datasets/mnist-background-images/mnist_background_images_train.amat")
test = np.loadtxt("./datasets/mnist-background-images/mnist_background_images_test.amat")
self.x_train = train[:, :-1]
self.x_test = test[:, :-1]
# Reshape images to 28x28
self.x_train = np.reshape(self.x_train, (-1, 28, 28))
self.x_test = np.reshape(self.x_test, (-1, 28, 28))
self.y_train = train[:, -1]
self.y_test = test[:, -1]
if dataset == "mnist-rotated-digits":
input_width = 28
input_height = 28
input_channels = 1
output_dim = 10
train = np.loadtxt("./datasets/mnist-rotated-digits/mnist_all_rotation_normalized_float_train_valid.amat")
test = np.loadtxt("./datasets/mnist-rotated-digits/mnist_all_rotation_normalized_float_test.amat")
self.x_train = train[:, :-1]
self.x_test = test[:, :-1]
# Reshape images to 28x28
self.x_train = np.reshape(self.x_train, (-1, 28, 28))
self.x_test = np.reshape(self.x_test, (-1, 28, 28))
self.y_train = train[:, -1]
self.y_test = test[:, -1]
if dataset == "mnist-random-background":
input_width = 28
input_height = 28
input_channels = 1
output_dim = 10
train = np.loadtxt("./datasets/mnist-random-background/mnist_background_random_train.amat")
test = np.loadtxt("./datasets/mnist-random-background/mnist_background_random_test.amat")
self.x_train = train[:, :-1]
self.x_test = test[:, :-1]
# Reshape images to 28x28
self.x_train = np.reshape(self.x_train, (-1, 28, 28))
self.x_test = np.reshape(self.x_test, (-1, 28, 28))
self.y_train = train[:, -1]
self.y_test = test[:, -1]
if dataset == "mnist-rotated-with-background":
input_width = 28
input_height = 28
input_channels = 1
output_dim = 10
train = np.loadtxt("./datasets/mnist-rotated-with-background/mnist_all_background_images_rotation_normalized_train_valid.amat")
test = np.loadtxt("./datasets/mnist-rotated-with-background/mnist_all_background_images_rotation_normalized_test.amat")
self.x_train = train[:, :-1]
self.x_test = test[:, :-1]
# Reshape images to 28x28
self.x_train = np.reshape(self.x_train, (-1, 28, 28))
self.x_test = np.reshape(self.x_test, (-1, 28, 28))
self.y_train = train[:, -1]
self.y_test = test[:, -1]
if dataset == "rectangles":
input_width = 28
input_height = 28
input_channels = 1
output_dim = 2
train = np.loadtxt("./datasets/rectangles/rectangles_train.amat")
test = np.loadtxt("./datasets/rectangles/rectangles_test.amat")
self.x_train = train[:, :-1]
self.x_test = test[:, :-1]
# Reshape images to 28x28
self.x_train = np.reshape(self.x_train, (-1, 28, 28))
self.x_test = np.reshape(self.x_test, (-1, 28, 28))
self.y_train = train[:, -1]
self.y_test = test[:, -1]
if dataset == "rectangles-images":
input_width = 28
input_height = 28
input_channels = 1
output_dim = 2
train = np.loadtxt("./datasets/rectangles-images/rectangles_im_train.amat")
test = np.loadtxt("./datasets/rectangles-images/rectangles_im_test.amat")
self.x_train = train[:, :-1]
self.x_test = test[:, :-1]
# Reshape images to 28x28
self.x_train = np.reshape(self.x_train, (-1, 28, 28))
self.x_test = np.reshape(self.x_test, (-1, 28, 28))
self.y_train = train[:, -1]
self.y_test = test[:, -1]
if dataset == "convex":
input_width = 28
input_height = 28
input_channels = 1
output_dim = 2
train = np.loadtxt("./datasets/convex/convex_train.amat")
test = np.loadtxt("./datasets/convex/convex_test.amat")
self.x_train = train[:, :-1]
self.x_test = test[:, :-1]
# Reshape images to 28x28
self.x_train = np.reshape(self.x_train, (-1, 28, 28))
self.x_test = np.reshape(self.x_test, (-1, 28, 28))
self.y_train = train[:, -1]
self.y_test = test[:, -1]
self.x_train = self.x_train.reshape(self.x_train.shape[0], self.x_train.shape[1], self.x_train.shape[2], input_channels)
self.x_test = self.x_test.reshape(self.x_test.shape[0], self.x_test.shape[1], self.x_test.shape[2], input_channels)
self.y_train = keras.utils.to_categorical(self.y_train, output_dim)
self.y_test = keras.utils.to_categorical(self.y_test, output_dim)
print("Initializing population...")
self.population = Population(pop_size, min_layer, max_layer, input_width, input_height, input_channels, conv_prob, pool_prob, fc_prob, max_conv_kernel, max_out_ch, max_fc_neurons, output_dim)
print("Verifying accuracy of the current gBest...")
print(self.population.particle[0])
self.gBest = deepcopy(self.population.particle[0])
self.gBest.model_compile(dropout_rate)
hist = self.gBest.model_fit(self.x_train, self.y_train, batch_size=batch_size, epochs=epochs)
test_metrics = self.gBest.model.evaluate(x=self.x_test, y=self.y_test, batch_size=batch_size)
self.gBest.model_delete()
self.gBest_acc[0] = hist.history['accuracy'][-1]
self.gBest_test_acc[0] = test_metrics[1]
self.population.particle[0].acc = hist.history['accuracy'][-1]
self.population.particle[0].pBest.acc = hist.history['accuracy'][-1]
print("Current gBest acc: " + str(self.gBest_acc[0]) + "\n")
print("Current gBest test acc: " + str(self.gBest_test_acc[0]) + "\n")
print("Looking for a new gBest in the population...")
for i in range(1, self.pop_size):
print('Initialization - Particle: ' + str(i+1))
print(self.population.particle[i])
self.population.particle[i].model_compile(dropout_rate)
hist = self.population.particle[i].model_fit(self.x_train, self.y_train, batch_size=batch_size, epochs=epochs)
self.population.particle[i].acc = hist.history['accuracy'][-1]
self.population.particle[i].pBest.acc = hist.history['accuracy'][-1]
if self.population.particle[i].pBest.acc >= self.gBest_acc[0]:
print("Found a new gBest.")
self.gBest = deepcopy(self.population.particle[i])
self.gBest_acc[0] = self.population.particle[i].pBest.acc
print("New gBest acc: " + str(self.gBest_acc[0]))
test_metrics = self.gBest.model.evaluate(x=self.x_test, y=self.y_test, batch_size=batch_size)
self.gBest_test_acc[0] = test_metrics[1]
print("New gBest test acc: " + str(self.gBest_acc[0]))
self.population.particle[i].model_delete()
self.gBest.model_delete()
def fit(self, Cg, dropout_rate):
for i in range(1, self.n_iter):
gBest_acc = self.gBest_acc[i-1]
gBest_test_acc = self.gBest_test_acc[i-1]
for j in range(self.pop_size):
print('Iteration: ' + str(i) + ' - Particle: ' + str(j+1))
# Update particle velocity
self.population.particle[j].velocity(self.gBest.layers, Cg)
# Update particle architecture
self.population.particle[j].update()
print('Particle NEW architecture: ')
print(self.population.particle[j])
# Compute the acc in the updated particle
self.population.particle[j].model_compile(dropout_rate)
hist = self.population.particle[j].model_fit(self.x_train, self.y_train, batch_size=self.batch_size, epochs=self.epochs)
self.population.particle[j].model_delete()
self.population.particle[j].acc = hist.history['accuracy'][-1]
f_test = self.population.particle[j].acc
pBest_acc = self.population.particle[j].pBest.acc
if f_test >= pBest_acc:
print("Found a new pBest.")
print("Current acc: " + str(f_test))
print("Past pBest acc: " + str(pBest_acc))
pBest_acc = f_test
self.population.particle[j].pBest = deepcopy(self.population.particle[j])
if pBest_acc >= gBest_acc:
print("Found a new gBest.")
gBest_acc = pBest_acc
self.gBest = deepcopy(self.population.particle[j])
self.gBest.model_compile(dropout_rate)
hist = self.gBest.model_fit(self.x_train, self.y_train, batch_size=self.batch_size, epochs=self.epochs)
test_metrics = self.gBest.model.evaluate(x=self.x_test, y=self.y_test, batch_size=self.batch_size)
self.gBest.model_delete()
gBest_test_acc = test_metrics[1]
self.gBest_acc[i] = gBest_acc
self.gBest_test_acc[i] = gBest_test_acc
print("Current gBest acc: " + str(self.gBest_acc[i]))
print("Current gBest test acc: " + str(self.gBest_test_acc[i]))
def fit_gBest(self, batch_size, epochs, dropout_rate):
print("\nFurther training gBest model...")
self.gBest.model_compile(dropout_rate)
trainable_count = int(np.sum([keras.backend.count_params(p) for p in set(self.gBest.model.trainable_weights)]))
print("gBest's number of trainable parameters: " + str(trainable_count))
self.gBest.model_fit_complete(self.x_train, self.y_train, batch_size=batch_size, epochs=epochs)
return trainable_count
def evaluate_gBest(self, batch_size):
print("\nEvaluating gBest model on the test set...")
metrics = self.gBest.model.evaluate(x=self.x_test, y=self.y_test, batch_size=batch_size)
print("\ngBest model loss in the test set: " + str(metrics[0]) + " - Test set accuracy: " + str(metrics[1]))
return metrics