-
Notifications
You must be signed in to change notification settings - Fork 0
/
models.py
305 lines (298 loc) · 14.3 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
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
297
298
299
300
301
302
303
304
305
import tensorflow as tf
import numpy as np
from tensorflow.keras import backend as K
import datetime
import os
import json
import random
import code # code.interact(local=dict(globals(), **locals()))
# imports from keras_transformer, which was adapted in order to fit the purpose
from keras_transformer.attention import MultiHeadSelfAttention
from keras_transformer.position import TransformerCoordinateEmbedding, AddPositionalEncoding
from keras_transformer.transformer import TransformerACT, TransformerBlock
# imports from own libraries
from losses import categorical_crossentropy, categorical_crossentropy_masked, focal_loss, focal_loss_masked
from metrics import accuracy
class ARCModel:
def __init__(self, config):
#
self.config = config
#
if not self.config['is_loaded']:
self.config['name'] = datetime.datetime.now().strftime("%Y_%m_%d__%H_%M_%S")
os.makedirs('saved_models/' + self.config['name'])
os.makedirs('saved_models/' + self.config['name'] + '/weights')
os.makedirs('saved_models/' + self.config['name'] + '/logs')
#
self.loss = categorical_crossentropy
#
self.optimizer = tf.keras.optimizers.SGD(learning_rate=0.0)
#
self.model = tf.keras.Sequential()
if self.config['model_type'] == 'conv':
size = 4
self.model.add(tf.keras.layers.Conv2D(size * 32, 3, activation='relu', padding='same'))
self.model.add(tf.keras.layers.Conv2D(size * 32, 3, activation='relu', padding='same'))
self.model.add(tf.keras.layers.Conv2D(size * 32, 3, activation='relu', padding='same'))
self.model.add(tf.keras.layers.Conv2D(size * 32, 3, activation='relu', padding='same', strides=(1, 2)))
self.model.add(tf.keras.layers.LayerNormalization(trainable=True))
#
self.model.add(tf.keras.layers.Conv2D(size * 64, 3, activation='relu', padding='same'))
self.model.add(tf.keras.layers.Conv2D(size * 64, 3, activation='relu', padding='same'))
self.model.add(tf.keras.layers.Conv2D(size * 64, 3, activation='relu', padding='same'))
self.model.add(tf.keras.layers.Conv2D(size * 64, 3, activation='relu', padding='same', strides=(2, 1)))
self.model.add(tf.keras.layers.LayerNormalization(trainable=True))
# transformer layer
self.model.add(tf.keras.layers.Conv2DTranspose(size * 64, 3, activation='relu', padding='same'))
self.model.add(tf.keras.layers.Conv2DTranspose(size * 64, 3, activation='relu', padding='same'))
self.model.add(tf.keras.layers.Conv2DTranspose(size * 64, 3, activation='relu', padding='same'))
self.model.add(tf.keras.layers.Conv2DTranspose(size * 64, 3, activation='relu', padding='same', strides=(2, 1)))
self.model.add(tf.keras.layers.LayerNormalization(trainable=True))
#
self.model.add(tf.keras.layers.Conv2DTranspose(size * 32, 3, activation='relu', padding='same'))
self.model.add(tf.keras.layers.Conv2DTranspose(size * 32, 3, activation='relu', padding='same'))
self.model.add(tf.keras.layers.Conv2DTranspose(size * 32, 3, activation='relu', padding='same'))
self.model.add(tf.keras.layers.Conv2DTranspose(size * 32, 3, activation='relu', padding='same', strides=(1, 2)))
self.model.add(tf.keras.layers.LayerNormalization(trainable=True))
#
self.model.add(tf.keras.layers.Conv2DTranspose(11, 1, activation='softmax', padding='same'))
elif self.config['model_type'] == 'transformer':
size = 4
self.model.add(tf.keras.layers.Conv2D(size * 32, 3, activation='relu', padding='same'))
self.model.add(tf.keras.layers.Conv2D(size * 32, 3, activation='relu', padding='same', strides=(1, 2)))
self.model.add(tf.keras.layers.LayerNormalization(trainable=True))
#
self.model.add(tf.keras.layers.Conv2D(size * 64, 3, activation='relu', padding='same'))
self.model.add(tf.keras.layers.Conv2D(size * 64, 3, activation='relu', padding='same', strides=(2, 1)))
self.model.add(tf.keras.layers.LayerNormalization(trainable=True))
# transformer layer
self.model.add(tf.keras.layers.Reshape([15 * 15, size * 64]))
#
self.model.add(TransformerBlock('Transformer1', num_heads=8, use_masking=False))
#self.model.add(TransformerBlock('Transformer2', num_heads=8, use_masking=False))
#self.model.add(TransformerBlock('Transformer3', num_heads=8, use_masking=False))
#self.model.add(TransformerBlock('Transformer4', num_heads=8, use_masking=False))
self.model.add(tf.keras.layers.Reshape([15, 15, size * 64]))
#
self.model.add(tf.keras.layers.Conv2DTranspose(size * 64, 3, activation='relu', padding='same'))
self.model.add(tf.keras.layers.Conv2DTranspose(size * 64, 3, activation='relu', padding='same', strides=(2, 1)))
self.model.add(tf.keras.layers.LayerNormalization(trainable=True))
#
self.model.add(tf.keras.layers.Conv2DTranspose(size * 32, 3, activation='relu', padding='same'))
self.model.add(tf.keras.layers.Conv2DTranspose(size * 32, 3, activation='relu', padding='same', strides=(1, 2)))
self.model.add(tf.keras.layers.LayerNormalization(trainable=True))
#
self.model.add(tf.keras.layers.Conv2DTranspose(11, 1, activation='softmax', padding='same'))
elif self.config['model_type'] == 'fc':
self.model = tf.keras.Sequential()
self.model.add(tf.keras.layers.Reshape([30 * 30 * 11]))
self.model.add(tf.keras.layers.Dense(30 * 30 * 11, activation='tanh'))
self.model.add(tf.keras.layers.Dense(30 * 30 * 11))
self.model.add(tf.keras.layers.Reshape([30, 30, 11]))
self.model.add(tf.keras.layers.Activation(tf.keras.activations.softmax))
#
self.model.compile(optimizer=self.optimizer, loss=categorical_crossentropy)
self.model.build(input_shape=[None,30,30,11])
#
if self.config['is_loaded']:
self.model.load_weights('saved_models/' + self.config['name'] + '/weights/' + str(self.config['current_epoch']))
else:
open('saved_models/' + self.config['name'] + '/config.json', 'w').write(json.dumps(self.config, sort_keys=True, indent=4))
self.model.save_weights('saved_models/' + self.config['name'] + '/weights/' + str(self.config['current_epoch']))
# e.g. train_on_one_example(train_data['x'][48], train_data['y'][48])
def train_on_one_example(self, x, y):
# self.model.fit(x, y, epochs=1000, verbose=1)
learning_rate = self.config['learning_rate']
ratio_correct = None
for i in range(1000):
print(i)
if i % 10 == 0:
print('decrease learning rate!')
learning_rate *= 0.5
learning_rate = max(learning_rate, 0.0001)
# loss = categorical_crossentropy
loss = categorical_crossentropy
#
list_of_variables = self.model.trainable_variables
print('loss: ' + str(float(loss(y, self.model(x)))))
with tf.GradientTape() as tape:
grads = tape.gradient(loss(y, self.model(x)), list_of_variables)
num_grads_zero = 0
num_grads_not_zero = 0
num_grads_nan = 0
num_grads_not_nan = 0
pre_update = []
new_weights = []
for i in range(len(self.model.trainable_weights)):
pre_update.append(np.array(self.model.trainable_weights[i][:]))
num_grads_zero += np.sum(grads[i] == 0)
num_grads_not_zero += np.sum(grads[i] != 0)
num_grads_nan += np.sum(np.isnan(grads[i]))
num_grads_not_nan += (np.size(grads[i]) - np.sum(np.isnan(grads[i])))
# list_of_variables[i].assign_sub(0.001 * grads[i])
new_weights.append(np.array(pre_update[i] - self.config['learning_rate'] * grads[i]))
self.model.set_weights(new_weights)
print('num_grads_zero: ' + str(num_grads_zero) + ', num_grads_not_zero:' + str(num_grads_not_zero))
print('num_grads_nan: ' + str(num_grads_nan) + ', num_grads_not_nan:' + str(num_grads_not_nan))
# code.interact(local=dict(globals(), **locals()))
#
post_update = self.model.trainable_weights
is_equal = 0
is_different = 0
for i in range(len(pre_update)):
is_equal += np.sum(pre_update[i] == post_update[i])
is_different += np.sum(pre_update[i] != post_update[i])
print('is_equal: ' + str(is_equal) + ', is_different:' + str(is_different))
ratio_correct, ratio_total, ratio_invalid = accuracy(self.model, x, y, self.config)
print('')
ratio_correct, ratio_total, ratio_invalid = accuracy(self.model, x, y, self.config)
print('ratio correct: ' + str(ratio_correct))
#print('in train_on_one_example')
#code.interact(local=dict(globals(), **locals()))
return ratio_correct, ratio_total, ratio_invalid
def unit_test(self, x_list, y_list):
ratio_correct_list = []
random_list = []
for i in range(self.config['num_unit_tests']):
n = len(x_list)
s = int(random.uniform(0,n))
random_list.append(s)
x = x_list[s]
y = y_list[s]
ratio_correct_list.append(self.train_on_one_example(x, y))
def train(self, train_data, val_data):
meta_weights = []
for i in range(len(self.model.trainable_weights)):
meta_weights.append(np.array(self.model.trainable_weights[i][:]))
for epoch in range(self.config['current_epoch'], self.config['max_epochs']):
print('epoch: ' + str(epoch))
meta_grads = []
current_train_losses = []
current_train_accuracies = []
current_train_total_accuracies = []
current_val_losses = []
current_val_accuracies = []
current_val_total_accuracies = []
for sample_nr in range(len(train_data['x'])): #for sample_nr in range(self.config['meta_batch_size']):
print(sample_nr)
# code.interact(local=dict(globals(), **locals()))
self.model.set_weights(meta_weights)
loss = categorical_crossentropy
x = train_data['x'][sample_nr]
y = train_data['y'][sample_nr]
x_val = train_data['x_val'][sample_nr]
y_val = train_data['y_val'][sample_nr]
meta_grads.append([])
#
list_of_variables = self.model.trainable_variables
current_loss = float(loss(y, self.model(x)))
with tf.GradientTape() as tape:
grads = tape.gradient(loss(y, self.model(x)), list_of_variables)
num_grads_zero = 0
num_grads_not_zero = 0
pre_update = []
new_weights = []
for i in range(len(self.model.trainable_weights)):
pre_update.append(np.array(self.model.trainable_weights[i][:]))
current_grad = np.array(grads[i])
#meta_grads[-1].append(current_grad)
num_grads_zero += np.sum(current_grad == 0)
num_grads_not_zero += np.sum(current_grad != 0)
new_weights.append(pre_update[i] - self.config['learning_rate'] * current_grad)
del(grads)
self.model.set_weights(new_weights)
#
post_update = self.model.trainable_weights
is_equal = 0
is_different = 0
for i in range(len(pre_update)):
is_equal += np.sum(pre_update[i] == post_update[i])
is_different += np.sum(pre_update[i] != post_update[i])
ratio_correct_val, ratio_total_val, ratio_invalid_val = accuracy(self.model, x_val, y_val, self.config)
meta_loss = float(loss(y_val, self.model(x_val)))
#
current_train_losses.append(meta_loss)
current_train_accuracies.append(ratio_correct_val)
current_train_total_accuracies.append(ratio_total_val)
#
with tf.GradientTape() as tape:
meta_grad = tape.gradient(loss(y_val, self.model(x_val)), list_of_variables)
for i in range(len(meta_weights)):
meta_grads[-1].append(np.array(meta_grad[i]))
del(meta_grad)
# apply meta step
if len(meta_grads) >= self.config['meta_batch_size'] or sample_nr == len(train_data['x']) - 1:
print('meta update')
for i in range(len(meta_weights)):
current_grad = meta_grads[0][i]
for j in range(1,len(meta_grads)):
current_grad += meta_grads[j][i]
current_grad /= len(meta_grads)
meta_weights[i] = meta_weights[i] - self.config['meta_learning_rate'] * current_grad
meta_grads = []
#
if self.config['verbosity'] >= 1:
ratio_correct, ratio_total, ratio_invalid = accuracy(self.model, x, y, self.config)
print('loss: ' + str(current_loss))
print('num_grads_zero: ' + str(num_grads_zero) + ', num_grads_not_zero:' + str(num_grads_not_zero))
print('is_equal: ' + str(is_equal) + ', is_different:' + str(is_different))
print('meta_loss: ' + str(meta_loss))
print('ratio_correct_meta: ' + str(ratio_correct_val))
print('ratio_total_meta: ' + str(ratio_total_val))
print('')
print('')
#
for sample_nr in range(len(val_data['x'])): #for sample_nr in range(self.config['meta_batch_size']):
print(sample_nr)
# code.interact(local=dict(globals(), **locals()))
self.model.set_weights(meta_weights)
loss = categorical_crossentropy
x = train_data['x'][sample_nr]
y = train_data['y'][sample_nr]
x_val = train_data['x_val'][sample_nr]
y_val = train_data['y_val'][sample_nr]
meta_grads.append([])
#
list_of_variables = self.model.trainable_variables
print('loss: ' + str(float(loss(y, self.model(x)))))
with tf.GradientTape() as tape:
grads = tape.gradient(loss(y, self.model(x)), list_of_variables)
num_grads_zero = 0
num_grads_not_zero = 0
pre_update = []
new_weights = []
for i in range(len(self.model.trainable_weights)):
pre_update.append(np.array(self.model.trainable_weights[i][:]))
current_grad = np.array(grads[i])
#meta_grads[-1].append(current_grad)
num_grads_zero += np.sum(current_grad == 0)
num_grads_not_zero += np.sum(current_grad != 0)
new_weights.append(pre_update[i] - self.config['learning_rate'] * current_grad)
del(grads)
self.model.set_weights(new_weights)
#
post_update = self.model.trainable_weights
is_equal = 0
is_different = 0
for i in range(len(pre_update)):
is_equal += np.sum(pre_update[i] == post_update[i])
is_different += np.sum(pre_update[i] != post_update[i])
ratio_correct_val, ratio_total_val, ratio_invalid_val = accuracy(self.model, x_val, y_val, self.config)
#
current_train_losses.append(meta_loss)
current_train_accuracies.append(ratio_correct_val)
current_train_total_accuracies.append(ratio_total_val)
#
if self.config['verbosity'] >= 1:
ratio_correct, ratio_total, ratio_invalid = accuracy(self.model, x, y, self.config)
print('num_grads_zero: ' + str(num_grads_zero) + ', num_grads_not_zero:' + str(num_grads_not_zero))
print('is_equal: ' + str(is_equal) + ', is_different:' + str(is_different))
print('meta_loss: ' + str(float(loss(y_val, self.model(x_val)))))
print('ratio_correct_val: ' + str(ratio_correct_val))
print('ratio_total_val: ' + str(ratio_total_val))
print('')
print('')
#
self.config['current_epoch'] += 1
open('saved_models/' + self.config['name'] + '/config.json', 'w').write(json.dumps(self.config, sort_keys=True, indent=4))
self.model.save_weights('saved_models/' + self.config['name'] + '/weights/' + str(self.config['current_epoch']))