-
Notifications
You must be signed in to change notification settings - Fork 7
/
train.py
195 lines (148 loc) · 5.08 KB
/
train.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
# use larger learning rate for PTB
import numpy as np
from layers import *
from network_tools import *
from climin import RmsProp, Adam, GradientDescent
from text import loader
import time
import pickle
import os
import matplotlib.pyplot as plt
plt.ion()
plt.style.use('kosh')
plt.figure(figsize=(12, 7))
plt.clf()
np.random.seed(np.random.randint(1213))
experiment_name = 'lstm_dropout'
text_file = 'ptb.txt'
periods = [1, 2, 10, 100]
vocabulary_size = 49
states = 512 # per clock
output = 512 # for all clocks
dinput = vocabulary_size
doutputs = vocabulary_size
sequence_length = 30
batch_size = 20
learning_rate = 5e-3
niterations = 10000
momentum = 0.9
dropout = 0.1
forget_every = 100
gradient_clip = (-1.0, 1.0)
sample_every = 1000
save_every = 1000
plot_every = 100
full_recurrence = False
learn_state = False
anneal = True
dynamic_forgetting = False
dynamic_mom = False
logs = {}
data = loader('data/' + text_file, sequence_length, batch_size)
def dW(W):
load_weights(model, W)
input, target = data.fetch()
output = forward(model, input)
backward(model, target)
gradients = extract_grads(model)
clipped_gradients = np.clip(gradients, gradient_clip[0], gradient_clip[1])
loss = -1.0 * np.sum(target * np.log(output)) / (sequence_length * batch_size)
gradient_norm = (gradients ** 2).sum() / gradients.size
clipped_gradient_norm = (clipped_gradients ** 2).sum() / gradients.size
logs['loss'].append(loss)
logs['smooth_loss'].append(loss * 0.01 + logs['smooth_loss'][-1] * 0.99)
logs['gradient_norm'].append(gradient_norm)
logs['clipped_gradient_norm'].append(clipped_gradient_norm)
return clipped_gradients
os.system('mkdir results/' + experiment_name)
path = 'results/' + experiment_name + '/'
logs['loss'] = []
logs['smooth_loss'] = [np.log(vocabulary_size)]
logs['gradient_norm'] = []
logs['clipped_gradient_norm'] = []
'''
model = [
Dropout(dropout),
CRNN_HSN(dinput, states, output, periods=periods, first_layer=True, sigma=0.1),
Dropout(dropout),
CRNN_HSN(output, states, output, periods=periods, sigma=0.1),
Linear(output, doutputs),
Softmax()
]
'''
model = [
Dropout(dropout),
LSTM(dinput, states, fbias=0.0, sigma=0.1),
Dropout(dropout),
LSTM(states, states, fbias=0.0, sigma=0.1),
Linear(states, doutputs),
Softmax()
]
W = extract_weights(model)
optimizer = Adam(W, dW, learning_rate, momentum=momentum)
config = 'experiment_name = ' + str(experiment_name) + '\n' + 'periods = ' + str(periods) + '\n' + 'vocabulary_size = ' + str(vocabulary_size) + '\n' + 'states = ' + str(states) + '\n' + 'output = ' + str(output) + '\n' + 'dinput = ' + str(dinput) + '\n' + 'doutputs = ' + str(doutputs) + '\n' + 'sequence_length = ' + str(sequence_length) + '\n' + 'batch_size = ' + str(batch_size) + '\n' + 'learning_rate = ' + str(learning_rate) + '\n' + 'niterations = ' + str(niterations) + '\n' + 'momentum = ' + str(momentum) + '\n' + 'forget_every = ' + str(forget_every) + '\n' + 'gradient_clip = ' + str(gradient_clip) + '\n' + 'sample_every = ' + str(sample_every) + '\n' + 'save_every = ' + str(save_every) + '\n' + 'plot_every = ' + str(plot_every) + '\n' + 'full_recurrence = ' + str(full_recurrence) + '\n' + 'learn_state = ' + str(learn_state) + '\n' + 'anneal = ' + str(anneal) + '\n' + 'dynamic_forgetting = ' + str(dynamic_forgetting) + '\n' + 'text = ' + str(text_file) + '\n'
config += 'optimizer = ' + str(optimizer.__class__.__name__) + '\n'
config += 'dropout = ' + str(dropout) + '\n'
f = open(path + 'config.txt', 'w')
f.write(config)
f.close()
print config, 'Approx. Parameters: ', W.size
for i in optimizer:
if i['n_iter'] > niterations:
break
print i['n_iter'], '\t',
print logs['loss'][-1], '\t',
print logs['gradient_norm'][-1]
if i['n_iter'] % forget_every == 0:
forget(model)
if dynamic_mom:
if i['n_iter'] > 100:
optimizer.momentum = 0.9
if dynamic_forgetting:
if i['n_iter'] > 100:
forget_every = 50
if i['n_iter'] > 1000:
forget_every = 500
if i['n_iter'] > 10000:
forget_every = 5000
if i['n_iter'] % sample_every == 0:
forget(model)
x = np.zeros((20, vocabulary_size, 1))
input, _ = data.fetch()
x[:20, :, :] = input[:20, :, 0:1]
ixes = []
for t in xrange(1000):
p = forward(model, np.array(x))
p = p[-1]
ix = np.random.choice(range(vocabulary_size), p=p.ravel())
x = np.zeros((1, vocabulary_size, 1))
x[0, ix, 0] = 1
ixes.append(ix)
sample = ''.join(data.decoder.to_c[ix] for ix in ixes)
print '----' * 20
print sample
print '----' * 20
forget(model)
if anneal:
if i['n_iter'] > 4000:
optimizer.step_rate *= 0.1
elif i['n_iter'] > 6000:
optimizer.step_rate *= 0.1
if i['n_iter'] % save_every == 0:
print 'serializing model... '
f = open(path + 'iter_' + str(i['n_iter']) +'.model', 'w')
pickle.dump(model, f)
f.close()
if i['n_iter'] % plot_every == 0:
plt.clf()
plt.plot(logs['smooth_loss'])
plt.draw()
print 'serializing logs... '
f = open(path + 'logs.logs', 'w')
pickle.dump(logs, f)
f.close()
print 'serializing final model... '
f = open(path + 'final.model', 'w')
pickle.dump(model, f)
f.close()
plt.savefig(path + 'loss_curve')