/
train_utils.py
executable file
·160 lines (136 loc) · 6.19 KB
/
train_utils.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
# Learning the Enigma with Recurrent Neural Networks
# Sam Greydanus. January 2017. MIT License.
import time
import sys
import copy
import numpy as np
import os
os.environ['TF_CPP_MIN_LOG_LEVEL']='2'
import tensorflow as tf
import matplotlib as mpl
mpl.use('Agg')
import matplotlib.pyplot as plt
# quantitative eval
def accuracy(model, data):
mean_acc = 0 ; trials = max(1, int(500/model.batch_size))
kl = data.key_len
for _ in range(trials):
batch = data.next_batch(model.batch_size)
y = np.stack(batch[1])
y_hat = model.decode(batch[0])
mean_acc += np.sum(y[:,kl:,:]*y_hat[:,kl:,:])/(y.shape[0]*(y.shape[1]-kl)) # TODO: make this less awk
return mean_acc * 100 / trials
# qualitative eval
def sample(model, data, FLAGS):
model.reset_states()
key = unicode("KEY", "utf-8") ; plaintext = unicode("YOUKNOWNOTHINGJONSNOW", "utf-8")
ciphertext = key + '-'*(data.key_len - len(key)) + data.encode(key, plaintext)
decoded = ''
for i in range(len(ciphertext)):
p = np.tile(data.one_hot(ciphertext[i]),[1,1,1])
c = model.step(p)
ix = np.where(c[0,0,:] == np.amax(c[0,0,:]))
ix = np.squeeze(ix[0])
if ix <= 25: # TODO: make this less awk
decoded += data.A[ix]
else:
decoded += '-'
print( "\t\tplaintext is: '{}'".format(key + '-'*(data.key_len - len(key)) + plaintext) )
print( "\t\tciphertext is: '{}'".format(ciphertext) )
print( "\t\tprediction is: '{}'".format(decoded) )
return plaintext, decoded
def train(model, data, log, FLAGS):
running_loss = None # for smoothing the loss over time
start = time.time()
# Restore train state. There's DEFINITELY a more efficient way to do this but I don't want to spend my
# time dealing with tedious file I/O. Still gonna send it! https://www.youtube.com/watch?v=WIrWyr3HgXI
global_step = model.try_load_model()
if global_step == 0:
log.clear_fs() ; print('\tresetting log files...')
else:
loss_hist = log.read('loss')
acc_hist = log.read('acc')
log.clear_fs()
for i in range(loss_hist.shape[0]):
if loss_hist[i,0] <= global_step:
log.write_loss('{},{},{}'.format(loss_hist[i,0], loss_hist[i,1],loss_hist[i,2]))
for i in range(acc_hist.shape[0]):
if acc_hist[i,0] <= global_step:
log.write_acc('{},{},{},{},{}'\
.format(acc_hist[i,0], acc_hist[i,1], acc_hist[i,2], acc_hist[i,3], acc_hist[i,4]))
print('training...')
while global_step < FLAGS.total_steps:
global_step += 1
data_start = time.time()
batch = data.next_batch(model.batch_size)
rnn_start = time.time()
train_loss = model.train_step(batch)
# bookkeeping starts here
data_time = rnn_start-data_start ; rnn_time = time.time() - rnn_start ; tot_time = time.time() - data_start
running_loss = train_loss if running_loss is None else 0.99*running_loss + 0.01*train_loss
log.write_loss('{},{},{}'.format(global_step, train_loss*100, tot_time))
if global_step%FLAGS.print_every == 0:
print( "\tstep: {:7d} | loss: {:.3f} | data time: {:.4f} sec | rnn time: {:.4f} sec | total time: {:.4f} sec"\
.format(global_step, running_loss*100, data_time, rnn_time, tot_time) )
if global_step%FLAGS.acc_every == 0:
train_acc = accuracy(model,data)
print('\t\taccuracy: {}'.format(train_acc))
sample(model, data, FLAGS)
log.write_acc('{},{},{},{},{}'.format(global_step, train_acc, data_time, rnn_time, tot_time))
if global_step%FLAGS.save_every == 0 and global_step!=0:
model.save(global_step)
print( "\t\t...SAVED MODEL #{}".format(global_step) )
class Logger():
def __init__(self, FLAGS):
self.lossf = FLAGS.meta_dir + 'loss_hist.txt'
self.accf = FLAGS.meta_dir + 'acc_hist.txt'
def clear_fs(self):
with open(self.accf, 'w') as f: f.write('')
with open(self.lossf, 'w') as f: f.write('')
def write_acc(self, s):
with open(self.accf, 'a') as f:
f.write(s + "\n")
def write_loss(self, s):
with open(self.lossf, 'a') as f:
f.write(s + "\n")
def read(self, mode='loss'):
file = None
if mode is 'loss':
file = self.lossf
elif mode is 'acc':
file = self.accf
else:
print('error: mode not recognized')
hist = []
with open(file, 'r') as f:
for line in f:
data = [float(s) for s in line.split(',')]
hist.append(data)
return np.vstack(hist)
def plot_stats(log, FLAGS):
loss_hist = log.read('loss')
acc_hist = log.read('acc')
f1 = plt.figure(figsize=[16,5])
plt.subplot(121)
plt.plot(loss_hist[:,0], loss_hist[:,1], linewidth=3.0, label='loss')
plt.title('Loss', fontsize=14)
plt.xlabel('train step', fontsize=14) ; plt.setp(plt.gca().axes.get_xticklabels(), fontsize=14)
plt.ylabel('loss', fontsize=14) ; plt.setp(plt.gca().axes.get_yticklabels(), fontsize=14)
plt.ylim([0,45])
plt.subplot(122)
plt.plot(acc_hist[:,0], acc_hist[:,1], linewidth=3.0, label='accuracy')
plt.title('Test accuracy', fontsize=14)
plt.xlabel('train step', fontsize=14) ; plt.setp(plt.gca().axes.get_xticklabels(), fontsize=14)
plt.ylabel('accuracy (%)', fontsize=14) ; plt.setp(plt.gca().axes.get_yticklabels(), fontsize=14)
tt_sec = np.sum(loss_hist[:,2])
m, s = divmod(tt_sec, 60)
h, m = divmod(m, 60)
tt_pretty = "%02dh %02dm %02ds"%(h, m, s)
results_msg = '\nlearning rate : {}\nbatch size: {}\ntrain time: {}\nfinal accuracy: {:.2f}%'\
.format(FLAGS.lr, FLAGS.batch_size, tt_pretty, acc_hist[-1,1])
print('\nTRAIN SUMMARY:' + results_msg + '\n')
f1.text(0.92, .50, results_msg, ha='left', va='center', fontsize=12)
plt.ylim([0,100])
title = "RNN on {} task".format(FLAGS.cipher[0].upper() + FLAGS.cipher[1:])
f1.text(0.4, .97, title, ha='left', va='center', fontsize=18)
f1.savefig('./{}train-{}.png'.format(FLAGS.meta_dir, FLAGS.cipher), bbox_inches='tight')