-
Notifications
You must be signed in to change notification settings - Fork 0
/
train.py
252 lines (203 loc) · 8.39 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
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
import cPickle as pickle
import string
import sys
import time
from itertools import izip
import numpy as np
from datetime import datetime, timedelta
import buffering
import utils
import logger
from configuration import config, set_configuration
import pathfinder
import app
import torch
from torch.autograd import Variable
if len(sys.argv) < 2:
sys.exit("Usage: CUDA_VISIBLE_DEVICES=<gpu_number> python train.py <configuration_name>")
config_name = sys.argv[1]
set_configuration('configs_pytorch', config_name)
expid = utils.generate_expid(config_name)
print
print "Experiment ID: %s" % expid
print
# metadata
metadata_dir = utils.get_dir_path('models', pathfinder.METADATA_PATH)
metadata_path = metadata_dir + '/%s.pkl' % expid
metadata_best_path = metadata_dir + '/%s-best.pkl' % expid
# logs
logs_dir = utils.get_dir_path('logs', pathfinder.METADATA_PATH)
sys.stdout = logger.Logger(logs_dir + '/%s.log' % expid)
sys.stderr = sys.stdout
print 'Build model'
model = config().build_model()
print model.l_out
model.l_out.cuda() # move to gpu
criterion = config().build_objective()
criterion2 = config().build_objective2()
learning_rate_schedule = config().learning_rate_schedule
learning_rate =(learning_rate_schedule[0])
optimizer = config().build_updates(model.l_out, learning_rate)
chunk_idxs = range(config().max_nchunks)
losses_eval_train = []
losses_eval_valid = []
losses_eval_train2 = []
losses_eval_valid2 = []
start_chunk_idx = 0
train_data_iterator = config().train_data_iterator
valid_data_iterator = config().valid_data_iterator
print
print 'Data'
print 'n train: %d' % train_data_iterator.nsamples
print 'n validation: %d' % valid_data_iterator.nsamples
print 'n chunks per epoch', config().nchunks_per_epoch
print
print 'Train model'
chunk_idx = 0
start_time = time.time()
prev_time = start_time
tmp_preds = []
tmp_gts = []
tmp_losses_train = []
tmp_losses_train2 = []
tmp_preds_train = []
tmp_gts_train = []
losses_train_print = []
losses_train_print2 = []
preds_train_print = []
gts_train_print = []
losses_time_print = []
best_valid_f2_score = 0
best_threshold = 0.91
# use buffering.buffered_gen_threaded()
for chunk_idx, (x_chunk_train, y_chunk_train, id_train) in izip(chunk_idxs, buffering.buffered_gen_threaded(
train_data_iterator.generate(), buffer_size=128)):
if chunk_idx in learning_rate_schedule:
lr = learning_rate_schedule[chunk_idx]
print ' setting learning rate to %.7f' % lr
print
for param_group in optimizer.param_groups:
param_group['lr'] = lr
for gt in y_chunk_train:
tmp_gts.append(gt)
tmp_gts_train.append(gt)
gts_train_print.append(gt)
# make nbatches_chunk iterations
model.l_out.train()
for b in xrange(config().nbatches_chunk):
losses_time_print.append(time.time())
# wrap them in Variable
inputs, labels = Variable(torch.from_numpy(x_chunk_train).cuda()), \
Variable(torch.from_numpy(y_chunk_train).cuda())
# zero the parameter gradients
optimizer.zero_grad()
# forward + backward + optimize
outputs = model.l_out(inputs)
loss = criterion(outputs, labels)
loss2 = criterion2(outputs, labels)
loss.backward()
optimizer.step()
pr=outputs.cpu().data.numpy()
tmp_preds.append(pr)
tmp_preds_train.append(pr)
preds_train_print.append(pr)
loss_out = loss.cpu().data.numpy()[0]
loss2_out = loss2.cpu().data.numpy()[0]
tmp_losses_train.append(loss_out)
tmp_losses_train2.append(loss2_out)
losses_train_print.append(loss_out)
losses_train_print2.append(loss2_out)
if (chunk_idx + 1) % 10 == 0:
print 'Chunk %d/%d %.1fHz' % (chunk_idx + 1, config().max_nchunks,
10. * config().nbatches_chunk * config().batch_size / (
time.time() - losses_time_print[0])),
print np.mean(losses_train_print), np.mean(losses_train_print2)
print 'score', config().score(gts_train_print, np.vstack(preds_train_print))
preds_train_print = []
gts_train_print = []
losses_train_print = []
losses_time_print = []
losses_train_print2 = []
losses_time_print2 = []
if ((chunk_idx + 1) % config().validate_every) == 0:
print
print 'Chunk %d/%d' % (chunk_idx + 1, config().max_nchunks)
# calculate mean train loss since the last validation phase
mean_train_loss = np.mean(tmp_losses_train)
mean_train_loss2 = np.mean(tmp_losses_train2)
mean_train_score = np.mean(config().score(tmp_gts_train, np.vstack(tmp_preds_train)))
print 'Mean train loss: %7f' % mean_train_loss, mean_train_loss2, mean_train_score
losses_eval_train.append(mean_train_loss)
tmp_losses_train = []
tmp_losses_train2 = []
tmp_preds_train = []
tmp_gts_train = []
# load validation data to GPU
tmp_losses_valid = []
tmp_losses_valid2 = []
tmp_preds_valid = []
tmp_gts_valid = []
model.l_out.eval()
for i, (x_chunk_valid, y_chunk_valid, ids_batch) in enumerate(
buffering.buffered_gen_threaded(valid_data_iterator.generate(),
buffer_size=2)):
inputs, labels = Variable(torch.from_numpy(x_chunk_valid).cuda(),volatile=True), Variable(
torch.from_numpy(y_chunk_valid).cuda(),volatile=True)
outputs = model.l_out(inputs)
loss = criterion(outputs, labels)
loss2 = criterion2(outputs, labels)
pr = outputs.cpu().data.numpy()
tmp_preds_valid.append(pr)
tmp_losses_valid.append(loss.cpu().data.numpy()[0])
tmp_losses_valid2.append(loss2.cpu().data.numpy()[0])
for gt in y_chunk_valid:
tmp_gts_valid.append(gt)
# calculate validation loss across validation set
valid_loss = np.mean(tmp_losses_valid)
valid_loss2 = np.mean(tmp_losses_valid2)
valid_score = np.mean(config().score(tmp_gts_valid, np.vstack(tmp_preds_valid)))
print 'Validation loss: ', valid_loss, valid_loss2, valid_score
losses_eval_valid.append(valid_loss)
losses_eval_valid2.append(valid_loss2)
if valid_score > best_threshold and valid_score > best_valid_f2_score:
with open(metadata_best_path, 'w') as f:
pickle.dump({
'configuration_file': config_name,
'git_revision_hash': utils.get_git_revision_hash(),
'experiment_id': expid,
'chunks_since_start': chunk_idx,
'losses_eval_train': losses_eval_train,
'losses_eval_valid': losses_eval_valid,
'param_values': model.l_out.state_dict(),
#'optimizer_values': optimizer.state_dict(),
}, f, pickle.HIGHEST_PROTOCOL)
print ' saved to %s' % metadata_best_path
print
best_valid_f2_score = valid_score
now = time.time()
time_since_start = now - start_time
time_since_prev = now - prev_time
prev_time = now
est_time_left = time_since_start * (config().max_nchunks - chunk_idx + 1.) / (chunk_idx + 1. - start_chunk_idx)
eta = datetime.now() + timedelta(seconds=est_time_left)
eta_str = eta.strftime("%c")
print " %s since start (%.2f s)" % (utils.hms(time_since_start), time_since_prev)
print " estimated %s to go (ETA: %s)" % (utils.hms(est_time_left), eta_str)
print
if ((chunk_idx + 1) % config().save_every) == 0:
print
print 'Chunk %d/%d' % (chunk_idx + 1, config().max_nchunks)
print 'Saving metadata, parameters'
with open(metadata_path, 'w') as f:
pickle.dump({
'configuration_file': config_name,
'git_revision_hash': utils.get_git_revision_hash(),
'experiment_id': expid,
'chunks_since_start': chunk_idx,
'losses_eval_train': losses_eval_train,
'losses_eval_valid': losses_eval_valid,
'param_values': model.l_out.state_dict(),
'optimizer_values': optimizer.state_dict(),
}, f, pickle.HIGHEST_PROTOCOL)
print ' saved to %s' % metadata_path
print