-
Notifications
You must be signed in to change notification settings - Fork 8
/
train.py
256 lines (222 loc) · 10.4 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
253
254
255
256
import argparse
import importlib
import json
import os
import pickle
import sys
import time
import numpy as np
import tensorflow as tf
import logger
import utils
# -----------------------------------------------------------------------------
parser = argparse.ArgumentParser()
parser.add_argument('--config_name', type=str, required=True, help='Configuration name')
parser.add_argument('--nr_gpu', type=int, default=1, help='How many GPUs to distribute the training across?')
parser.add_argument('--resume', type=int, default=0, help='Resume training from a checkpoint?')
parser.add_argument('--tf_seed', type=int, default=0, help='tf rng seed')
args = parser.parse_args()
print('input args:\n', json.dumps(vars(args), indent=4, separators=(',', ':')))
assert args.nr_gpu == len(''.join(filter(str.isdigit, os.environ["CUDA_VISIBLE_DEVICES"])))
# -----------------------------------------------------------------------------
np.random.seed(seed=42)
tf.reset_default_graph()
tf.set_random_seed(args.tf_seed)
# config
configs_dir = __file__.split('/')[-2]
config = importlib.import_module('%s.%s' % (configs_dir, args.config_name))
if not args.resume:
experiment_id = '%s-%s' % (args.config_name.split('.')[-1], time.strftime("%Y_%m_%d", time.localtime()))
utils.autodir('metadata')
save_dir = 'metadata/' + experiment_id
utils.autodir(save_dir)
else:
save_dir = utils.find_model_metadata('metadata/', args.config_name)
experiment_id = os.path.dirname(save_dir).split('/')[-1]
with open(save_dir + '/meta.pkl', 'rb') as f:
resumed_metadata = pickle.load(f)
last_lr = resumed_metadata['lr']
last_iteration = resumed_metadata['iteration']
print('Last iteration', last_iteration)
print('Last learning rate', last_lr)
# logs
utils.autodir('logs')
sys.stdout = logger.Logger('logs/%s.log' % experiment_id)
sys.stderr = sys.stdout
print('exp_id', experiment_id)
if args.resume:
print('Resuming training')
# create the model
model = tf.make_template('model', config.build_model)
# run once for data dependent initialization of parameters
x_init = tf.placeholder(tf.float32, shape=(config.batch_size,) + config.obs_shape)
init_pass = model(x_init, init=True)[0]
all_params = tf.trainable_variables()
n_parameters = 0
for variable in all_params:
shape = variable.get_shape()
variable_parameters = 1
for dim in shape:
variable_parameters *= dim.value
n_parameters += variable_parameters
print('Number of parameters', n_parameters)
# get loss gradients over multiple GPUs
xs = []
grads = []
train_losses = []
# evaluation in case we want to validate
x_in_eval = tf.placeholder(tf.float32, shape=(config.batch_size,) + config.obs_shape)
log_probs = model(x_in_eval)[0]
eval_loss = config.eval_loss(log_probs) if hasattr(config, 'eval_loss') else config.loss(log_probs)
for i in range(args.nr_gpu):
xs.append(tf.placeholder(tf.float32, shape=(config.batch_size / args.nr_gpu,) + config.obs_shape))
with tf.device('/gpu:%d' % i):
# train
with tf.variable_scope('gpu_%d' % i):
with tf.variable_scope('train'):
log_probs = model(xs[i])[0]
train_losses.append(config.loss(log_probs))
grads.append(tf.gradients(train_losses[i], all_params))
# add gradients together and get training updates
tf_lr = tf.placeholder(tf.float32, shape=[])
tf_student_grad_scale = tf.placeholder(tf.float32, shape=[])
with tf.device('/gpu:0'):
for i in range(1, args.nr_gpu):
train_losses[0] += train_losses[i]
for j in range(len(grads[0])):
grads[0][j] += grads[i][j]
# average over gpus
train_losses[0] /= args.nr_gpu
for j in range(len(grads[0])):
grads[0][j] /= args.nr_gpu
# scale gradients of student_params
student_params = ['prior_nu', 'prior_mean', 'prior_var', 'prior_corr']
for j in range(len(grads[0])):
if any(name in all_params[j].name for name in student_params):
grads[0][j] *= tf_student_grad_scale
# training op
grads_and_vars = zip(grads[0], all_params)
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
update_ops_gpu0 = []
for u in update_ops:
if u.name.startswith('gpu_0/train'):
update_ops_gpu0.append(u)
with tf.control_dependencies(update_ops_gpu0):
if hasattr(config, 'optimizer') and config.optimizer == 'rmsprop':
print('using rmsprop')
train_step = tf.train.RMSPropOptimizer(learning_rate=tf_lr).apply_gradients(
grads_and_vars=grads_and_vars,
global_step=None, name='rmsprop')
else:
print('using adam')
train_step = tf.train.AdamOptimizer(learning_rate=tf_lr, beta1=0.95, beta2=0.9995).apply_gradients(
grads_and_vars=grads_and_vars,
global_step=None, name='adam')
train_loss = train_losses[0]
# init & save
initializer = tf.global_variables_initializer()
saver = tf.train.Saver()
print('\n Start training')
train_data_iter = config.train_data_iter
lr = config.learning_rate
student_grad_scale = config.scale_student_grad
batch_idxs = range(0, config.max_iter)
print_every = 100
train_iter_losses = []
if args.resume:
losses_eval_valid = resumed_metadata['losses_eval_valid']
losses_eval_train = resumed_metadata['losses_eval_train']
losses_avg_train = resumed_metadata['losses_avg_train']
else:
losses_eval_valid, losses_eval_train, losses_avg_train = [], [], []
start_time = time.time()
with tf.Session() as sess:
if args.resume:
ckpt_file = save_dir + 'params.ckpt'
print('restoring parameters from', ckpt_file)
saver.restore(sess, tf.train.latest_checkpoint(save_dir))
prev_time = time.time()
for iteration, x_batch in zip(batch_idxs, train_data_iter.generate()):
if hasattr(config, 'learning_rate_schedule') and iteration in config.learning_rate_schedule:
lr = np.float32(config.learning_rate_schedule[iteration])
elif hasattr(config, 'lr_decay'):
lr *= config.lr_decay
if hasattr(config, 'student_grad_schedule') and iteration in config.student_grad_schedule:
student_grad_scale = np.float32(config.student_grad_schedule[iteration])
print('setting student grad scale to %.7f' % config.student_grad_schedule[iteration])
if args.resume and iteration < last_iteration:
if iteration % (print_every * 10) == 0:
print(iteration, 'skipping training')
continue
# init
if iteration == 0:
print('initializing the model...')
sess.run(initializer)
init_loss = sess.run(init_pass, {x_init: x_batch})
sess.graph.finalize()
else:
xfs = np.split(x_batch, args.nr_gpu)
feed_dict = {tf_lr: lr, tf_student_grad_scale: student_grad_scale}
feed_dict.update({xs[i]: xfs[i] for i in range(args.nr_gpu)})
l, _ = sess.run([train_loss, train_step], feed_dict)
train_iter_losses.append(l)
if np.isnan(l):
print('Loss is NaN')
sys.exit(0)
if (iteration + 1) % print_every == 0:
avg_train_loss = np.mean(train_iter_losses)
losses_avg_train.append(avg_train_loss)
train_iter_losses = []
print('%d/%d train_loss=%6.8f bits/value=%.3f' % (
iteration + 1, config.max_iter, avg_train_loss, avg_train_loss / config.ndim / np.log(2.)))
corr = config.student_layer.corr.eval().flatten()
if hasattr(config, 'validate_every') and (iteration + 1) % config.validate_every == 0:
print('\n Validating ...')
losses = []
rng = np.random.RandomState(42)
for _, x_valid_batch in zip(range(0, config.n_valid_batches),
config.valid_data_iter.generate(rng)):
feed_dict = {x_in_eval: x_valid_batch}
l = sess.run([eval_loss], feed_dict)
losses.append(l)
avg_loss = np.mean(np.asarray(losses), axis=0)
losses_eval_valid.append(avg_loss)
print('valid loss', avg_loss)
losses = []
rng = np.random.RandomState(42)
for _, x_valid_batch in zip(range(0, config.n_valid_batches * 10),
train_data_iter.generate(rng)):
feed_dict = {x_in_eval: x_valid_batch}
l = sess.run([eval_loss], feed_dict)
losses.append(l)
avg_loss = np.mean(np.asarray(losses), axis=0)
losses_eval_train.append(avg_loss)
print('train loss', avg_loss)
if (iteration + 1) % config.save_every == 0:
current_time = time.time()
eta_time = (config.max_iter - iteration) / config.save_every * (current_time - prev_time)
prev_time = current_time
print('ETA: ', time.strftime("%H:%M:%S", time.gmtime(eta_time)))
print('Saving model (iteration %s):' % iteration, experiment_id)
print('current learning rate:', lr)
saver.save(sess, save_dir + '/params.ckpt')
with open(save_dir + '/meta.pkl', 'wb') as f:
pickle.dump({'lr': lr,
'iteration': iteration + 1,
'losses_avg_train': losses_avg_train,
'losses_eval_valid': losses_eval_valid,
'losses_eval_train': losses_eval_train}, f)
corr = config.student_layer.corr.eval().flatten()
print('0.01', np.sum(corr > 0.01))
print('0.1', np.sum(corr > 0.1))
print('0.2', np.sum(corr > 0.2))
print('0.3', np.sum(corr > 0.3))
print('0.5', np.sum(corr > 0.5))
print('0.7', np.sum(corr > 0.7))
print('corr min-max:', np.min(corr), np.max(corr))
var = config.student_layer.var.eval().flatten()
print('var min-max:', np.min(var), np.max(var))
if hasattr(config.student_layer, 'nu'):
nu = config.student_layer.nu.eval().flatten()
print('nu median-min-max:', np.median(nu), np.min(nu), np.max(nu))
print('Total time: ', time.strftime("%H:%M:%S", time.gmtime(time.time() - start_time)))