-
Notifications
You must be signed in to change notification settings - Fork 57
/
run_utils.py
137 lines (118 loc) · 5.84 KB
/
run_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
import sys
import time
import numpy as np
import tensorflow as tf
from tensorflow.python.client import timeline
from language_model import LM
from common import CheckpointLoader
def run_train(dataset, hps, logdir, ps_device, task=0, master=""):
with tf.variable_scope("model"):
model = LM(hps, "train", ps_device)
print("ALL VARIABLES")
for v in tf.all_variables():
print("%s %s %s" % (v.name, v.get_shape(), v.device))
print("TRAINABLE VARIABLES")
for v in tf.trainable_variables():
print("%s %s %s" % (v.name, v.get_shape(), v.device))
print("LOCAL VARIABLES")
for v in tf.local_variables():
print("%s %s %s" % (v.name, v.get_shape(), v.device))
sv = tf.train.Supervisor(is_chief=(task == 0),
logdir=logdir,
summary_op=None, # Automatic summaries don't work with placeholders.
global_step=model.global_step,
save_summaries_secs=30,
save_model_secs=120 * 5)
config = tf.ConfigProto(allow_soft_placement=True,
intra_op_parallelism_threads=2,
inter_op_parallelism_threads=20)
with sv.managed_session(master, config=config) as sess:
# Slowly increase the number of workers during beginning of the training.
while not sv.should_stop():
step = int(sess.run(model.global_step))
waiting_until_step = task * hps.num_delayed_steps
if step >= waiting_until_step:
break
else:
print("Current step is %d. Waiting until: %d" % (step, waiting_until_step))
time.sleep(10.0)
local_step = 0
prev_global_step = sess.run(model.global_step)
prev_time = time.time()
data_iterator = dataset.iterate_forever(hps.batch_size * hps.num_gpus, hps.num_steps)
while not sv.should_stop():
fetches = [model.global_step, model.loss, model.train_op]
# Chief worker computes summaries every 20 steps.
should_compute_summary = (task == 0 and local_step > 0 and local_step % 20 == 0)
if should_compute_summary:
fetches += [model.summary_op]
x, y, w = next(data_iterator)
should_run_profiler = (hps.run_profiler and task == 0 and local_step % 1000 == 13)
if should_run_profiler:
run_options = tf.RunOptions(trace_level=tf.RunOptions.FULL_TRACE)
run_metadata = tf.RunMetadata()
fetched = sess.run(fetches, {model.x: x, model.y: y, model.w: w},
options=run_options, run_metadata=run_metadata)
# Create the Timeline object, and write it to a json
tl = timeline.Timeline(run_metadata.step_stats)
ctf = tl.generate_chrome_trace_format()
print("Running profiler")
with open(logdir + "/timeline.json", 'w') as f:
f.write(ctf)
print("Finished profiling!")
else:
fetched = sess.run(fetches, {model.x: x, model.y: y, model.w: w})
local_step += 1
if should_compute_summary:
sv.summary_computed(sess, fetched[-1])
if local_step < 10 or local_step % 20 == 0:
cur_time = time.time()
num_words = hps.batch_size * hps.num_gpus * hps.num_steps
wps = (fetched[0] - prev_global_step) * num_words / (cur_time - prev_time)
prev_global_step = fetched[0]
print("Iteration %d, time = %.2fs, wps = %.0f, train loss = %.4f" % (
fetched[0], cur_time - prev_time, wps, fetched[1]))
prev_time = cur_time
sv.stop()
def run_eval(dataset, hps, logdir, mode, num_eval_steps):
with tf.variable_scope("model"):
hps.num_sampled = 0 # Always using full softmax at evaluation.
hps.keep_prob = 1.0
model = LM(hps, "eval", "/cpu:0")
if hps.average_params:
print("Averaging parameters for evaluation.")
saver = tf.train.Saver(model.avg_dict)
else:
saver = tf.train.Saver()
# Use only 4 threads for the evaluation.
config = tf.ConfigProto(allow_soft_placement=True,
intra_op_parallelism_threads=20,
inter_op_parallelism_threads=1)
sess = tf.Session(config=config)
sw = tf.train.SummaryWriter(logdir + "/" + mode, sess.graph)
ckpt_loader = CheckpointLoader(saver, model.global_step, logdir + "/train")
with sess.as_default():
while ckpt_loader.load_checkpoint():
global_step = ckpt_loader.last_global_step
data_iterator = dataset.iterate_once(hps.batch_size * hps.num_gpus, hps.num_steps)
tf.initialize_local_variables().run()
loss_nom = 0.0
loss_den = 0.0
for i, (x, y, w) in enumerate(data_iterator):
if i >= num_eval_steps:
break
loss = sess.run(model.loss, {model.x: x, model.y: y, model.w: w})
loss_nom += loss
loss_den += w.mean()
loss = loss_nom / loss_den
sys.stdout.write("%d: %.3f (%.3f) ... " % (i, loss, np.exp(loss)))
sys.stdout.flush()
sys.stdout.write("\n")
log_perplexity = loss_nom / loss_den
print("Results at %d: log_perplexity = %.3f perplexity = %.3f" % (
global_step, log_perplexity, np.exp(log_perplexity)))
summary = tf.Summary()
summary.value.add(tag='eval/log_perplexity', simple_value=log_perplexity)
summary.value.add(tag='eval/perplexity', simple_value=np.exp(log_perplexity))
sw.add_summary(summary, global_step)
sw.flush()