/
train.py
212 lines (179 loc) · 6.02 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
import os
from algorithms import Adasecant
from blocks import initialization
from blocks.algorithms import (
Adam, CompositeRule, GradientDescent, StepClipping)
from blocks.extensions import (Printing, Timing)
from blocks.extensions.monitoring import (
DataStreamMonitoring, TrainingDataMonitoring)
from blocks.extensions.predicates import OnLogRecord
from blocks.extensions.saveload import Checkpoint, Load
from blocks.extensions.training import TrackTheBest
from blocks.graph import ComputationGraph
from blocks.main_loop import MainLoop
from blocks.model import Model
import cPickle
from extensions import LearningRateSchedule, Plot, TimedFinish, Write
from iam_on_line import stream_handwriting
from model import Scribe
from theano import function
from utils import train_parse
parser = train_parse()
args = parser.parse_args()
if args.algorithm == "adasecant":
args.lr_schedule = False
rec_h_dim = args.rnn_size
att_size = args.size_attention
k = args.num_mixture
exp_name = args.experiment_name
save_dir = args.save_dir
print "Saving config ..."
with open(os.path.join(save_dir, 'config', exp_name + '.pkl'), 'w') as f:
cPickle.dump(args, f)
print "Finished saving."
w_init = initialization.IsotropicGaussian(0.01)
b_init = initialization.Constant(0.)
train_stream = stream_handwriting(
which_sets=('train',),
batch_size=args.batch_size,
seq_size=args.train_seq_length,
num_letters=args.num_letters,
sorting_mult=args.sort_mult)
valid_stream = stream_handwriting(
which_sets=('valid',),
batch_size=args.batch_size,
seq_size=args.valid_seq_length,
num_letters=args.num_letters,
sorting_mult=5)
x_tr, x_mask_tr, context_tr, context_mask_tr, flag_tr = \
next(train_stream.get_epoch_iterator())
scribe = Scribe(
k=args.num_mixture,
rec_h_dim=args.rnn_size,
att_size=args.size_attention,
num_letters=args.num_letters,
sampling_bias=0.,
attention_type=args.attention_type,
attention_alignment=args.attention_alignment,
weights_init=w_init,
biases_init=b_init,
name="scribe")
scribe.initialize()
data, data_mask, context, context_mask, start_flag = \
scribe.symbolic_input_variables()
cost, extra_updates = scribe.compute_cost(
data, data_mask, context, context_mask, start_flag, args.batch_size)
cost.name = 'nll'
valid_cost = cost
cg = ComputationGraph(cost)
model = Model(cost)
parameters = cg.parameters
if args.algorithm == "adam":
step_rule = CompositeRule(
[StepClipping(10. * args.grad_clip), Adam(args.learning_rate)])
elif args.algorithm == "adasecant":
step_rule = Adasecant(grad_clip=args.grad_clip)
algorithm = GradientDescent(
cost=cost,
parameters=parameters,
step_rule=step_rule,
on_unused_sources='warn')
algorithm.add_updates(extra_updates)
monitoring_vars = [cost]
if args.lr_schedule:
lr = algorithm.step_rule.components[1].learning_rate
monitoring_vars.append(lr)
train_monitor = TrainingDataMonitoring(
variables=monitoring_vars,
every_n_batches=args.save_every,
after_epoch=False,
prefix="train")
valid_monitor = DataStreamMonitoring(
[valid_cost],
valid_stream,
every_n_batches=args.save_every,
after_epoch=False,
prefix="valid")
# Multi GPU
worker = None
if args.platoon_port:
from blocks_extras.extensions.synchronization import (
Synchronize, SynchronizeWorker)
from platoon.param_sync import ASGD
sync_rule = ASGD()
worker = SynchronizeWorker(
sync_rule, control_port=args.platoon_port, socket_timeout=2000)
extensions = []
if args.load_experiment and (not worker or worker.is_main_worker):
extensions += [Load(os.path.join(
save_dir, "pkl", "best_" + args.load_experiment + ".tar"))]
extensions += [
Timing(every_n_batches=args.save_every),
train_monitor]
if not worker or worker.is_main_worker:
extensions += [
valid_monitor,
TrackTheBest(
'valid_nll',
every_n_batches=args.save_every,
before_first_epoch=True),
Plot(
os.path.join(save_dir, "progress", exp_name + ".png"),
[['train_nll', 'valid_nll']],
every_n_batches=args.save_every,
email=False),
Checkpoint(
os.path.join(save_dir, "pkl", "best_" + exp_name + ".tar"),
after_training=False,
save_separately=['log'],
use_cpickle=True,
save_main_loop=False)
.add_condition(
["after_batch", "before_training"],
predicate=OnLogRecord('valid_nll_best_so_far')),
Checkpoint(
os.path.join(save_dir, "pkl", "last_" + exp_name + ".tar"),
after_training=True,
save_separately=['log'],
use_cpickle=True,
every_n_batches=args.save_every,
save_main_loop=False)]
if args.lr_schedule:
extensions += [
LearningRateSchedule(
lr, 'valid_nll',
os.path.join(save_dir, "pkl", "best_" + exp_name + ".tar"),
patience=10,
num_cuts=5,
every_n_batches=args.save_every)]
if args.plot_every:
sample_x, sample_pi, sample_phi, sample_pi_att, updates_sample = \
scribe.sample_model(
context, context_mask, args.num_steps, args.num_samples)
sampling_function = function(
[context, context_mask], sample_x, updates=updates_sample)
extensions += [
Write(
sampling_function,
every_n_batches=args.plot_every,
n_samples=args.num_samples,
save_name=os.path.join(save_dir, "samples", exp_name))]
if worker:
extensions += [
Synchronize(
worker,
after_batch=True,
before_epoch=True)]
extensions += [
Printing(
after_epoch=False,
every_n_batches=args.save_every)]
if args.time_limit:
extensions += [TimedFinish(args.time_limit)]
main_loop = MainLoop(
model=model,
data_stream=train_stream,
algorithm=algorithm,
extensions=extensions)
print "Training starting."
main_loop.run()