-
Notifications
You must be signed in to change notification settings - Fork 87
/
trainer.py
186 lines (154 loc) · 6.37 KB
/
trainer.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
# Copyright 2021, Google LLC.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""Module to train a condititonal flow prediction model."""
from absl import app
from absl import flags
from mint.core import inputs
from mint.core import learning_schedules
from mint.core import model_builder
from mint.ctl import single_task_trainer
from mint.utils import config_util
from third_party.tf_models import orbit
import tensorflow as tf
TRAIN_STRATEGY = ['tpu', 'gpu']
FLAGS = flags.FLAGS
flags.DEFINE_enum('train_strategy', TRAIN_STRATEGY[1], TRAIN_STRATEGY,
'Whether to train with TPUs or Mirrored GPUs.')
flags.DEFINE_string('master', None, 'BNS name of the TensorFlow tpu to use.')
flags.DEFINE_string('config_path', None, 'Path to the config file.')
flags.DEFINE_string('model_dir', None,
'Directory to write training checkpoints and logs')
flags.DEFINE_float('initial_learning_rate', 0.1,
'Initial learning rate for cosine decay schedule')
flags.DEFINE_string(
'head_initializer', 'he_normal',
'Initializer for prediction head. Valid options are any '
'of the tf.keras.initializers.')
flags.DEFINE_integer('steps', 2400000, 'Number of training steps')
flags.DEFINE_integer('warmup_steps', 1000,
'Number of learning rate warmup steps')
flags.DEFINE_float('weight_decay', None, 'L2 regularization penalty to apply.')
flags.DEFINE_float('grad_clip_norm', 0., 'Clip gradients by norm.')
def _create_learning_rate(learning_rate_config):
"""Create optimizer learning rate based on config.
Args:
learning_rate_config: A LearningRate proto message.
Returns:
A learning rate schedule.
Raises:
ValueError: when using an unsupported input data type.
"""
lr_schedule = None
learning_rate_type = learning_rate_config.WhichOneof('learning_rate')
if learning_rate_type == 'exponential_decay_learning_rate':
config = learning_rate_config.exponential_decay_learning_rate
lr_schedule = tf.keras.optimizers.schedules.PolynomialDecay(
FLAGS.initial_learning_rate,
decay_steps=config.decay_steps,
end_learning_rate=config.min_learning_rate,
power=config.decay_factor)
if FLAGS.warmup_steps:
lr_schedule = learning_schedules.WarmUp(
FLAGS.initial_learning_rate,
decay_schedule_fn=lr_schedule,
warmup_steps=FLAGS.warmup_steps)
if learning_rate_type == 'manual_step_learning_rate':
config = learning_rate_config.manual_step_learning_rate
if not config.schedule:
raise ValueError('Empty learning rate schedule.')
learning_rate_step_boundaries = [x.step for x in config.schedule]
learning_rate_sequence = [config.initial_learning_rate]
learning_rate_sequence += [x.learning_rate for x in config.schedule]
lr_schedule = learning_schedules.ManualStepping(
learning_rate_step_boundaries, learning_rate_sequence,
config.warmup)
if learning_rate_type == 'cosine_decay_learning_rate':
config = learning_rate_config.cosine_decay_learning_rate
lr_schedule = learning_schedules.CosineDecayWithWarmup(
FLAGS.initial_learning_rate, config.total_steps, FLAGS.warmup_steps)
if lr_schedule is None:
raise ValueError('Learning_rate %s not supported.' % learning_rate_type)
return lr_schedule
def get_dataset_fn(configs):
"""Returns tf dataset."""
def dataset_fn(input_context=None):
del input_context
train_config = configs['train_config']
train_dataset_config = configs['train_dataset']
use_tpu = (FLAGS.train_strategy == TRAIN_STRATEGY[0])
dataset = inputs.create_input(
train_config, train_dataset_config, use_tpu=use_tpu)
return dataset
return dataset_fn
def summary_fn(loss_dict, global_step):
"""Function to summarize model input and output tensors.
Args:
loss_dict: A dictionary of the losses.
global_step: Global step value.
"""
for loss_type in loss_dict:
tf.summary.scalar(loss_type, loss_dict[loss_type], step=global_step)
def distribution_strategy():
"""Returns strategy for distributed training."""
if FLAGS.train_strategy == TRAIN_STRATEGY[0]:
resolver = tf.distribute.cluster_resolver.TPUClusterResolver(
tpu=FLAGS.master)
tf.config.experimental_connect_to_cluster(resolver)
tf.tpu.experimental.initialize_tpu_system(resolver)
strategy = tf.distribute.TPUStrategy(resolver)
else:
strategy = tf.distribute.MirroredStrategy()
return strategy
def train():
"""Trains model."""
configs = config_util.get_configs_from_pipeline_file(FLAGS.config_path)
model_config = configs['model']
train_config = configs['train_config']
strategy = distribution_strategy()
dataset = strategy.distribute_datasets_from_function(get_dataset_fn(configs))
with strategy.scope():
model_ = model_builder.build(model_config, True)
lr_schedule = _create_learning_rate(train_config.learning_rate)
optimizer = tf.keras.optimizers.Adam(lr_schedule)
model_.global_step = optimizer.iterations
summaryfn = None
if FLAGS.train_strategy == TRAIN_STRATEGY[1]:
summaryfn = summary_fn
trainer = single_task_trainer.SingleTaskTrainer(
dataset,
label_key='target',
model=model_,
loss_fn=model_.loss,
optimizer=optimizer,
summary_fn=summaryfn,
grad_clip_norm=FLAGS.grad_clip_norm)
controller = orbit.Controller(
trainer=trainer,
strategy=strategy,
steps_per_loop=10,
checkpoint_manager=tf.train.CheckpointManager(
tf.train.Checkpoint(optimizer=optimizer, model=model_),
directory=FLAGS.model_dir,
checkpoint_interval=1000,
step_counter=trainer.optimizer.iterations,
max_to_keep=5),
summary_dir=FLAGS.model_dir,
summary_interval=10,
global_step=trainer.optimizer.iterations)
controller.train(1)
controller.train(FLAGS.steps - 1)
def main(_):
train()
if __name__ == '__main__':
app.run(main)