This repository has been archived by the owner on Apr 2, 2022. It is now read-only.
-
Notifications
You must be signed in to change notification settings - Fork 232
/
task.py
291 lines (261 loc) · 9.25 KB
/
task.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
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
#!/usr/bin/env python
# Copyright 2019 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.
"""This is the Chicago Taxi example for AI Platform training."""
import argparse
import logging
import os
import sys
from datetime import datetime
import tensorflow as tf
from . import model
from . import experiment
def get_args():
"""Define the task arguments with the default values.
Returns:
experiment parameters
"""
args_parser = argparse.ArgumentParser()
# Data files arguments
args_parser.add_argument(
'--train-files',
help='GCS or local paths to training data.',
nargs='+',
required=True)
args_parser.add_argument(
'--eval-files',
help='GCS or local paths to evaluation data.',
nargs='+',
required=True)
# Experiment arguments
args_parser.add_argument(
'--train-steps',
help="""
Steps to run the training job for.
If --num-epochs and --train-size are not specified,
this must be specified;
otherwise the training job will run indefinitely.
if --num-epochs and --train-size are specified,
then --train-steps will default to:
(train-size/train-batch-size) * num-epochs
""",
default=0,
type=int)
args_parser.add_argument(
'--eval-steps',
help="""
Number of steps to run evaluation for at each checkpoint.,
Set to None to evaluate on the whole evaluation data.
""",
default=None,
type=int)
args_parser.add_argument(
'--batch-size',
help='Batch size for each training and evaluation step.',
type=int,
default=128)
args_parser.add_argument(
'--train-size',
help="""
Size of the training data (instance count).
If both --train-size and --num-epochs are specified,
--train-steps will default to:
(train-size/train-batch-size) * num-epochs.
""",
type=int,
default=None)
args_parser.add_argument(
'--num-epochs',
help="""
Maximum number of training data epochs on which to train.
If both --train-size and --num-epochs are specified,
--train-steps will default to:
(train-size/train-batch-size) * num-epochs.
""",
default=100,
type=int,
)
args_parser.add_argument(
'--eval-frequency-secs',
help='How many seconds to wait before running the next evaluation.',
default=15,
type=int)
# Feature columns arguments
args_parser.add_argument(
'--embed-categorical-columns',
help="""
If set to True, the categorical columns will be embedded
and used in the deep part of the model.
The embedding size = sqrt(vocab_size).
""",
action='store_true',
default=True,
)
args_parser.add_argument(
'--use-indicator-columns',
help="""
If set to True, the categorical columns will be encoded
as One-Hot indicators in the deep part of the model.
""",
action='store_true',
default=False,
)
args_parser.add_argument(
'--use-wide-columns',
help="""
If set to True, the categorical columns will be used in the
wide part of the model.
""",
action='store_true',
default=False,
)
# Estimator arguments
args_parser.add_argument(
'--learning-rate',
help='Learning rate value for the optimizers.',
default=0.1,
type=float)
args_parser.add_argument(
'--learning-rate-decay-factor',
help="""
The factor by which the learning rate should decay by the end of the
training.
decayed_learning_rate = learning_rate * decay_rate ^ (global_step /
decay_steps).
If set to 1.0 (default), then no decay will occur.
If set to 0.5, then the learning rate should reach 0.5 of its original
value at the end of the training.
Note that decay_steps is set to train_steps.
""",
default=1.0,
type=float)
args_parser.add_argument(
'--hidden-units',
help="""
Hidden layer sizes to use for DNN feature columns, provided in
comma-separated layers.
If --scale-factor > 0, then only the size of the first layer will be
used to compute the sizes of subsequent layers.
""",
default='30,30,30')
args_parser.add_argument(
'--layer-sizes-scale-factor',
help="""
Determine how the size of the layers in the DNN are scaled per
successive layer.
If value = 0 then the provided --hidden-units will be the
same per layer.
""",
default=0.7,
type=float)
args_parser.add_argument(
'--num-layers',
help="""
Number of layers in the DNN. If --layer-sizes-scale-factor > 0,
then this parameter is ignored.
""",
default=4,
type=int)
args_parser.add_argument(
'--dropout-prob',
help='The probability we will drop out a given coordinate.',
default=None)
# Saved model arguments
args_parser.add_argument(
'--job-dir',
help='GCS location to write checkpoints and export models.',
required=True)
args_parser.add_argument(
'--reuse-job-dir',
action='store_true',
default=False,
help="""
Flag to decide if the model checkpoint should be
re-used from the job-dir.
If set to False then the job-dir will be deleted.
""")
args_parser.add_argument(
'--serving-export-format',
help='The input format of the exported serving SavedModel.',
choices=['JSON', 'CSV', 'EXAMPLE'],
default='JSON')
args_parser.add_argument(
'--eval-export-format',
help='The input format of the exported evaluating SavedModel.',
choices=['CSV', 'EXAMPLE'],
default='CSV')
return args_parser.parse_args()
def _setup_logging():
"""Sets up logging."""
root_logger = logging.getLogger()
root_logger_previous_handlers = list(root_logger.handlers)
for h in root_logger_previous_handlers:
root_logger.removeHandler(h)
root_logger.setLevel(logging.INFO)
root_logger.propagate = False
# Set tf logging to avoid duplicate logging. If the handlers are not removed
# then we will have duplicate logging
tf_logger = logging.getLogger('TensorFlow')
while tf_logger.handlers:
tf_logger.removeHandler(tf_logger.handlers[0])
# Redirect INFO logs to stdout
stdout_handler = logging.StreamHandler(sys.stdout)
stdout_handler.setLevel(logging.INFO)
root_logger.addHandler(stdout_handler)
# Suppress C++ level warnings.
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3'
def main():
args = get_args()
_setup_logging()
# If job_dir_reuse is False then remove the job_dir if it exists.
logging.info('Resume training: {}'.format(args.reuse_job_dir))
if not args.reuse_job_dir:
if tf.io.gfile.exists(args.job_dir):
tf.io.gfile.rmtree(args.job_dir)
logging.info(
'Deleted job_dir {} to avoid re-use'.format(args.job_dir))
else:
logging.info('Reusing job_dir {} if it exists'.format(args.job_dir))
run_config = experiment.create_run_config(args)
logging.info('Job directory: {}'.format(run_config.model_dir))
# Compute the number of Training steps.
if args.train_size is not None and args.num_epochs is not None:
args.train_steps = int(
(args.train_size / args.batch_size) * args.num_epochs)
else:
args.train_steps = args.train_steps
logging.info('Train size: {}.'.format(args.train_size))
logging.info('Epoch count: {}.'.format(args.num_epochs))
logging.info('Batch size: {}.'.format(args.batch_size))
logging.info('Training steps: {} ({}).'.format(
args.train_steps,
'supplied' if args.train_size is None else 'computed'))
logging.info('Evaluate every {} steps.'.format(args.eval_frequency_secs))
# Create the Estimator
estimator = model.create(args, run_config)
logging.info('Creating an Estimator: {}'.format(type(estimator)))
# Run the train and evaluate experiment
time_start = datetime.utcnow()
logging.info('Experiment started...')
logging.info('.......................................')
# Run experiment
experiment.run(estimator, args)
time_end = datetime.utcnow()
logging.info('.......................................')
logging.info('Experiment finished.')
time_elapsed = time_end - time_start
logging.info('Experiment elapsed time: {} seconds'.format(
time_elapsed.total_seconds()))
if __name__ == '__main__':
main()