Repository navigation
Expand file tree
/
Copy pathmain.py
More file actions
157 lines (127 loc) · 5.93 KB
/
Copy pathmain.py
File metadata and controls
157 lines (127 loc) · 5.93 KB
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
# Copyright 2018 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.
import argh
import argparse
import os.path
import sys
import tempfile
import dual_net
import evaluation
import utils
import cloud_logging
import tensorflow as tf
from absl import flags
from tensorflow import gfile
# How many positions we should aggregate per 'chunk'.
EXAMPLES_PER_RECORD = 10000
# How many positions to draw from for our training window.
# AGZ used the most recent 500k games, which, assuming 250 moves/game = 125M
WINDOW_SIZE = 125000000
def bootstrap(
working_dir: 'tf.estimator working directory. If not set, defaults to a random tmp dir'=None,
model_save_path: 'Where to export the first bootstrapped generation'=None):
if working_dir is None:
with tempfile.TemporaryDirectory() as working_dir:
utils.ensure_dir_exists(working_dir)
utils.ensure_dir_exists(os.path.dirname(model_save_path))
dual_net.bootstrap(working_dir)
dual_net.export_model(working_dir, model_save_path)
else:
utils.ensure_dir_exists(working_dir)
utils.ensure_dir_exists(os.path.dirname(model_save_path))
dual_net.bootstrap(working_dir)
dual_net.export_model(working_dir, model_save_path)
freeze_graph(model_save_path)
def train_dir(
chunk_dir: 'Directory where training chunks are.',
model_save_path: 'Where to export the completed generation.'):
tf_records = sorted(gfile.Glob(os.path.join(chunk_dir, '*.tfrecord.zz')))
tf_records = tf_records[-1 * (WINDOW_SIZE // EXAMPLES_PER_RECORD):]
train(tf_records, model_save_path)
def train(tf_records: 'list of files of tf_records to train on',
model_save_path: 'Where to export the completed generation.'):
if tf_records:
print("Training on {} records: {} to {}".format(
len(tf_records), tf_records[0], tf_records[-1]))
with utils.logged_timer("Training"):
dual_net.train(*tf_records)
print("== Training done. Exporting model to ", model_save_path)
dual_net.export_model(flags.FLAGS.model_dir, model_save_path)
freeze_graph(model_save_path)
def validate(
*tf_record_dirs: 'Directories where holdout data are',
validate_name: 'Name for validation set (i.e., selfplay or human)'=None):
tf_records = []
with utils.logged_timer("Building lists of holdout files"):
for record_dir in tf_record_dirs:
tf_records.extend(gfile.Glob(os.path.join(record_dir, '*.zz')))
first_record = os.path.basename(tf_records[0])
last_record = os.path.basename(tf_records[-1])
with utils.logged_timer("Validating from {} to {}".format(first_record, last_record)):
dual_net.validate(tf_records, validate_name=validate_name)
def evaluate(
black_model: 'The path to the model to play black',
white_model: 'The path to the model to play white',
output_dir: 'Where to write the evaluation results'='sgf/evaluate',
games: 'the number of games to play'=16,
verbose: 'How verbose the players should be (see selfplay)' = 1):
utils.ensure_dir_exists(output_dir)
with utils.logged_timer("Loading weights"):
black_net = dual_net.DualNetwork(black_model)
white_net = dual_net.DualNetwork(white_model)
with utils.logged_timer("Playing game"):
evaluation.play_match(
black_net, white_net, games, output_dir, verbose)
def convert(load_file, dest_file):
from tensorflow.python.framework import meta_graph
features, labels = dual_net.get_inference_input()
dual_net.model_fn(features, labels, tf.estimator.ModeKeys.PREDICT)
sess = tf.Session()
# retrieve the global step as a python value
ckpt = tf.train.load_checkpoint(load_file)
global_step_value = ckpt.get_tensor('global_step')
# restore all saved weights, except global_step
meta_graph_def = meta_graph.read_meta_graph_file(
load_file + '.meta')
stored_var_names = set([n.name
for n in meta_graph_def.graph_def.node
if n.op == 'VariableV2'])
stored_var_names.remove('global_step')
var_list = [v for v in tf.global_variables()
if v.op.name in stored_var_names]
tf.train.Saver(var_list=var_list).restore(sess, load_file)
# manually set the global step
global_step_tensor = tf.train.get_or_create_global_step()
assign_op = tf.assign(global_step_tensor, global_step_value)
sess.run(assign_op)
# export a new savedmodel that has the right global step type
tf.train.Saver().save(sess, dest_file)
sess.close()
tf.reset_default_graph()
def freeze_graph(load_file):
""" Loads a network and serializes just the inference parts for use by e.g. the C++ binary """
n = dual_net.DualNetwork(load_file)
out_graph = tf.graph_util.convert_variables_to_constants(
n.sess, n.sess.graph.as_graph_def(), ["policy_output", "value_output"])
with gfile.GFile(os.path.join(load_file + '.pb'), 'wb') as f:
f.write(out_graph.SerializeToString())
parser = argparse.ArgumentParser()
argh.add_commands(parser, [bootstrap, train, train_dir, freeze_graph,
evaluate, validate, convert])
if __name__ == '__main__':
cloud_logging.configure()
# Let absl.flags parse known flags from argv, then pass the remaining flags
# into argh for dispatching.
remaining_argv = flags.FLAGS(sys.argv, known_only=True)
argh.dispatch(parser, argv=remaining_argv[1:])