def main(cmd):
cfg = yaml.safe_load(cmd.cfg.read())
print(yaml.dump(cfg, default_flow_style=False))
num_chunks = cfg['dataset']['num_chunks']
allow_less = cfg['dataset'].get('allow_less_chunks', False)
train_ratio = cfg['dataset']['train_ratio']
num_train = int(num_chunks*train_ratio)
num_test = num_chunks - num_train
if 'input_test' in cfg['dataset']:
train_chunks = get_latest_chunks(cfg['dataset']['input_train'], num_train, allow_less)
test_chunks = get_latest_chunks(cfg['dataset']['input_test'], num_test, allow_less)
else:
chunks = get_latest_chunks(cfg['dataset']['input'], num_chunks, allow_less)
if allow_less:
num_train = int(len(chunks)*train_ratio)
num_test = len(chunks) - num_train
train_chunks = chunks[:num_train]
test_chunks = chunks[num_train:]
shuffle_size = cfg['training']['shuffle_size']
total_batch_size = cfg['training']['batch_size']
batch_splits = cfg['training'].get('num_batch_splits', 1)
if total_batch_size % batch_splits != 0:
raise ValueError('num_batch_splits must divide batch_size evenly')
split_batch_size = total_batch_size // batch_splits
# Load data with split batch size, which will be combined to the total batch size in tfprocess.
ChunkParser.BATCH_SIZE = split_batch_size
root_dir = os.path.join(cfg['training']['path'], cfg['name'])
if not os.path.exists(root_dir):
os.makedirs(root_dir)
train_parser = ChunkParser(FileDataSrc(train_chunks),
shuffle_size=shuffle_size, sample=SKIP, batch_size=ChunkParser.BATCH_SIZE)
dataset = tf.data.Dataset.from_generator(
train_parser.parse, output_types=(tf.string, tf.string, tf.string, tf.string))
dataset = dataset.map(ChunkParser.parse_function)
dataset = dataset.prefetch(4)
train_iterator = dataset.make_one_shot_iterator()
shuffle_size = int(shuffle_size*(1.0-train_ratio))
test_parser = ChunkParser(FileDataSrc(test_chunks),
shuffle_size=shuffle_size, sample=SKIP, batch_size=ChunkParser.BATCH_SIZE)
dataset = tf.data.Dataset.from_generator(
test_parser.parse, output_types=(tf.string, tf.string, tf.string, tf.string))
dataset = dataset.map(ChunkParser.parse_function)
dataset = dataset.prefetch(4)
test_iterator = dataset.make_one_shot_iterator()
tfprocess = TFProcess(cfg)
tfprocess.init(dataset, train_iterator, test_iterator)
if os.path.exists(os.path.join(root_dir, 'checkpoint')):
cp = tf.train.latest_checkpoint(root_dir)
tfprocess.restore(cp)
# If number of test positions is not given
# sweeps through all test chunks statistically
# Assumes average of 10 samples per test game.
# For simplicity, testing can use the split batch size instead of total batch size.
# This does not affect results, because test results are simple averages that are independent of batch size.
num_evals = cfg['training'].get('num_test_positions', len(test_chunks) * 10)
num_evals = max(1, num_evals // ChunkParser.BATCH_SIZE)
print("Using {} evaluation batches".format(num_evals))
tfprocess.process_loop(total_batch_size, num_evals, batch_splits=batch_splits)
if cmd.output is not None:
if cfg['training'].get('swa_output', False):
tfprocess.save_swa_weights(cmd.output)
else:
tfprocess.save_leelaz_weights(cmd.output)
tfprocess.session.close()
train_parser.shutdown()
test_parser.shutdown()
def main(cmd):
cfg = yaml.safe_load(cmd.cfg.read())
print(yaml.dump(cfg, default_flow_style=False))
num_chunks = cfg['dataset']['num_chunks']
allow_less = cfg['dataset'].get('allow_less_chunks', False)
train_ratio = cfg['dataset']['train_ratio']
num_train = int(num_chunks * train_ratio)
num_test = num_chunks - num_train
sort_type = cfg['dataset'].get('sort_type', 'mtime')
if sort_type == 'mtime':
sort_key_fn = os.path.getmtime
elif sort_type == 'number':
sort_key_fn = game_number_for_name
elif sort_type == 'name':
sort_key_fn = identity_function
else:
raise ValueError('Unknown dataset sort_type: {}'.format(sort_type))
if 'input_test' in cfg['dataset']:
train_chunks = get_latest_chunks(cfg['dataset']['input_train'],
num_train, allow_less, sort_key_fn)
test_chunks = get_latest_chunks(cfg['dataset']['input_test'], num_test,
allow_less, sort_key_fn)
else:
chunks = get_latest_chunks(cfg['dataset']['input'], num_chunks,
allow_less, sort_key_fn)
if allow_less:
num_train = int(len(chunks) * train_ratio)
num_test = len(chunks) - num_train
train_chunks = chunks[:num_train]
test_chunks = chunks[num_train:]
shuffle_size = cfg['training']['shuffle_size']
total_batch_size = cfg['training']['batch_size']
batch_splits = cfg['training'].get('num_batch_splits', 1)
train_workers = cfg['dataset'].get('train_workers', None)
test_workers = cfg['dataset'].get('test_workers', None)
if total_batch_size % batch_splits != 0:
raise ValueError('num_batch_splits must divide batch_size evenly')
split_batch_size = total_batch_size // batch_splits
diff_focus_min = cfg['training'].get('diff_focus_min', 1)
diff_focus_slope = cfg['training'].get('diff_focus_slope', 0)
diff_focus_q_weight = cfg['training'].get('diff_focus_q_weight', 6.0)
diff_focus_pol_scale = cfg['training'].get('diff_focus_pol_scale', 3.5)
root_dir = os.path.join(cfg['training']['path'], cfg['name'])
if not os.path.exists(root_dir):
os.makedirs(root_dir)
train_parser = ChunkParser(train_chunks,
get_input_mode(cfg),
shuffle_size=shuffle_size,
sample=SKIP,
batch_size=split_batch_size,
diff_focus_min=diff_focus_min,
diff_focus_slope=diff_focus_slope,
diff_focus_q_weight=diff_focus_q_weight,
diff_focus_pol_scale=diff_focus_pol_scale,
workers=train_workers)
test_shuffle_size = int(shuffle_size * (1.0 - train_ratio))
# no diff focus for test_parser
test_parser = ChunkParser(test_chunks,
get_input_mode(cfg),
shuffle_size=test_shuffle_size,
sample=SKIP,
batch_size=split_batch_size,
workers=test_workers)
if 'input_validation' in cfg['dataset']:
valid_chunks = get_all_chunks(cfg['dataset']['input_validation'])
validation_parser = ChunkParser(valid_chunks,
get_input_mode(cfg),
sample=1,
batch_size=split_batch_size,
workers=0)
import tensorflow as tf
from chunkparsefunc import parse_function
from tfprocess import TFProcess
tfprocess = TFProcess(cfg)
train_dataset = tf.data.Dataset.from_generator(
train_parser.parse,
output_types=(tf.string, tf.string, tf.string, tf.string, tf.string))
train_dataset = train_dataset.map(parse_function)
test_dataset = tf.data.Dataset.from_generator(
test_parser.parse,
output_types=(tf.string, tf.string, tf.string, tf.string, tf.string))
test_dataset = test_dataset.map(parse_function)
validation_dataset = None
if 'input_validation' in cfg['dataset']:
validation_dataset = tf.data.Dataset.from_generator(
validation_parser.sequential,
output_types=(tf.string, tf.string, tf.string, tf.string,
tf.string))
validation_dataset = validation_dataset.map(parse_function)
if tfprocess.strategy is None: #Mirrored strategy appends prefetch itself with a value depending on number of replicas
train_dataset = train_dataset.prefetch(4)
test_dataset = test_dataset.prefetch(4)
if validation_dataset is not None:
validation_dataset = validation_dataset.prefetch(4)
else:
options = tf.data.Options()
options.experimental_distribute.auto_shard_policy = tf.data.experimental.AutoShardPolicy.OFF
train_dataset = train_dataset.with_options(options)
test_dataset = test_dataset.with_options(options)
if validation_dataset is not None:
validation_dataset = validation_dataset.with_options(options)
tfprocess.init(train_dataset, test_dataset, validation_dataset)
tfprocess.restore()
# If number of test positions is not given
# sweeps through all test chunks statistically
# Assumes average of 10 samples per test game.
# For simplicity, testing can use the split batch size instead of total batch size.
# This does not affect results, because test results are simple averages that are independent of batch size.
num_evals = cfg['training'].get('num_test_positions',
len(test_chunks) * 10)
num_evals = max(1, num_evals // split_batch_size)
print("Using {} evaluation batches".format(num_evals))
tfprocess.total_batch_size = total_batch_size
tfprocess.process_loop(total_batch_size,
num_evals,
batch_splits=batch_splits)
if cmd.output is not None:
if cfg['training'].get('swa_output', False):
tfprocess.save_swa_weights(cmd.output)
else:
tfprocess.save_leelaz_weights(cmd.output)
train_parser.shutdown()
test_parser.shutdown()