def main(args):
trainning = GameArchive('../../data/train1.data')
train_parser = ChunkParser(trainning.games)
dataset = tf.data.Dataset.from_generator(train_parser.parse_chunk,
output_types=(tf.string))
dataset = dataset.shuffle(1 << 18)
dataset = dataset.map(_parse_function)
dataset = dataset.batch(BATCH_SIZE)
dataset = dataset.prefetch(4)
train_iterator = dataset.make_one_shot_iterator()
testing = GameArchive('../../data/test1.data')
test_parser = ChunkParser(testing.games)
dataset = tf.data.Dataset.from_generator(test_parser.parse_chunk,
output_types=(tf.string))
dataset = dataset.map(_parse_function)
dataset = dataset.batch(BATCH_SIZE)
dataset = dataset.prefetch(4)
test_iterator = dataset.make_one_shot_iterator()
tfprocess = TFProcess()
tfprocess.init(dataset, train_iterator, test_iterator)
if args:
restore_file = args.pop(0)
tfprocess.restore(restore_file)
while True:
tfprocess.process(BATCH_SIZE)
def main(args):
train_data_prefix = args.pop(0)
chunks = get_chunks(train_data_prefix)
print("Found {0} chunks".format(len(chunks)))
if not chunks:
return
parser = ChunkParser(chunks)
run_test(parser)
#benchmark(parser)
dataset = tf.data.Dataset.from_generator(
parser.parse_chunk, output_types=(tf.string))
dataset = dataset.shuffle(65536)
dataset = dataset.map(_parse_function)
dataset = dataset.batch(BATCH_SIZE)
dataset = dataset.prefetch(16)
iterator = dataset.make_one_shot_iterator()
next_batch = iterator.get_next()
tfprocess = TFProcess(next_batch)
if args:
restore_file = args.pop(0)
tfprocess.restore(restore_file)
while True:
tfprocess.process(BATCH_SIZE)
def main(cmd):
cfg = yaml.safe_load(cmd.cfg.read())
print(yaml.dump(cfg, default_flow_style=False))
root_dir = os.path.join(cfg['training']['path'], cfg['name'])
if not os.path.exists(root_dir):
os.makedirs(root_dir)
x = [
tf.placeholder(tf.float32, [None, 112, 8 * 8]),
tf.placeholder(tf.float32, [None, 1858]),
tf.placeholder(tf.float32, [None, 3]),
tf.placeholder(tf.float32, [None, 3]),
]
tfprocess = TFProcess(cfg)
tfprocess.init_net(x)
if os.path.exists(os.path.join(root_dir, 'checkpoint')):
cp = tf.train.latest_checkpoint(root_dir)
tfprocess.restore(cp)
START_FROM = cmd.start
update_global_step = tfprocess.global_step.assign(START_FROM)
tfprocess.session.run(update_global_step)
path = os.path.join(root_dir, cfg['name'])
save_path = tfprocess.saver.save(tfprocess.session,
path,
global_step=START_FROM)
tfprocess.session.close()
def main(args):
train_data_prefix = args.pop(0)
chunks = get_chunks(train_data_prefix)
print("Found {0} chunks".format(len(chunks)))
if not chunks:
return
parser = ChunkParser(chunks)
dataset = tf.data.Dataset.from_generator(parser.parse_chunk,
output_types=(tf.float32,
tf.float32,
tf.float32))
dataset = dataset.shuffle(65536)
dataset = dataset.batch(BATCH_SIZE)
dataset = dataset.prefetch(16)
iterator = dataset.make_one_shot_iterator()
next_batch = iterator.get_next()
tfprocess = TFProcess(next_batch)
if args:
restore_file = args.pop(0)
tfprocess.restore(restore_file)
while True:
tfprocess.process(BATCH_SIZE)
def main(args):
chunks = latest_chunks()
parser, next_batch = chunks2batches(chunks)
tfprocess = TFProcess(next_batch)
if args:
restore_file = args.pop(0)
print("Restoring weights ....")
tfprocess.restore(restore_file)
print("Training starts ....")
while True:
change_data, run_val = tfprocess.process()
if change_data:
chunks = latest_chunks()
parser.chunk_switch(chunks)
if run_val:
best_net = leela_conf.SAVE_DIR + "/best.txt"
last_net = leela_conf.SAVE_DIR + "/latest.txt"
cmd = leela_conf.VALIDATION_COMMAND % \
(last_net,
best_net)
print(cmd)
subprocess.call(cmd.split(" ")) #, stdout=subprocess.PIPE)
with open(leela_conf.VALIDATION_LOG, "r") as f:
better = int(f.readlines()[-1].split("\t")[0])
if better:
print("---------------- Better Network Found! --------------")
copy2(last_net, best_net)
else:
print("------------- Checkout best net so far. -------------")
tfprocess.replace_weights(get_weights(best_net))
def main(cmd):
cfg = yaml.safe_load(cmd.cfg.read())
print(yaml.dump(cfg, default_flow_style=False))
num_chunks = cfg['dataset']['num_chunks']
chunks = get_latest_chunks(cfg['dataset']['input'], num_chunks)
train_ratio = cfg['dataset']['train_ratio']
num_train = int(num_chunks * train_ratio)
shuffle_size = cfg['training']['shuffle_size']
ChunkParser.BATCH_SIZE = cfg['training']['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(chunks[:num_train]),
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))
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(chunks[num_train:]),
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))
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 = get_checkpoint(root_dir)
tfprocess.restore(cp)
# Sweeps through all test chunks statistically
num_evals = (num_chunks - num_train) * 10 // ChunkParser.BATCH_SIZE
print("Using {} evaluation batches".format(num_evals))
for _ in range(cfg['training']['total_steps']):
tfprocess.process(ChunkParser.BATCH_SIZE, num_evals)
tfprocess.save_leelaz_weights(cmd.output)
tfprocess.session.close()
train_parser.shutdown()
test_parser.shutdown()
def main():
if len(sys.argv) != 2:
print("Usage: {} config.yaml".format(sys.argv[0]))
return 1
cfg = yaml.safe_load(open(sys.argv[1], 'r').read())
print(yaml.dump(cfg, default_flow_style=False))
batch_size = cfg['training']['batch_size']
filename = os.path.join(cfg['dataset']['path'], 'train.bin')
train_next_batch, parser = dataset_iterator(filename, batch_size)
print("Creating trainingset from {}".format(filename))
num_eval = parser.num_samples() // batch_size
print("Train epoch in {} steps".format(num_eval))
filename = os.path.join(cfg['dataset']['path'], 'test.bin')
test_next_batch, parser = dataset_iterator(filename, batch_size)
print("Creating testset from {}".format(filename))
num_eval = parser.num_samples() // batch_size
print("Test epoch in {} steps".format(num_eval))
tfprocess = TFProcess(cfg, train_next_batch, test_next_batch, num_eval)
root_dir = os.path.join(cfg['training']['path'], cfg['name'])
if os.path.exists(os.path.join(root_dir, 'checkpoint')):
checkpoint = parse.get_checkpoint(root_dir)
tfprocess.restore(checkpoint)
if not os.path.exists(root_dir):
os.makedirs(root_dir)
print("Created output directory: {}".format(root_dir))
while True:
tfprocess.process(batch_size)
def main():
parser = argparse.ArgumentParser(
description='Train network from game data.')
parser.add_argument("trainpref",
help='Training file prefix', nargs='?', type=str)
parser.add_argument("restorepref",
help='Training snapshot prefix', nargs='?', type=str)
parser.add_argument("--train", '-t',
help="Training file prefix", type=str)
parser.add_argument("--test", help="Test file prefix", type=str)
parser.add_argument("--restore", type=str,
help="Prefix of tensorflow snapshot to restore from")
parser.add_argument("--logbase", default='leelalogs', type=str,
help="Log file prefix (for tensorboard)")
parser.add_argument("--sample", default=DOWN_SAMPLE, type=int,
help="Rate of data down-sampling to use")
args = parser.parse_args()
train_data_prefix = args.train or args.trainpref
restore_prefix = args.restore or args.restorepref
training = get_chunks(train_data_prefix)
if not args.test:
# Generate test by taking 10% of the training chunks.
random.shuffle(training)
training, test = split_chunks(training, 0.1)
else:
test = get_chunks(args.test)
if not training:
print("No data to train on!")
return
print("Training with {0} chunks, validating on {1} chunks".format(
len(training), len(test)))
train_parser = ChunkParser(FileDataSrc(training),
shuffle_size=1<<20, # 2.2GB of RAM.
sample=args.sample,
batch_size=RAM_BATCH_SIZE).parse()
test_parser = ChunkParser(FileDataSrc(test),
shuffle_size=1<<19,
sample=args.sample,
batch_size=RAM_BATCH_SIZE).parse()
tfprocess = TFProcess()
tfprocess.init(RAM_BATCH_SIZE,
logbase=args.logbase,
macrobatch=BATCH_SIZE // RAM_BATCH_SIZE)
#benchmark1(tfprocess)
if restore_prefix:
tfprocess.restore(restore_prefix)
tfprocess.process(train_parser, test_parser)
def main(args):
train_data_prefix = args.pop(0)
chunks = get_chunks(train_data_prefix)
print("Found {0} chunks".format(len(chunks)))
if chunks:
tfprocess = TFProcess()
if args:
restore_file = args.pop(0)
tfprocess.restore(restore_file)
do_train_loop(chunks, tfprocess)
def main(cmd):
cfg = yaml.safe_load(cmd.cfg.read())
print(yaml.dump(cfg, default_flow_style=False))
num_chunks = cfg['dataset']['num_chunks']
chunks = get_latest_chunks(cfg['dataset']['input'], num_chunks)
train_ratio = cfg['dataset']['train_ratio']
num_train = int(num_chunks*train_ratio)
shuffle_size = cfg['training']['shuffle_size']
ChunkParser.BATCH_SIZE = cfg['training']['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(chunks[:num_train]),
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))
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(chunks[num_train:]),
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))
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 = get_checkpoint(root_dir)
tfprocess.restore(cp)
# Sweeps through all test chunks statistically
num_evals = (num_chunks-num_train)*10 // ChunkParser.BATCH_SIZE
print("Using {} evaluation batches".format(num_evals))
for _ in range(cfg['training']['total_steps']):
tfprocess.process(ChunkParser.BATCH_SIZE, num_evals)
tfprocess.save_leelaz_weights(cmd.output)
tfprocess.session.close()
train_parser.shutdown()
test_parser.shutdown()
def main():
if len(sys.argv) != 2:
print("Usage: {} config.yaml".format(sys.argv[0]))
return 1
cfg = yaml.safe_load(open(sys.argv[1], 'r').read())
print(yaml.dump(cfg, default_flow_style=False))
num_chunks = cfg['dataset']['num_chunks']
chunks = get_latest_chunks(cfg['dataset']['input'], num_chunks)
num_train = int(num_chunks*cfg['dataset']['train_ratio'])
shuffle_size = cfg['training']['shuffle_size']
ChunkParser.BATCH_SIZE = cfg['training']['batch_size']
root_dir = os.path.join(cfg['training']['path'], cfg['name'])
if not os.path.exists(root_dir):
os.makedirs(root_dir)
#bench_parser = ChunkParser(FileDataSrc(chunks[:1000]), shuffle_size=1<<14, sample=SKIP, batch_size=ChunkParser.BATCH_SIZE)
#benchmark(bench_parser)
train_parser = ChunkParser(FileDataSrc(chunks[:num_train]),
shuffle_size=shuffle_size, sample=SKIP, batch_size=ChunkParser.BATCH_SIZE)
#benchmark(train_parser)
dataset = tf.data.Dataset.from_generator(
train_parser.parse, output_types=(tf.string, tf.string, tf.string))
dataset = dataset.map(ChunkParser.parse_function)
dataset = dataset.prefetch(4)
train_iterator = dataset.make_one_shot_iterator()
test_parser = ChunkParser(FileDataSrc(chunks[num_train:]), batch_size=ChunkParser.BATCH_SIZE)
dataset = tf.data.Dataset.from_generator(
test_parser.parse, output_types=(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 = get_checkpoint(root_dir)
tfprocess.restore(cp)
# Sweeps through all test chunks statistically
num_evals = int(round(((num_chunks-num_train) * (200 / SKIP)) / ChunkParser.BATCH_SIZE))
print("Using {} evaluation batches".format(num_evals))
# while True:
for _ in range(cfg['training']['total_steps']):
tfprocess.process(ChunkParser.BATCH_SIZE, num_evals)
def main(args):
train_data_prefix = args.pop(0)
chunks = get_chunks(train_data_prefix)
print("Found {0} chunks".format(len(chunks)))
if not chunks:
return
# The following assumes positions from one game are not
# spread through chunks.
random.shuffle(chunks)
training, test = split_chunks(chunks, 0.1)
print("Training with {0} chunks, validating on {1} chunks".format(
len(training), len(test)))
train_parser = ChunkParser(FileDataSrc(training),
shuffle_size=1 << 19,
sample=DOWN_SAMPLE,
batch_size=BATCH_SIZE)
#benchmark(train_parser)
dataset = tf.data.Dataset.from_generator(train_parser.parse,
output_types=(tf.string,
tf.string,
tf.string))
dataset = dataset.map(_parse_function)
dataset = dataset.prefetch(4)
train_iterator = dataset.make_one_shot_iterator()
test_parser = ChunkParser(FileDataSrc(test),
shuffle_size=1 << 19,
sample=DOWN_SAMPLE,
batch_size=BATCH_SIZE)
dataset = tf.data.Dataset.from_generator(test_parser.parse,
output_types=(tf.string,
tf.string,
tf.string))
dataset = dataset.map(_parse_function)
dataset = dataset.prefetch(4)
test_iterator = dataset.make_one_shot_iterator()
tfprocess = TFProcess()
tfprocess.init(dataset, train_iterator, test_iterator)
#benchmark1(tfprocess)
if args:
restore_file = args.pop(0)
tfprocess.restore(restore_file)
while True:
tfprocess.process(BATCH_SIZE)
def main(args):
train_data_prefix = args.pop(0)
chunks = get_chunks(train_data_prefix)
print("Found {0} chunks".format(len(chunks)))
if not chunks:
return
# The following assumes positions from one game are not
# spread through chunks.
random.shuffle(chunks)
training, test = split_chunks(chunks, 0.1)
print("Training with {0} chunks, validating on {1} chunks".format(
len(training), len(test)))
#run_test(parser)
#benchmark(parser)
train_parser = ChunkParser(training)
dataset = tf.data.Dataset.from_generator(train_parser.parse_chunk,
output_types=(tf.string))
dataset = dataset.shuffle(1 << 18)
dataset = dataset.map(_parse_function)
dataset = dataset.batch(BATCH_SIZE)
dataset = dataset.prefetch(4)
train_iterator = dataset.make_one_shot_iterator()
test_parser = ChunkParser(test)
dataset = tf.data.Dataset.from_generator(test_parser.parse_chunk,
output_types=(tf.string))
dataset = dataset.map(_parse_function)
dataset = dataset.batch(BATCH_SIZE)
dataset = dataset.prefetch(4)
test_iterator = dataset.make_one_shot_iterator()
tfprocess = TFProcess()
tfprocess.init(dataset, train_iterator, test_iterator)
if args:
restore_file = args.pop(0)
tfprocess.restore(restore_file)
for _ in range(12001):
tfprocess.process(BATCH_SIZE)
for x in train_parser.mp_instances:
x.terminate()
x.join()
os.killpg(0, signal.SIGTERM)
def main(cmd):
cfg = yaml.safe_load(cmd.cfg.read())
print(yaml.dump(cfg, default_flow_style=False))
root_dir = os.path.join(cfg['training']['path'], cfg['name'])
if not os.path.exists(root_dir):
os.makedirs(root_dir)
tfprocess = TFProcess(cfg)
tfprocess.init_net()
tfprocess.restore()
START_FROM = cmd.start
tfprocess.global_step.assign(START_FROM)
tfprocess.manager.save(checkpoint_number=START_FROM)
def main():
if len(sys.argv) != 2:
print("Usage: {} config.yaml".format(sys.argv[0]))
return 1
cfg = yaml.safe_load(open(sys.argv[1], 'r').read())
print(yaml.dump(cfg, default_flow_style=False))
chunks = get_chunks(cfg['dataset']['input'])
print("Found {0} chunks".format(len(chunks)))
if not chunks:
return
parser = ChunkParser(chunks)
run_test(parser)
#benchmark(parser)
dataset = tf.data.Dataset.from_generator(parser.parse_chunk,
output_types=(tf.string))
dataset = dataset.shuffle(65536)
dataset = dataset.map(_parse_function)
dataset = dataset.batch(cfg['training']['batch_size'])
dataset = dataset.prefetch(16)
iterator = dataset.make_one_shot_iterator()
next_batch = iterator.get_next()
tfprocess = TFProcess(cfg, next_batch)
root_dir = os.path.join(cfg['training']['path'], cfg['name'])
if os.path.exists(os.path.join(root_dir, 'checkpoint')):
checkpoint = get_checkpoint(root_dir)
tfprocess.restore(checkpoint)
while True:
tfprocess.process(cfg['training']['batch_size'])
def main(cmd):
cfg = yaml.safe_load(cmd.cfg.read())
print(yaml.dump(cfg, default_flow_style=False))
num_chunks = cfg['dataset']['num_chunks']
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)
test_chunks = get_latest_chunks(cfg['dataset']['input_test'], num_test)
else:
chunks = get_latest_chunks(cfg['dataset']['input'], num_chunks)
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', num_test * 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:
tfprocess.save_leelaz_weights(cmd.output)
tfprocess.session.close()
train_parser.shutdown()
test_parser.shutdown()
def main():
parser = argparse.ArgumentParser(
description='Train network from game data.')
parser.add_argument("blockspref",
help="Number of blocks",
nargs='?',
type=int)
parser.add_argument("filterspref",
help="Number of filters",
nargs='?',
type=int)
parser.add_argument("trainpref",
help='Training file prefix',
nargs='?',
type=str)
parser.add_argument("restorepref",
help='Training snapshot prefix',
nargs='?',
type=str)
parser.add_argument("--blocks", '-b', help="Number of blocks", type=int)
parser.add_argument("--filters", '-f', help="Number of filters", type=int)
parser.add_argument("--train", '-t', help="Training file prefix", type=str)
parser.add_argument("--test", help="Test file prefix", type=str)
parser.add_argument("--restore",
type=str,
help="Prefix of tensorflow snapshot to restore from")
parser.add_argument(
"--logbase",
default='leelalogs',
type=str,
help="Log file prefix (for tensorboard) (default: %(default)s)")
parser.add_argument(
"--sample",
default=DOWN_SAMPLE,
type=int,
help="Rate of data down-sampling to use (default: %(default)d)")
args = parser.parse_args()
blocks = args.blocks or args.blockspref
filters = args.filters or args.filterspref
train_data_prefix = args.train or args.trainpref
restore_prefix = args.restore or args.restorepref
if not blocks or not filters:
print("Must supply number of blocks and filters")
return
training = get_chunks(train_data_prefix)
if not args.test:
# Generate test by taking 10% of the training chunks.
random.shuffle(training)
print("here1")
training, test = split_chunks(training, 0.1)
else:
test = get_chunks(args.test)
if not training:
print("No data to train on!")
return
print("Training with {0} chunks, validating on {1} chunks".format(
len(training), len(test)))
train_parser = ChunkParser(
FileDataSrc(training),
shuffle_size=1 << 20, # 2.2GB of RAM.
sample=args.sample,
batch_size=RAM_BATCH_SIZE).parse()
test_parser = ChunkParser(FileDataSrc(test),
shuffle_size=1 << 19,
sample=args.sample,
batch_size=RAM_BATCH_SIZE).parse()
tfprocess = TFProcess(blocks, filters)
tfprocess.init(RAM_BATCH_SIZE,
logbase=args.logbase,
macrobatch=BATCH_SIZE // RAM_BATCH_SIZE)
#benchmark1(tfprocess)
if restore_prefix:
tfprocess.restore(restore_prefix)
tfprocess.process(train_parser, test_parser)
def main(cmd):
cfg = yaml.safe_load(cmd.cfg.read())
print(yaml.dump(cfg, default_flow_style=False))
num_chunks = cfg['dataset']['num_chunks']
chunks = get_latest_chunks(cfg['dataset']['input'], num_chunks)
train_ratio = cfg['dataset']['train_ratio']
num_train = int(num_chunks*train_ratio)
shuffle_size = cfg['training']['shuffle_size']
ChunkParser.BATCH_SIZE = cfg['training']['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(chunks[:num_train]),
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))
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(chunks[num_train:]),
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))
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 = get_checkpoint(root_dir)
tfprocess.restore(cp)
# Sweeps through all test chunks statistically
num_evals = (num_chunks-num_train)*10 // ChunkParser.BATCH_SIZE
print("Using {} evaluation batches".format(num_evals))
for _ in range(cfg['training']['total_steps']):
tfprocess.process(ChunkParser.BATCH_SIZE, num_evals)
tfprocess.save_leelaz_weights('/tmp/weights.txt')
with open('/tmp/weights.txt', 'rb') as f:
m = hashlib.sha256()
w = f.read()
m.update(w)
digest = m.hexdigest()
filename = '/tmp/{}.gz'.format(digest)
with gzip.open(filename, 'wb') as f:
f.write(w)
if cmd.upload:
metadata = {'training_id':'1', 'layers':cfg['model']['residual_blocks'],
'filters':cfg['model']['filters']}
print("\nUploading `{}'...".format(digest[:8]), end='')
upload(cmd.upload, metadata, filename)
print("[done]\n")
else:
print("\nStored `{}'\n".format(filename))
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()
def main():
parser = argparse.ArgumentParser(
description='Train network from game data.')
parser.add_argument("blockspref",
help="Number of blocks", nargs='?', type=int)
parser.add_argument("filterspref",
help="Number of filters", nargs='?', type=int)
parser.add_argument("trainpref",
help='Training file prefix', nargs='?', type=str)
parser.add_argument("restorepref",
help='Training snapshot prefix', nargs='?', type=str)
parser.add_argument("--blocks", '-b',
help="Number of blocks", type=int)
parser.add_argument("--filters", '-f',
help="Number of filters", type=int)
parser.add_argument("--train", '-t',
help="Training file prefix", type=str)
parser.add_argument("--test", help="Test file prefix", type=str)
parser.add_argument("--restore", type=str,
help="Prefix of tensorflow snapshot to restore from")
parser.add_argument("--logbase", default='leelalogs', type=str,
help="Log file prefix (for tensorboard) (default: %(default)s)")
parser.add_argument("--sample", default=DOWN_SAMPLE, type=int,
help="Rate of data down-sampling to use (default: %(default)d)")
parser.add_argument("--bufferbits", default=TRAIN_SHUFFLE_BITS, type=int,
help="Train shuffle-buffer size in bits (default: %(default)d)")
parser.add_argument("--rate", default=LEARN_RATE, type=float,
help="Learning rate (default: %(default)f)")
parser.add_argument("--steps", default=TRAINING_STEPS, type=int,
help="Training step before writing a network (default: %(default)d)")
parser.add_argument("--maxsteps", default=MAX_TRAINING_STEPS, type=int,
help="Terminates after this many steps (default: %(default)d)")
parser.add_argument("--maxkeep", default=MAX_SAVER_TO_KEEP, type=int,
help="Keeps meta files for at most this many networks (default: %(default)d)")
parser.add_argument("--policyloss", default=POLICY_LOSS_WT, type=float,
help="Coefficient for policy term in loss function (default: %(default)f)")
parser.add_argument("--mseloss", default=MSE_LOSS_WT, type=float,
help="Coefficient for mse term in loss function (default: %(default)f)")
parser.add_argument("--regloss", default=REG_LOSS_WT, type=float,
help="Coefficient for regularizing term in loss function (default: %(default)f)")
args = parser.parse_args()
blocks = args.blocks or args.blockspref
filters = args.filters or args.filterspref
train_data_prefix = args.train or args.trainpref
restore_prefix = args.restore or args.restorepref
if not blocks or not filters:
print("Must supply number of blocks and filters")
return
training = get_chunks(train_data_prefix)
if not args.test:
# Generate test by taking 10% of the training chunks.
random.shuffle(training)
training, test = split_chunks(training, 0.1)
else:
test = get_chunks(args.test)
if not training:
print("No data to train on!")
return
print("Training with {0} chunks, validating on {1} chunks".format(
len(training), len(test)))
train_parser = ChunkParser(FileDataSrc(training),
shuffle_size=1<<args.bufferbits, # was 20 -- 2.2GB of RAM.
sample=args.sample,
batch_size=RAM_BATCH_SIZE).parse()
test_parser = ChunkParser(FileDataSrc(test),
shuffle_size=1<<(args.bufferbits-3), # was 19
sample=args.sample,
batch_size=RAM_BATCH_SIZE).parse()
tfprocess = TFProcess(blocks, filters,
args.rate, args.steps, args.maxsteps, args.maxkeep,
args.policyloss, args.mseloss, args.regloss)
tfprocess.init(RAM_BATCH_SIZE,
logbase=args.logbase,
macrobatch=BATCH_SIZE // RAM_BATCH_SIZE)
#benchmark1(tfprocess)
if restore_prefix:
tfprocess.restore(restore_prefix)
tfprocess.process(train_parser, test_parser)
def main(cmd):
cfg = yaml.safe_load(cmd.cfg.read())
print(yaml.dump(cfg, default_flow_style=False))
num_chunks = cfg['dataset']['num_chunks']
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)
test_chunks = get_latest_chunks(cfg['dataset']['input_test'], num_test)
else:
chunks = get_latest_chunks(cfg['dataset']['input'], num_chunks)
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)
def map_fn(example_proto):
"""
planes <tf.Tensor 'Reshape:0' shape=(2048, 112, 64) dtype=float32>
probs <tf.Tensor 'Reshape_1:0' shape=(2048, 1858) dtype=float32>
winner <tf.Tensor 'Reshape_2:0' shape=(2048, 1) dtype=float32>
"""
tfrecord_features = {
'planes': tf.FixedLenFeature((112, 64), tf.float32),
'probs': tf.FixedLenFeature((1858), tf.float32),
'winner': tf.FixedLenFeature([], tf.float32)
}
parsed_features = tf.parse_single_example(example_proto,
tfrecord_features)
planes, probs, winner = parsed_features["planes"], parsed_features[
"probs"], parsed_features["winner"]
return planes, probs, winner
filenames = [
"gs://jeremylorino-staging-bq-data/chess/lczero_data/train.tfrecords",
"gs://jeremylorino-staging-bq-data/chess/lczero_data/train.tfrecords"
]
dataset = tf.data.TFRecordDataset(filenames)
dataset = dataset.map(map_fn, num_parallel_calls=10)
dataset = dataset.prefetch(4)
train_iterator = dataset.make_one_shot_iterator()
# shuffle_size = int(shuffle_size*(1.0-train_ratio))
filenames = [
"gs://jeremylorino-staging-bq-data/chess/lczero_data/test.tfrecords",
"gs://jeremylorino-staging-bq-data/chess/lczero_data/test.tfrecords"
]
dataset = tf.data.TFRecordDataset(filenames)
dataset = dataset.map(
map_fn,
num_parallel_calls=10) # .take(int(num_chunks*(1.0-train_ratio)))
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(
'gs://jeremylorino-staging-bq-data/chess/lczero_model/64x6-test-6/')
tfprocess.restore(cp)
# 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 = num_test * 10 // 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:
tfprocess.save_leelaz_weights(cmd.output)
tfprocess.session.close()
