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']
experimental_parser = cfg['dataset'].get('experimental_v5_only_dataset',
False)
# num_train = int(num_chunks * train_ratio)
# we just need to use one data loader, just put everything into train
num_train = int(num_chunks)
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']
shuffle_size = 1
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
# Load data with split batch size, which will be combined to the total batch size in tfprocess.
ChunkParser.BATCH_SIZE = split_batch_size
value_focus_min = cfg['training'].get('value_focus_min', 1)
value_focus_slope = cfg['training'].get('value_focus_slope', 0)
root_dir = os.path.join(cfg['training']['path'], cfg['name'])
if not os.path.exists(root_dir):
os.makedirs(root_dir)
tfprocess = TFProcess(cfg)
experimental_reads = max(2, mp.cpu_count() - 2) // 2
extractor = select_extractor(tfprocess.INPUT_MODE)
if experimental_parser and (value_focus_min != 1
or value_focus_slope != 0):
raise ValueError(
'Experimental parser does not support non-default value \
focus parameters.')
def read(x):
return tf.data.FixedLengthRecordDataset(
x,
8308,
compression_type='GZIP',
num_parallel_reads=experimental_reads)
if experimental_parser:
# train_dataset = tf.data.Dataset.from_tensor_slices(train_chunks).shuffle(len(train_chunks)).repeat().batch(256)\
train_dataset = tf.data.Dataset.from_tensor_slices(train_chunks).repeat().batch(256)\
.interleave(read, num_parallel_calls=1)\
.batch(SKIP_MULTIPLE*SKIP).map(semi_sample).unbatch()\
.batch(split_batch_size).map(extractor)
# .shuffle(shuffle_size)\
# .batch(split_batch_size).map(extractor)
else:
train_parser = ChunkParser(train_chunks,
tfprocess.INPUT_MODE,
shuffle_size=shuffle_size,
sample=SKIP,
batch_size=ChunkParser.BATCH_SIZE,
value_focus_min=value_focus_min,
value_focus_slope=value_focus_slope,
workers=train_workers)
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(ChunkParser.parse_function)
shuffle_size = int(shuffle_size * (1.0 - train_ratio))
if experimental_parser:
# test_dataset = tf.data.Dataset.from_tensor_slices(test_chunks).shuffle(len(test_chunks)).repeat().batch(256)\
test_dataset = tf.data.Dataset.from_tensor_slices(test_chunks).repeat().batch(256)\
.interleave(read, num_parallel_calls=2)\
.batch(SKIP_MULTIPLE*SKIP).map(semi_sample).unbatch()\
.batch(split_batch_size).map(extractor)
# .shuffle(shuffle_size)\
# .batch(split_batch_size).map(extractor)
else:
# no value focus for test_parser
test_parser = ChunkParser(test_chunks,
tfprocess.INPUT_MODE,
shuffle_size=shuffle_size,
sample=SKIP,
batch_size=ChunkParser.BATCH_SIZE,
workers=test_workers)
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(ChunkParser.parse_function)
validation_dataset = None
if 'input_validation' in cfg['dataset']:
valid_chunks = get_all_chunks(cfg['dataset']['input_validation'])
validation_dataset = tf.data.FixedLengthRecordDataset(valid_chunks, 8308, compression_type='GZIP', num_parallel_reads=experimental_reads)\
.batch(split_batch_size, drop_remainder=True).map(extractor)
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)
##########################
# Custom Additions #
##########################
tfprocess.init_v2(train_dataset, test_dataset, validation_dataset)
# load net from weights file given in yaml config
tfprocess.replace_weights_v2(proto_filename=cmd.net, ignore_errors=False)
tfprocess.model.summary()
for layer_name, path in zip(cmd.layer, cmd.path):
# sort data files
train_chunks = sorted(train_chunks)
# create predictor that gives access to specific intermediate layer
layer = tfprocess.model.get_layer(layer_name)
earlyPredictor = tf.keras.models.Model(
tfprocess.model.inputs,
[tfprocess.model.inputs, tfprocess.model.outputs, layer.output])
# create custom iterator which doesn't shuffle the data etc
custom_parse_gen = train_parser.custom_parse(train_chunks)
turn_counter = 0
custom_iter = iter(custom_parse_gen)
# prepare dataframe
df = pd.DataFrame()
# iterate entire dataset generator / iterator
for data in custom_iter: #i in range(30):
# data = next(custom_iter)
planes, probs, winner, best_q = train_parser.custom_get_batch(data)
x = planes
print('predicting...')
_, _, layer_results = earlyPredictor.predict(x)
# append to dataframe
# df = df.append(pd.DataFrame(activation_31.reshape(-1,128*8*8)))
shape_tuple = (-1, np.prod(layer.output_shape[1:]))
df = df.append(pd.DataFrame(layer_results.reshape(shape_tuple)))
turn_counter += len(x)
df.info()
df.to_csv(path)
print('done')
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']
experimental_parser = cfg['dataset'].get('experimental_v5_only_dataset',
False)
# num_train = int(num_chunks * train_ratio)
# we just need to use one data loader, just put everything into train
num_train = int(num_chunks)
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']
shuffle_size = 1
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
# Load data with split batch size, which will be combined to the total batch size in tfprocess.
ChunkParser.BATCH_SIZE = split_batch_size
value_focus_min = cfg['training'].get('value_focus_min', 1)
value_focus_slope = cfg['training'].get('value_focus_slope', 0)
root_dir = os.path.join(cfg['training']['path'], cfg['name'])
if not os.path.exists(root_dir):
os.makedirs(root_dir)
tfprocess = TFProcess(cfg)
experimental_reads = max(2, mp.cpu_count() - 2) // 2
extractor = select_extractor(tfprocess.INPUT_MODE)
if experimental_parser and (value_focus_min != 1
or value_focus_slope != 0):
raise ValueError(
'Experimental parser does not support non-default value \
focus parameters.')
def read(x):
return tf.data.FixedLengthRecordDataset(
x,
8308,
compression_type='GZIP',
num_parallel_reads=experimental_reads)
if experimental_parser:
# train_dataset = tf.data.Dataset.from_tensor_slices(train_chunks).shuffle(len(train_chunks)).repeat().batch(256)\
train_dataset = tf.data.Dataset.from_tensor_slices(train_chunks).repeat().batch(256)\
.interleave(read, num_parallel_calls=1)\
.batch(SKIP_MULTIPLE*SKIP).map(semi_sample).unbatch()\
.batch(split_batch_size).map(extractor)
# .shuffle(shuffle_size)\
# .batch(split_batch_size).map(extractor)
else:
train_parser = ChunkParser(train_chunks,
tfprocess.INPUT_MODE,
shuffle_size=shuffle_size,
sample=SKIP,
batch_size=ChunkParser.BATCH_SIZE,
value_focus_min=value_focus_min,
value_focus_slope=value_focus_slope,
workers=train_workers)
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(ChunkParser.parse_function)
shuffle_size = int(shuffle_size * (1.0 - train_ratio))
if experimental_parser:
# test_dataset = tf.data.Dataset.from_tensor_slices(test_chunks).shuffle(len(test_chunks)).repeat().batch(256)\
test_dataset = tf.data.Dataset.from_tensor_slices(test_chunks).repeat().batch(256)\
.interleave(read, num_parallel_calls=2)\
.batch(SKIP_MULTIPLE*SKIP).map(semi_sample).unbatch()\
.batch(split_batch_size).map(extractor)
# .shuffle(shuffle_size)\
# .batch(split_batch_size).map(extractor)
else:
# no value focus for test_parser
test_parser = ChunkParser(test_chunks,
tfprocess.INPUT_MODE,
shuffle_size=shuffle_size,
sample=SKIP,
batch_size=ChunkParser.BATCH_SIZE,
workers=test_workers)
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(ChunkParser.parse_function)
validation_dataset = None
if 'input_validation' in cfg['dataset']:
valid_chunks = get_all_chunks(cfg['dataset']['input_validation'])
validation_dataset = tf.data.FixedLengthRecordDataset(valid_chunks, 8308, compression_type='GZIP', num_parallel_reads=experimental_reads)\
.batch(split_batch_size, drop_remainder=True).map(extractor)
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_v2(train_dataset, test_dataset, validation_dataset)
# load net from weights file given in yaml config
tfprocess.replace_weights_v2(proto_filename=cmd.net, ignore_errors=False)
train_chunks = sorted(train_chunks)
custom_parse_gen = train_parser.custom_parse(train_chunks)
print(train_chunks)
counter = 0
custom_iter = iter(custom_parse_gen)
for data in custom_iter:#i in range(30):
# data = next(custom_iter)
planes, probs, winner, best_q = train_parser.custom_get_batch(data)
print(planes.shape)
x = planes
# TODO make sure no shuffling happens. the following output should clearly show the first few moves of the game w.r.t. pawn placement
for i in range(len(x)):
counter += 1
print('move no.:', counter)
print(x[i, 0].reshape(8,8))
print()
print()
# TODO
# print(tfprocess.model.summary())
# print(tfprocess.train_dataset)
# data_iter = iter(tfprocess.train_dataset)
# nxt = next(data_iter)
# x, _, _, _, _ = nxt
# x2_0_r = tf.reshape(x[0], [1, 112, 64])
# pred = tfprocess.model.predict(x)
earlyPredictor = tf.keras.models.Model(tfprocess.model.inputs, [tfprocess.model.inputs, tfprocess.model.outputs, tfprocess.model.get_layer('activation_31').output])
early_pred_single = earlyPredictor.predict(x)
# print(np.array(early_pred_single[0]).shape)
input = np.array(early_pred_single[0])
print(input.shape)
# print(early_pred_single[0]) # input
# print(early_pred_single[1]) # output
# print(early_pred_single[2]) # intermediate layer
# print(train_parser.sample_record())
# print(next(tfprocess.train_iter))
# tfprocess.restore_v2()
# 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.total_batch_size = total_batch_size
# tfprocess.process_loop_v2(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_v2(cmd.output)
# else:
# tfprocess.save_leelaz_weights_v2(cmd.output)
train_parser.shutdown()
test_parser.shutdown()