def gen_golden_chunk(files, state):
buffer = example_buffer.ExampleBuffer(sampling_frac=1.0)
buffer.parallel_fill(files[1], threads=1)
buffer.flush(
os.path.join(
fsdb.golden_chunk_dir(),
state.output_model_name + '-{}.tfrecord.zz'.format(files[0])))
def selfplay(state):
output_dir = os.path.join(fsdb.selfplay_dir(), state.output_model_name)
holdout_dir = os.path.join(fsdb.holdout_dir(), state.output_model_name)
model_path = os.path.join(fsdb.models_dir(), state.best_model_name)
result = checked_run([
'bazel-bin/cc/selfplay', '--parallel_games=2048', '--num_readouts=100',
'--model={}.pb'.format(model_path),
'--output_dir={}'.format(output_dir),
'--holdout_dir={}'.format(holdout_dir)
] + cc_flags(state), 'selfplay')
logging.info(get_lines(result, make_slice[-2:]))
# Write examples to a single record.
pattern = os.path.join(output_dir, '*', '*.zz')
random.seed(state.seed)
tf.set_random_seed(state.seed)
np.random.seed(state.seed)
# TODO(tommadams): This method of generating one golden chunk per generation
# is sub-optimal because each chunk gets reused multiple times for training,
# introducing bias. Instead, a fresh dataset should be uniformly sampled out
# of *all* games in the training window before the start of each training run.
buffer = example_buffer.ExampleBuffer(sampling_frac=1.0)
# TODO(tommadams): parallel_fill is currently non-deterministic. Make it not
# so.
logging.info('Writing golden chunk from "{}"'.format(pattern))
buffer.parallel_fill(tf.gfile.Glob(pattern))
buffer.flush(
os.path.join(fsdb.golden_chunk_dir(),
state.output_model_name + '.tfrecord.zz'))
def initialize_from_checkpoint(state, out_files_number):
"""Initialize the reinforcement learning loop from a checkpoint."""
# The checkpoint's work_dir should contain the most recently trained model.
model_paths = glob.glob(
os.path.join(FLAGS.checkpoint_dir, 'work_dir/model.ckpt-*.pb'))
if len(model_paths) != 1:
raise RuntimeError(
'Expected exactly one model in the checkpoint work_dir, '
'got [{}]'.format(', '.join(model_paths)))
start_model_path = model_paths[0]
golden_chunks_dir = os.path.join(FLAGS.checkpoint_dir, 'golden_chunks')
for basename in os.listdir(golden_chunks_dir):
path = os.path.join(golden_chunks_dir, basename)
out_path = os.path.join(fsdb.golden_chunk_dir(), basename)
buffer = example_buffer.ExampleBuffer(sampling_frac=1.0)
example_num = buffer.parallel_fill(tf.gfile.Glob(path),
FLAGS.physical_cores)
buffer.flush_new(out_path, example_num, out_files_number, 1)
# Copy the latest trained model into the models directory and use it on the
# first round of selfplay.
state.best_model_name = 'checkpoint'
best_model_path = os.path.join(fsdb.models_dir(), state.best_model_name)
dual_net.optimize_graph(start_model_path, best_model_path,
FLAGS.quantization,
fsdb.golden_chunk_dir() + '/*.zz*',
FLAGS.eval_min_max_every_epoch)
# Copy the training files.
work_dir = os.path.join(FLAGS.checkpoint_dir, 'work_dir')
for basename in os.listdir(work_dir):
path = os.path.join(work_dir, basename)
shutil.copy(path, fsdb.working_dir())
def initialize_from_checkpoint(state):
"""Initialize the reinforcement learning loop from a checkpoint."""
# The checkpoint's work_dir should contain the most recently trained model.
model_paths = glob.glob(
os.path.join(FLAGS.checkpoint_dir, 'work_dir/model.ckpt-*.pb'))
if len(model_paths) != 1:
raise RuntimeError(
'Expected exactly one model in the checkpoint work_dir, '
'got [{}]'.format(', '.join(model_paths)))
start_model_path = model_paths[0]
# Copy the latest trained model into the models directory and use it on the
# first round of selfplay.
state.best_model_name = 'checkpoint'
shutil.copy(start_model_path,
os.path.join(fsdb.models_dir(), state.best_model_name + '.pb'))
shutil.copy(
start_model_path + '.og',
os.path.join(fsdb.models_dir(), state.best_model_name + '.pb.og'))
# Copy the training chunks.
golden_chunks_dir = os.path.join(FLAGS.checkpoint_dir, 'golden_chunks')
for basename in os.listdir(golden_chunks_dir):
path = os.path.join(golden_chunks_dir, basename)
out_path = os.path.join(fsdb.golden_chunk_dir(), basename)
buffer = example_buffer.ExampleBuffer(sampling_frac=1.0)
buffer.parallel_fill(tf.gfile.Glob(path))
buffer.flush(out_path, FLAGS.num_gpus_train)
# Copy the training files.
work_dir = os.path.join(FLAGS.checkpoint_dir, 'work_dir')
for basename in os.listdir(work_dir):
path = os.path.join(work_dir, basename)
shutil.copy(path, fsdb.working_dir())
async def selfplay(state, flagfile='selfplay', seed_factor=0):
"""Run selfplay and write a training chunk to the fsdb golden_chunk_dir.
Args:
state: the RL loop State instance.
flagfile: the name of the flagfile to use for selfplay, either 'selfplay'
(the default) or 'boostrap'.
seed_factor: Factor to increase seed.
"""
output_dir = os.path.join(fsdb.selfplay_dir(), state.output_model_name)
holdout_dir = os.path.join(fsdb.holdout_dir(), state.output_model_name)
lines = await run(
'bazel-bin/cc/selfplay',
'--flagfile={}.flags'.format(os.path.join(FLAGS.flags_dir, flagfile)),
'--model={}'.format(get_ckpt_path(state.best_model_path)),
'--output_dir={}'.format(output_dir),
'--holdout_dir={}'.format(holdout_dir),
'--seed={}'.format(state.seed+100*seed_factor))
result = '\n'.join(lines[-6:])
logging.info(result)
result = '\n'.join(lines[-50:])
try:
stats = parse_win_stats_table(result, 1)[0]
num_games = stats.total_wins
logging.info('Black won %0.3f, white won %0.3f',
stats.black_wins.total / num_games,
stats.white_wins.total / num_games)
except AssertionError:
# Poplar logging might screw up lines extraction approach.
logging.error("No results to parse: \n %s" % lines[-50:])
if not MULTI_SP:
# Write examples to a single record.
pattern = os.path.join(output_dir, '*', '*.zz')
random.seed(state.seed)
tf.set_random_seed(state.seed)
np.random.seed(state.seed)
# TODO(tommadams): This method of generating one golden chunk per generation
# is sub-optimal because each chunk gets reused multiple times for training,
# introducing bias. Instead, a fresh dataset should be uniformly sampled out
# of *all* games in the training window before the start of each training run.
buffer = example_buffer.ExampleBuffer(sampling_frac=1.0)
# TODO(tommadams): parallel_fill is currently non-deterministic. Make it not
# so.
logging.info('Writing golden chunk from "{}"'.format(pattern))
buffer.parallel_fill(tf.gfile.Glob(pattern))
buffer.flush(os.path.join(fsdb.golden_chunk_dir(),
state.output_model_name + '.tfrecord.zz'))
def divide_record(state, pattern, num_out, rank):
if rank < 0:
rank_str = ''
else:
rank_str = '-mpirank-' + str(rank)
buffer = example_buffer.ExampleBuffer(sampling_frac=1.0)
buffer.parallel_fill(tf.gfile.Glob(pattern))
output = os.path.join(fsdb.golden_chunk_dir(),
state.output_model_name + rank_str + '.tfrecord.zz')
buffer.flush(output, num_out)
if rank >= 0:
##put files to exchange
output = output + '*'
put_files_exchange(state, rank, fileout=output)
return
async def selfplay_multi(state, num_ipus):
""" Start *num_ipu* selfplay processes """
output_dir = os.path.join(fsdb.selfplay_dir(), state.output_model_name)
holdout_dir = os.path.join(fsdb.holdout_dir(), state.output_model_name)
flagfile = 'selfplay'
all_tasks = []
loop = asyncio.get_event_loop()
for i in range(num_ipus):
all_tasks.append(loop.create_task(selfplay_sub(state, output_dir, holdout_dir, flagfile, i)))
all_lines = await asyncio.gather(*all_tasks, return_exceptions=True)
black_wins_total = white_wins_total = num_games = 0
for lines in all_lines:
if type(lines) == RuntimeError or type(lines) == OSError:
raise lines
result = '\n'.join(lines[-6:])
logging.info(result)
stats = parse_win_stats_table(result, 1)[0]
num_games += stats.total_wins
black_wins_total += stats.black_wins.total
white_wins_total += stats.white_wins.total
logging.info('Black won %0.3f, white won %0.3f',
black_wins_total / num_games,
white_wins_total / num_games)
# copy paste from selfplay to aggregate results
# potentially should be parallized to training?
# Write examples to a single record.
pattern = os.path.join(output_dir, '*', '*.zz')
random.seed(state.seed)
tf.set_random_seed(state.seed)
np.random.seed(state.seed)
# TODO(tommadams): This method of generating one golden chunk per generation
# is sub-optimal because each chunk gets reused multiple times for training,
# introducing bias. Instead, a fresh dataset should be uniformly sampled out
# of *all* games in the training window before the start of each training run.
buffer = example_buffer.ExampleBuffer(sampling_frac=1.0)
# TODO(tommadams): parallel_fill is currently non-deterministic. Make it not
# so.
logging.info('Writing golden chunk from "{}"'.format(pattern))
buffer.parallel_fill(tf.gfile.Glob(pattern))
buffer.flush(os.path.join(fsdb.golden_chunk_dir(),
state.output_model_name + '.tfrecord.zz'))
def main(unused_argv):
mpi_comm = MPI.COMM_WORLD
mpi_rank = mpi_comm.Get_rank()
mpi_size = mpi_comm.Get_size()
# avoid seed out of range
random.seed(FLAGS.seed % 1048576)
tf.set_random_seed(FLAGS.seed % 1048576)
np.random.seed(FLAGS.seed % 1048576)
pattern = os.path.join(FLAGS.read_path, '*.zz')
files = tf.gfile.Glob(pattern)
buffer = example_buffer.ExampleBuffer(sampling_frac=1.0)
example_num = buffer.parallel_fill(files, threads=FLAGS.physical_cores)
# make sure all nodes generate same number of examples
example_num = int(mpi_comm.allreduce(example_num, op=MPI.MIN))
buffer.flush_new(FLAGS.write_path + '_{}'.format(mpi_rank),
example_num,
FLAGS.out_files_number,
threads=1)
shutil.rmtree('/tmp/minigo/home', ignore_errors=True)
def selfplay(state):
play_output_name = state.play_output_name
play_output_dir = os.path.join(fsdb.selfplay_dir(), play_output_name)
play_holdout_dir = os.path.join(fsdb.holdout_dir(), play_output_name)
result = checked_run([
'external/minigo/cc/main', '--mode=selfplay', '--parallel_games=2048',
'--num_readouts=100', '--model={}'.format(
state.play_model_path), '--output_dir={}'.format(play_output_dir),
'--holdout_dir={}'.format(play_holdout_dir)
] + cc_flags(state), 'selfplay')
logging.info(get_lines(result, make_slice[-2:]))
# Write examples to a single record.
logging.info('Extracting examples')
random.seed(state.seed)
tensorflow.set_random_seed(state.seed)
numpy.random.seed(state.seed)
buffer = example_buffer.ExampleBuffer(sampling_frac=1.0)
buffer.parallel_fill(
tensorflow.gfile.Glob(os.path.join(play_output_dir, '*.zz')))
buffer.flush(
os.path.join(fsdb.golden_chunk_dir(), play_output_name + '.tfrecord.zz'))
