def main(unused_argv):
"""Run the reinforcement learning loop."""
print('Wiping dir %s' % FLAGS.base_dir, flush=True)
shutil.rmtree(FLAGS.base_dir, ignore_errors=True)
dirs = [fsdb.models_dir(), fsdb.selfplay_dir(), fsdb.holdout_dir(),
fsdb.eval_dir(), fsdb.golden_chunk_dir(), fsdb.working_dir()]
for d in dirs:
ensure_dir_exists(d);
# Copy the flag files so there's no chance of them getting accidentally
# overwritten while the RL loop is running.
flags_dir = os.path.join(FLAGS.base_dir, 'flags')
shutil.copytree(FLAGS.flags_dir, flags_dir)
FLAGS.flags_dir = flags_dir
# Copy the target model to the models directory so we can find it easily.
for file_name in [
"target.pb", "target_raw.ckpt.data-00000-of-00001",
"target_raw.ckpt.index", "target_raw.ckpt.meta"]:
shutil.copy(FLAGS.target_path[:-len("target.pb")] + file_name,
os.path.join(fsdb.models_dir(), file_name))
logging.getLogger().addHandler(
logging.FileHandler(os.path.join(FLAGS.base_dir, 'rl_loop.log')))
formatter = logging.Formatter('[%(asctime)s] %(message)s',
'%Y-%m-%d %H:%M:%S')
for handler in logging.getLogger().handlers:
handler.setFormatter(formatter)
with logged_timer('Total time'):
try:
rl_loop()
finally:
asyncio.get_event_loop().close()
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())
def main(unused_argv):
"""Run the reinforcement learning loop."""
print('Wiping dir %s' % FLAGS.base_dir, flush=True)
shutil.rmtree(FLAGS.base_dir, ignore_errors=True)
utils.ensure_dir_exists(fsdb.models_dir())
utils.ensure_dir_exists(fsdb.selfplay_dir())
utils.ensure_dir_exists(fsdb.holdout_dir())
utils.ensure_dir_exists(fsdb.eval_dir())
utils.ensure_dir_exists(fsdb.golden_chunk_dir())
utils.ensure_dir_exists(fsdb.working_dir())
# Copy the flag files so there's no chance of them getting accidentally
# overwritten while the RL loop is running.
flags_dir = os.path.join(FLAGS.base_dir, 'flags')
shutil.copytree(FLAGS.flags_dir, flags_dir)
FLAGS.flags_dir = flags_dir
# Copy the target model to the models directory so we can find it easily.
shutil.copy('ml_perf/target.pb', fsdb.models_dir())
logging.getLogger().addHandler(
logging.FileHandler(os.path.join(FLAGS.base_dir, 'rl_loop.log')))
formatter = logging.Formatter('[%(asctime)s] %(message)s',
'%Y-%m-%d %H:%M:%S')
for handler in logging.getLogger().handlers:
handler.setFormatter(formatter)
with utils.logged_timer('Total time'):
try:
rl_loop()
finally:
asyncio.get_event_loop().close()
async def train(state, tf_records):
"""Run training and write a new model to the fsdb models_dir.
Args:
state: the RL loop State instance.
tf_records: a list of paths to TensorFlow records to train on.
"""
model_path = os.path.join(fsdb.models_dir(), state.train_model_name)
await run(
'python3', 'train.py',
'--gpu_device_list={}'.format(','.join(FLAGS.train_devices)),
'--flagfile={}'.format(os.path.join(FLAGS.flags_dir, 'train.flags')),
'--work_dir={}'.format(fsdb.working_dir()),
'--export_path={}'.format(model_path),
'--use_extra_features={}'.format(FLAGS.use_extra_features),
'--freeze=true',
*tf_records)
# Append the time elapsed from when the RL was started to when this model
# was trained.
elapsed = time.time() - state.start_time
timestamps_path = os.path.join(fsdb.models_dir(), 'train_times.txt')
with gfile.Open(timestamps_path, 'a') as f:
print('{:.3f} {}'.format(elapsed, state.train_model_name), file=f)
def main(unused_argv):
"""Run the reinforcement learning loop."""
print('Wiping dir %s' % FLAGS.base_dir, flush=True)
shutil.rmtree(FLAGS.base_dir, ignore_errors=True)
utils.ensure_dir_exists(fsdb.models_dir())
utils.ensure_dir_exists(fsdb.selfplay_dir())
utils.ensure_dir_exists(fsdb.holdout_dir())
utils.ensure_dir_exists(fsdb.eval_dir())
utils.ensure_dir_exists(fsdb.golden_chunk_dir())
utils.ensure_dir_exists(fsdb.working_dir())
# Copy the target model to the models directory so we can find it easily.
shutil.copy('ml_perf/target.pb', fsdb.models_dir())
logging.getLogger().addHandler(
logging.FileHandler(os.path.join(FLAGS.base_dir, 'reinforcement.log')))
formatter = logging.Formatter('[%(asctime)s] %(message)s',
'%Y-%m-%d %H:%M:%S')
for handler in logging.getLogger().handlers:
handler.setFormatter(formatter)
with utils.logged_timer('Total time'):
for target_win_rate in rl_loop():
if target_win_rate > 0.5:
return logging.info('Passed exit criteria.')
logging.info('Failed to converge.')
async def train(state, selfplay_processes):
"""Run training and write a new model to the fsdb models_dir.
Args:
state: the RL loop State instance.
tf_records: a list of paths to TensorFlow records to train on.
"""
wait_for_training_examples(state, selfplay_processes, FLAGS.min_games_per_iteration)
tf_records = await sample_training_examples(state)
model_path = os.path.join(fsdb.models_dir(), state.train_model_name)
await run(
'python3', 'train.py',
'--gpu_device_list={}'.format(','.join(FLAGS.train_devices)),
'--flagfile={}'.format(os.path.join(FLAGS.flags_dir, 'train.flags')),
'--work_dir={}'.format(fsdb.working_dir()),
'--export_path={}'.format(model_path),
'--use_extra_features={}'.format(FLAGS.use_extra_features),
'--freeze=true',
*tf_records)
# Append the time elapsed from when the RL was started to when this model
# was trained.
elapsed = time.time() - state.start_time
timestamps_path = os.path.join(fsdb.models_dir(), 'train_times.txt')
with gfile.Open(timestamps_path, 'a') as f:
print('{:.3f} {}'.format(elapsed, state.train_model_name), file=f)
if FLAGS.validate and state.iter_num > 1:
try:
await validate(state)
except Exception as e:
logging.error(e)
def load_train_times():
models = []
path = os.path.join(fsdb.models_dir(), 'train_times.txt')
with gfile.Open(path, 'r') as f:
for line in f.readlines():
line = line.strip()
if line:
timestamp, name = line.split(' ')
path = 'tf,' + os.path.join(fsdb.models_dir(), name + '.pb')
models.append((float(timestamp), name, path))
return models
def post_train(state):
model_path = os.path.join(fsdb.models_dir(), state.train_model_name)
dual_net.optimize_graph(model_path + '.pb', model_path, FLAGS.quantization, fsdb.golden_chunk_dir()+'/*.zz', FLAGS.eval_min_max_every_epoch)
mll.save_model(state.iter_num-1)
# Append the time elapsed from when the RL was started to when this model
# was trained.
elapsed = time.time() - state.start_time
timestamps_path = os.path.join(fsdb.models_dir(), 'train_times.txt')
with gfile.Open(timestamps_path, 'a') as f:
print('{:.3f} {}'.format(elapsed, state.train_model_name), file=f)
def main(unused_argv):
"""Run the reinforcement learning loop."""
logging.getLogger('mlperf_compliance').propagate = False
##-->multi-node setup
if FLAGS.use_multinode:
mpi_comm = MPI.COMM_WORLD
mpi_rank = mpi_comm.Get_rank()
mpi_size = mpi_comm.Get_size()
print('[MPI Init] MPI rank {}, mpi size is {} host is {}'.format(
mpi_rank, mpi_size, socket.gethostname()))
else:
mpi_comm = None
mpi_rank = 0
mpi_size = 1
print('Wiping dir %s' % FLAGS.base_dir, flush=True)
shutil.rmtree(FLAGS.base_dir, ignore_errors=True)
dirs = [
fsdb.models_dir(),
fsdb.selfplay_dir(),
fsdb.holdout_dir(),
fsdb.eval_dir(),
fsdb.golden_chunk_dir(),
fsdb.working_dir()
]
##-->sharedFS for dataExchange. tmp solution 5/6/2019
if FLAGS.use_multinode:
ensure_dir_exists(FLAGS.shared_dir_exchange)
for d in dirs:
ensure_dir_exists(d)
# Copy the flag files so there's no chance of them getting accidentally
# overwritten while the RL loop is running.
flags_dir = os.path.join(FLAGS.base_dir, 'flags')
shutil.copytree(FLAGS.flags_dir, flags_dir)
FLAGS.flags_dir = flags_dir
# Copy the target model to the models directory so we can find it easily.
shutil.copy(FLAGS.target_path, os.path.join(fsdb.models_dir(),
'target.pb'))
shutil.copy(FLAGS.target_path + '.og',
os.path.join(fsdb.models_dir(), 'target.pb.og'))
with logged_timer('Total time from mpi_rank={}'.format(mpi_rank)):
try:
rl_loop(mpi_comm, mpi_rank, mpi_size)
finally:
asyncio.get_event_loop().close()
def get_files_exchange(state, mpi_rank):
##-->Train gets selfplay
##-->Self-play gets eval-model
if mpi_rank == FLAGS.train_rank:
selfplay_files = glob.glob(
os.path.join(FLAGS.shared_dir_exchange,
state.output_model_name + '-mpirank-*.zz*'))
for filename in selfplay_files:
print('Rank = {}, Getting file={} iter={} from SharedFS'.format(
mpi_rank, filename, state.iter_num))
shutil.copy(filename, fsdb.golden_chunk_dir())
else:
##self-play needs to get training eval model
dst_dir = os.path.join(fsdb.models_dir())
src_file = os.path.join(FLAGS.shared_dir_exchange,
state.train_model_name + '.pb')
print('Rank = {}, Getting file={} iter={} from SharedFS'.format(
mpi_rank, src_file, state.iter_num))
shutil.copy(src_file, dst_dir)
src_file = os.path.join(FLAGS.shared_dir_exchange,
state.train_model_name + '.pb' + '.og')
print('Rank = {}, Getting file={} iter={} from SharedFS'.format(
mpi_rank, src_file, state.iter_num))
shutil.copy(src_file, dst_dir)
async def convert(state):
"""Freeze the trained model and convert to TRT.
Args:
state: the RL loop State instance.
"""
# set to use only second from last GPU
new_env = os.environ.copy()
new_env['CUDA_VISIBLE_DEVICES'] = str(FLAGS.num_gpus_train - 2)
model_path = os.path.join(fsdb.models_dir(), state.train_model_name)
if FLAGS.use_multinode:
##convert is running in parallel with eval, and val
cores_per_worker = (FLAGS.cores_per_socket *
FLAGS.num_socket) // FLAGS.num_gpus_selfplay
start = (FLAGS.num_gpus_train - 2) * cores_per_worker
end = start + cores_per_worker - 1
##cpu-str
cpus = str(start) + '-' + str(end)
await run(new_env, 'taskset', '-c', cpus, 'python3', 'freeze_graph.py',
'--model_path={}'.format(model_path),
'--trt_batch={}'.format(FLAGS.trt_batch))
else:
await run(new_env, 'python3', 'freeze_graph.py',
'--model_path={}'.format(model_path),
'--trt_batch={}'.format(FLAGS.trt_batch))
async def start_selfplay():
output_dir = os.path.join(fsdb.selfplay_dir(), "$MODEL")
holdout_dir = os.path.join(fsdb.holdout_dir(), "$MODEL")
model_pattern = os.path.join(fsdb.models_dir(), '%d.pb')
logs = []
processes = []
loop = asyncio.get_event_loop()
for i, device in enumerate(FLAGS.selfplay_devices):
cmd = [
'bazel-bin/cc/concurrent_selfplay',
'--flagfile={}'.format(os.path.join(FLAGS.flags_dir,
'selfplay.flags')),
'--run_forever=1',
'--device={}'.format(device),
'--model={}'.format(model_pattern),
'--output_dir={}/{}'.format(output_dir, i),
'--holdout_dir={}/{}'.format(holdout_dir, i)]
cmd_str = await expand_cmd_str(cmd)
f = open(os.path.join(FLAGS.base_dir, 'selfplay_%d.log' % i), 'w')
f.write(cmd_str + '\n\n')
f.flush()
logging.info('Running: %s', cmd_str)
processes.append(await asyncio.create_subprocess_exec(
*cmd, stdout=f, stderr=asyncio.subprocess.STDOUT))
logs.append(f)
return (processes, logs)
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 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)
shutil.copy(path, fsdb.golden_chunk_dir())
# 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 eval_pv(eval_positions):
model_paths = oneoff_utils.get_model_paths(fsdb.models_dir())
idx_start = FLAGS.idx_start
eval_every = FLAGS.eval_every
print("Evaluating models {}-{}, eval_every={}".format(
idx_start, len(model_paths), eval_every))
for idx in tqdm(range(idx_start, len(model_paths), eval_every)):
if idx == idx_start:
player = oneoff_utils.load_player(model_paths[idx])
else:
oneoff_utils.restore_params(model_paths[idx], player)
mcts = strategies.MCTSPlayer(player.network, resign_threshold=-1)
for name, position in eval_positions:
mcts.initialize_game(position)
mcts.suggest_move(position)
path = []
node = mcts.root
while node.children:
node = node.children.get(node.best_child())
path.append("{},{}".format(node.fmove, int(node.N)))
save_file = os.path.join(FLAGS.data_dir,
"pv-{}-{}".format(name, idx))
with open(save_file, "w") as data:
data.write("{}, {}\n".format(idx, ",".join(path)))
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 same_run_eval(black_num=0, white_num=0):
"""Shorthand to spawn a job matching up two models from the same run,
identified by their model number """
if black_num <= 0 or white_num <= 0:
print("Need real model numbers")
return
b = fsdb.get_model(black_num)
w = fsdb.get_model(white_num)
b_model_path = os.path.join(fsdb.models_dir(), b)
w_model_path = os.path.join(fsdb.models_dir(), w)
return launch_eval_job(b_model_path + ".pb", w_model_path + ".pb",
"{:d}-{:d}".format(black_num, white_num),
flags.FLAGS.bucket_name)
def main(unused_argv):
logging.getLogger('mlperf_compliance').propagate = False
sgf_dir = os.path.join(fsdb.eval_dir(), 'target')
target = 'tf,' + os.path.join(fsdb.models_dir(), 'target.pb.og')
models = load_train_times()
timestamp_to_log = 0
iter_evaluated = 0
for i, (timestamp, name, path) in enumerate(models):
minigo_print(key=constants.EVAL_START, metadata={'epoch_num': i + 1})
iter_evaluated += 1
winrate = wait(evaluate_model(path + '.og', target, sgf_dir, i + 1))
minigo_print(key=constants.EVAL_ACCURACY,
value=winrate,
metadata={'epoch_num': i + 1})
minigo_print(key=constants.EVAL_STOP, metadata={'epoch_num': i + 1})
if winrate >= 0.50:
timestamp_to_log = timestamp
print('Model {} beat target after {}s'.format(name, timestamp))
break
minigo_print(key='eval_result',
metadata={
'iteration': iter_evaluated,
'timestamp': timestamp_to_log
})
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 best_model_path(self):
if self.best_model_name is None:
# We don't have a good model yet, use a random fake model implementation.
return 'random:0,0.4:0.4'
else:
return '{},{}.pb'.format(
FLAGS.engine, os.path.join(fsdb.models_dir(), self.best_model_name))
def bootstrap(unused_argv):
bootstrap_name = shipname.generate(0)
bootstrap_model_path = os.path.join(fsdb.models_dir(), bootstrap_name)
mask_flags.checked_run([
'python', 'bootstrap.py',
'--export_path={}'.format(bootstrap_model_path),
'--flagfile=rl_loop/distributed_flags'
])
def main(unused_argv):
sgf_dir = os.path.join(fsdb.eval_dir(), 'target')
target = 'tf,' + os.path.join(fsdb.models_dir(), 'target.pb')
models = load_train_times()
for i, (timestamp, name, path) in enumerate(models):
winrate = wait(evaluate_model(path, name, target, sgf_dir))
if winrate >= 0.50:
break
def train(state, tf_records):
model_path = os.path.join(fsdb.models_dir(), state.train_model_name)
checked_run(
'training', 'python3', 'train.py', *tf_records,
'--flagfile={}'.format(os.path.join(FLAGS.flags_dir, 'train.flags')),
'--work_dir={}'.format(fsdb.working_dir()),
'--export_path={}'.format(model_path),
'--training_seed={}'.format(state.seed), '--freeze=true')
def get_mg_path(model_run, model_num):
"""
model_run = integer, e.g. 15, 16, corresponding to the v-number
model_num = integer, e.g 939, for the model number in that run
"""
fsdb.switch_base("minigo-pub/v{:d}-19x19".format(model_run))
model = fsdb.get_model(model_num)
return os.path.join(fsdb.models_dir(), model)
def main(unused_argv):
sgf_dir = os.path.join(fsdb.eval_dir(), 'target')
target = 'tf,' + os.path.join(fsdb.models_dir(), 'target.pb')
models = load_train_times()
for i, (timestamp, name, path) in enumerate(models):
winrate = wait(evaluate_model(path, target, sgf_dir, i + 1))
if winrate >= 0.50:
print('Model {} beat target after {}s'.format(name, timestamp))
break
def evaluate(state):
eval_model = state.train_model_name
best_model = state.best_model_name
eval_model_path = os.path.join(fsdb.models_dir(), eval_model)
best_model_path = os.path.join(fsdb.models_dir(), best_model)
sgf_dir = os.path.join(fsdb.eval_dir(), eval_model)
result = checked_run(
'evaluation', 'bazel-bin/cc/eval',
'--flagfile={}'.format(os.path.join(FLAGS.flags_dir, 'eval.flags')),
'--model={}.pb'.format(eval_model_path),
'--model_two={}.pb'.format(best_model_path),
'--sgf_dir={}'.format(sgf_dir), '--seed={}'.format(state.seed))
result = get_lines(result, make_slice[-7:])
logging.info(result)
pattern = '{}\s+\d+\s+(\d+\.\d+)%'.format(eval_model)
win_rate = float(re.search(pattern, result).group(1)) * 0.01
logging.info('Win rate %s vs %s: %.3f', eval_model, best_model, win_rate)
return win_rate
def evaluate(state, against_model):
eval_model = state.train_model_name
eval_model_path = os.path.join(fsdb.models_dir(), eval_model)
against_model_path = os.path.join(fsdb.models_dir(), against_model)
sgf_dir = os.path.join(fsdb.eval_dir(), eval_model)
result = checked_run([
'bazel-bin/cc/eval', '--num_readouts=100', '--parallel_games=100',
'--model={}.pb'.format(eval_model_path),
'--model_two={}.pb'.format(against_model_path),
'--sgf_dir={}'.format(sgf_dir)
] + cc_flags(state), 'evaluation against ' + against_model)
result = get_lines(result, make_slice[-7:])
logging.info(result)
pattern = '{}\s+\d+\s+(\d+\.\d+)%'.format(eval_model)
win_rate = float(re.search(pattern, result).group(1)) * 0.01
logging.info('Win rate %s vs %s: %.3f', eval_model, against_model,
win_rate)
return win_rate
def main(unusedargv):
sgf_files = oneoff_utils.find_and_filter_sgf_files(FLAGS.sgf_dir,
FLAGS.min_year,
FLAGS.komi)
pos_data, move_data, result_data, move_idxs = sample_positions_from_games(
sgf_files=sgf_files, num_positions=FLAGS.num_positions)
df = get_training_curve_data(fsdb.models_dir(), pos_data, move_data,
result_data, FLAGS.idx_start,
FLAGS.eval_every)
save_plots(FLAGS.plot_dir, df)
def main(unusedargv):
model_paths = oneoff_utils.get_model_paths(fsdb.models_dir())
# List vars constructed when using dual_net.
dual_net_list(model_paths[0])
# Calculate l2 cost over a sequence of our models.
df = get_l2_cost_data(model_paths, FLAGS.idx_start, FLAGS.eval_every)
print(df)
save_plots(FLAGS.plot_dir, df)
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'))
print(os.path.join(FLAGS.checkpoint_dir, 'work_dir/model.ckpt-*.pb'))
print(os.getcwd())
if len(model_paths) != 1:
raise RuntimeError(
'Expected exactly one model in the checkpoint work_dir'
'({}), got [{}]'.format(
os.path.join(FLAGS.checkpoint_dir, 'work_dir'), ', '.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'))
start_model_files = glob.glob(os.path.join(
FLAGS.checkpoint_dir, 'work_dir/model.ckpt-9383_raw.ckpt*'))
for file_name in start_model_files:
shutil.copy(file_name,
os.path.join(fsdb.models_dir(),
state.best_model_name +
os.path.basename(file_name)[len("model.ckpt-9383"):]))
# 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)
shutil.copy(path, fsdb.golden_chunk_dir())
# 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 train(state, tf_records):
"""Run training and write a new model to the fsdb models_dir.
Args:
state: the RL loop State instance.
tf_records: a list of paths to TensorFlow records to train on.
"""
new_env = os.environ.copy()
model_path = os.path.join(fsdb.models_dir(), state.train_model_name)
if FLAGS.use_mgpu_horovod:
# assign leading cores of sockets to train
await run(
new_env, 'mpiexec', '--allow-run-as-root', '--map-by',
'ppr:{}:socket,pe=2'.format(
str(FLAGS.num_gpus_train // FLAGS.num_socket)), '-np',
str(FLAGS.num_gpus_train), 'python3', 'train.py', *tf_records,
'--flagfile={}'.format(os.path.join(FLAGS.flags_dir,
'train.flags')),
'--work_dir={}'.format(fsdb.working_dir()),
'--export_path={}'.format(model_path), '--training_seed={}'.format(
state.seed), '--use_mgpu_horovod=true', '--freeze=true')
else:
new_env['CUDA_VISIBLE_DEVICES'] = '0'
await run(
new_env, 'python3', 'train.py', *tf_records,
'--flagfile={}'.format(os.path.join(FLAGS.flags_dir,
'train.flags')),
'--work_dir={}'.format(fsdb.working_dir()),
'--export_path={}'.format(model_path),
'--training_seed={}'.format(state.seed), '--freeze=true')
minigo_print(key='save_model', value={'iteration': state.iter_num})
# Append the time elapsed from when the RL was started to when this model
# was trained.
elapsed = time.time() - state.start_time
timestamps_path = os.path.join(fsdb.models_dir(), 'train_times.txt')
with gfile.Open(timestamps_path, 'a') as f:
print('{:.3f} {}'.format(elapsed, state.train_model_name), file=f)
