def rl_loop():
"""Run the reinforcement learning loop
This tries to create a realistic way to run the reinforcement learning with
all default parameters.
"""
if goparams.DUMMY_MODEL:
# monkeypatch the hyperparams so that we get a quickly executing network.
dual_net.get_default_hyperparams = lambda **kwargs: {
'k': 8, 'fc_width': 16, 'num_shared_layers': 1, 'l2_strength': 1e-4, 'momentum': 0.9}
dual_net.TRAIN_BATCH_SIZE = 16
dual_net.EXAMPLES_PER_GENERATION = 64
#monkeypatch the shuffle buffer size so we don't spin forever shuffling up positions.
preprocessing.SHUFFLE_BUFFER_SIZE = 1000
qmeas.stop_time('selfplay_wait')
print("Gathering game output...")
gather()
print("Training on gathered game data...")
_, model_name = get_latest_model()
new_model = train()
if goparams.EVALUATE_PUZZLES:
qmeas.start_time('puzzle')
new_model_path = os.path.join(MODELS_DIR, new_model)
sgf_files = [
'./benchmark_sgf/9x9_pro_YKSH.sgf',
'./benchmark_sgf/9x9_pro_IYMD.sgf',
'./benchmark_sgf/9x9_pro_YSIY.sgf',
'./benchmark_sgf/9x9_pro_IYHN.sgf',
]
result, total_pct = predict_games.report_for_puzzles(new_model_path, sgf_files, 2, tries_per_move=1)
print('accuracy = ', total_pct)
qmeas.record('puzzle_total', total_pct)
qmeas.record('puzzle_result', repr(result))
qmeas.record('puzzle_summary', {'results': repr(result), 'total_pct': total_pct, 'model': new_model})
qmeas._flush()
with open(os.path.join(BASE_DIR, new_model + '-puzzles.txt'), 'w') as f:
f.write(repr(result))
f.write('\n' + str(total_pct) + '\n')
qmeas.stop_time('puzzle')
if total_pct >= goparams.TERMINATION_ACCURACY:
print('Reaching termination accuracy; ', goparams.TERMINATION_ACCURACY)
with open('TERMINATE_FLAG', 'w') as f:
f.write(repr(result))
f.write('\n' + str(total_pct) + '\n')
if goparams.EVALUATE_MODELS:
if not evaluate(model_name, new_model):
bury_latest_model()
def rl_loop_eval():
"""Run the reinforcement learning loop
This tries to create a realistic way to run the reinforcement learning with
all default parameters.
"""
(_, new_model) = get_latest_model()
qmeas.start_time('puzzle')
new_model_path = os.path.join(MODELS_DIR, new_model)
sgf_files = [
'./benchmark_sgf/9x9_pro_YKSH.sgf',
'./benchmark_sgf/9x9_pro_IYMD.sgf',
'./benchmark_sgf/9x9_pro_YSIY.sgf',
'./benchmark_sgf/9x9_pro_IYHN.sgf',
]
result, total_pct = predict_games.report_for_puzzles_parallel(
new_model_path, sgf_files, 2, tries_per_move=1)
#result, total_pct = predict_games.report_for_puzzles(new_model_path, sgf_files, 2, tries_per_move=1)
print('accuracy = ', total_pct)
print('result = ', result)
mlperf_log.minigo_print(key=mlperf_log.EVAL_ACCURACY,
value={
"epoch": iteration,
"value": total_pct
})
mlperf_log.minigo_print(key=mlperf_log.EVAL_TARGET,
value=goparams.TERMINATION_ACCURACY)
qmeas.record('puzzle_total', total_pct)
qmeas.record('puzzle_result', repr(result))
qmeas.record('puzzle_summary', {
'results': repr(result),
'total_pct': total_pct,
'model': new_model
})
qmeas._flush()
with open(os.path.join(BASE_DIR, new_model + '-puzzles.txt'), 'w') as f:
f.write(repr(result))
f.write('\n' + str(total_pct) + '\n')
qmeas.stop_time('puzzle')
if total_pct >= goparams.TERMINATION_ACCURACY:
print('Reaching termination accuracy; ', goparams.TERMINATION_ACCURACY)
mlperf_log.minigo_print(key=mlperf_log.RUN_STOP,
value={"success": True})
with open('TERMINATE_FLAG', 'w') as f:
f.write(repr(result))
f.write('\n' + str(total_pct) + '\n')
qmeas.end()
def rl_loop():
"""Run the reinforcement learning loop
This is meant to be more of an integration test than a realistic way to run
the reinforcement learning.
"""
# monkeypatch the hyperparams so that we get a quickly executing network.
dual_net.get_default_hyperparams = lambda **kwargs: {
'k': 8,
'fc_width': 16,
'num_shared_layers': 1,
'l2_strength': 1e-4,
'momentum': 0.9
}
dual_net.TRAIN_BATCH_SIZE = 16
dual_net.EXAMPLES_PER_GENERATION = 64
#monkeypatch the shuffle buffer size so we don't spin forever shuffling up positions.
preprocessing.SHUFFLE_BUFFER_SIZE = 1000
# with tempfile.TemporaryDirectory() as base_dir:
base_dir = "/tmp/minigo"
with open('/tmp/foo', 'w') as fff:
working_dir = os.path.join(base_dir, 'models_in_training')
model_save_path = os.path.join(base_dir, 'models', '000000-bootstrap')
next_model_save_file = os.path.join(base_dir, 'models',
'000001-nextmodel')
selfplay_dir = os.path.join(base_dir, 'data', 'selfplay')
model_selfplay_dir = os.path.join(selfplay_dir, '000000-bootstrap')
gather_dir = os.path.join(base_dir, 'data', 'training_chunks')
holdout_dir = os.path.join(base_dir, 'data', 'holdout',
'000000-bootstrap')
sgf_dir = os.path.join(base_dir, 'sgf', '000000-bootstrap')
os.makedirs(os.path.join(base_dir, 'data'), exist_ok=True)
print("Creating random initial weights...")
main.bootstrap(working_dir, model_save_path)
for i in range(100):
qmeas.start_time('main-loop')
print("Playing some games...")
# Do two selfplay runs to test gather functionality
qmeas.start_time('main-loop-self-play')
for j in range(2):
main.selfplay(load_file=model_save_path,
output_dir=model_selfplay_dir,
output_sgf=sgf_dir,
holdout_pct=0,
readouts=10)
qmeas.stop_time('main-loop-self-play')
# Do one holdout run to test validation
qmeas.start_time('main-loop-self-play-holdout')
main.selfplay(load_file=model_save_path,
holdout_dir=holdout_dir,
output_dir=model_selfplay_dir,
output_sgf=sgf_dir,
holdout_pct=100,
readouts=10)
qmeas.stop_time('main-loop-self-play-holdout')
print("See sgf files here?")
sgf_listing = subprocess.check_output(
["ls", "-l", sgf_dir + "/full"])
print(sgf_listing.decode("utf-8"))
print("Gathering game output...")
qmeas.start_time('main-loop-gather')
main.gather(input_directory=selfplay_dir,
output_directory=gather_dir)
qmeas.stop_time('main-loop-gather')
print("Training on gathered game data...")
qmeas.start_time('main-loop-train')
main.train(working_dir,
gather_dir,
next_model_save_file,
generation_num=1)
qmeas.stop_time('main-loop-train')
print("Trying validate on 'holdout' game...")
qmeas.start_time('main-loop-validate')
main.validate(working_dir, holdout_dir)
qmeas.stop_time('main-loop-validate')
print("Verifying that new checkpoint is playable...")
main.selfplay(load_file=next_model_save_file,
holdout_dir=holdout_dir,
output_dir=model_selfplay_dir,
output_sgf=sgf_dir,
readouts=10)
qmeas.stop_time('main-loop')
qmeas._flush()
def rl_loop():
"""Run the reinforcement learning loop
This is meant to be more of an integration test than a realistic way to run
the reinforcement learning.
"""
# monkeypatch the hyperparams so that we get a quickly executing network.
dual_net.get_default_hyperparams = lambda **kwargs: {
'k': 8, 'fc_width': 16, 'num_shared_layers': 1, 'l2_strength': 1e-4, 'momentum': 0.9}
dual_net.TRAIN_BATCH_SIZE = 16
dual_net.EXAMPLES_PER_GENERATION = 64
#monkeypatch the shuffle buffer size so we don't spin forever shuffling up positions.
preprocessing.SHUFFLE_BUFFER_SIZE = 1000
# with tempfile.TemporaryDirectory() as base_dir:
base_dir = "/tmp/minigo"
with open('/tmp/foo', 'w') as fff:
working_dir = os.path.join(base_dir, 'models_in_training')
model_save_path = os.path.join(base_dir, 'models', '000000-bootstrap')
next_model_save_file = os.path.join(base_dir, 'models', '000001-nextmodel')
selfplay_dir = os.path.join(base_dir, 'data', 'selfplay')
model_selfplay_dir = os.path.join(selfplay_dir, '000000-bootstrap')
gather_dir = os.path.join(base_dir, 'data', 'training_chunks')
holdout_dir = os.path.join(
base_dir, 'data', 'holdout', '000000-bootstrap')
sgf_dir = os.path.join(base_dir, 'sgf', '000000-bootstrap')
os.makedirs(os.path.join(base_dir, 'data'), exist_ok=True)
print("Creating random initial weights...")
main.bootstrap(working_dir, model_save_path)
for i in range(100):
qmeas.start_time('main-loop')
print("Playing some games...")
# Do two selfplay runs to test gather functionality
qmeas.start_time('main-loop-self-play')
for j in range(2):
main.selfplay(
load_file=model_save_path,
output_dir=model_selfplay_dir,
output_sgf=sgf_dir,
holdout_pct=0,
readouts=10)
qmeas.stop_time('main-loop-self-play')
# Do one holdout run to test validation
qmeas.start_time('main-loop-self-play-holdout')
main.selfplay(
load_file=model_save_path,
holdout_dir=holdout_dir,
output_dir=model_selfplay_dir,
output_sgf=sgf_dir,
holdout_pct=100,
readouts=10)
qmeas.stop_time('main-loop-self-play-holdout')
print("See sgf files here?")
sgf_listing = subprocess.check_output(["ls", "-l", sgf_dir + "/full"])
print(sgf_listing.decode("utf-8"))
print("Gathering game output...")
qmeas.start_time('main-loop-gather')
main.gather(input_directory=selfplay_dir, output_directory=gather_dir)
qmeas.stop_time('main-loop-gather')
print("Training on gathered game data...")
qmeas.start_time('main-loop-train')
main.train(working_dir, gather_dir, next_model_save_file, generation_num=1)
qmeas.stop_time('main-loop-train')
print("Trying validate on 'holdout' game...")
qmeas.start_time('main-loop-validate')
main.validate(working_dir, holdout_dir)
qmeas.stop_time('main-loop-validate')
print("Verifying that new checkpoint is playable...")
main.selfplay(
load_file=next_model_save_file,
holdout_dir=holdout_dir,
output_dir=model_selfplay_dir,
output_sgf=sgf_dir,
readouts=10)
qmeas.stop_time('main-loop')
qmeas._flush()