def evaluate(prev_model,
cur_model,
readouts=200,
verbose=1,
resign_threshold=0.95):
''' returns True if cur model should be used in future games '''
prev_model_save_path = os.path.join(MODELS_DIR, prev_model)
cur_model_save_path = os.path.join(MODELS_DIR, cur_model)
game_output_dir = os.path.join(SELFPLAY_DIR, cur_model)
game_holdout_dir = os.path.join(HOLDOUT_DIR, cur_model)
sgf_dir = os.path.join(SGF_DIR, cur_model)
cur_win_pct = main.evaluate_evenly(prev_model_save_path,
cur_model_save_path,
game_output_dir,
readouts=readouts,
games=goparams.EVAL_GAMES_PER_SIDE)
print('Evalute Win Pct = ', cur_win_pct)
qmeas.record('evaluate_win_pct', cur_win_pct)
keep = False
if cur_win_pct >= goparams.EVAL_WIN_PCT_FOR_NEW_MODEL:
qmeas.record('evaluate_choice', 'new')
keep = True
else:
qmeas.record('evaluate_choice', 'old')
keep = False
qmeas.record('eval_summary', {
'win_pct': cur_win_pct,
'model': cur_model,
'keep': keep
})
# return keep
return False
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 evaluate(prev_model, cur_model, readouts=200, verbose=1, resign_threshold=0.95):
''' returns True if cur model should be used in future games '''
prev_model_save_path = os.path.join(MODELS_DIR, prev_model)
cur_model_save_path = os.path.join(MODELS_DIR, cur_model)
game_output_dir = os.path.join(SELFPLAY_DIR, cur_model)
game_holdout_dir = os.path.join(HOLDOUT_DIR, cur_model)
sgf_dir = os.path.join(SGF_DIR, cur_model)
cur_win_pct = main.evaluate_evenly(prev_model_save_path, cur_model_save_path, game_output_dir, readouts=readouts, games=goparams.EVAL_GAMES_PER_SIDE)
print('Evalute Win Pct = ', cur_win_pct)
qmeas.record('evaluate_win_pct', cur_win_pct)
keep = False
if cur_win_pct >= goparams.EVAL_WIN_PCT_FOR_NEW_MODEL:
qmeas.record('evaluate_choice', 'new')
keep = True
else:
qmeas.record('evaluate_choice', 'old')
keep = False
qmeas.record('eval_summary', {'win_pct': cur_win_pct, 'model': cur_model, 'keep': keep})
return keep
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()