def get_option_spec(cls):
spec = PyOptionSpec()
spec.addFloatOption(
'discount',
'exponential discount rate',
0.99)
return spec
def get_option_spec(cls, model_class=None, model_idx=None):
spec = PyOptionSpec()
spec.addStrOption('load', 'load model', '')
spec.addStrListOption(
'onload',
('functions to call after loading. e.g., reset,zero_first_layer. '
'These functions are specified in the model'), [])
spec.addStrListOption('omit_keys', 'omitted keys when loading', [])
spec.addStrListOption('replace_prefix', 'replace prefix', [])
spec.addIntOption('gpu', 'which GPU to use', -1)
spec.addBoolOption(
'check_loaded_options',
'Toggles consistency check of loaded vs. current model options.',
True)
spec.addBoolOption('use_fp16', 'use_fp16', False)
spec.addFloatOption(
'load_model_sleep_interval',
('If zero, has no effect. If positive, then before loading the '
'model, we will sleep for an interval of '
'duration (secs) ~ Uniform[0, load_model_sleep_interval]'), 0.0)
if model_class is not None and hasattr(model_class, 'get_option_spec'):
spec.merge(model_class.get_option_spec())
idx_suffix = '' if model_idx is None else str(model_idx)
spec.addPrefixSuffixToOptionNames('', idx_suffix)
return spec
def get_option_spec(cls):
spec = PyOptionSpec()
spec.addBoolOption('bn', 'toggles batch norm', True)
spec.addBoolOption('leaky_relu', 'toggles leaky ReLU', False)
spec.addFloatOption('bn_momentum',
'batch norm momentum (pytorch style)', 0.1)
spec.addIntOption('dim', 'model dimension', 128)
return spec
def get_option_spec(cls):
spec = PyOptionSpec()
spec.addStrOption('preload_sgf', 'TODO: fill this help message in', '')
spec.addIntOption('preload_sgf_move_to',
'TODO: fill this help message in', -1)
spec.addStrOption('mode', 'TODO: fill this help message in', "online")
spec.addBoolOption('actor_only', 'TODO: fill this help message in',
False)
spec.addIntOption('num_reset_ranking',
'TODO: fill this help message in', 5000)
spec.addBoolOption('verbose', 'TODO: fill this help message in', False)
spec.addBoolOption('print_result', 'TODO: fill this help message in',
False)
spec.addIntOption('data_aug',
'specify data augumentation, 0-7, -1 mean random',
-1)
spec.addIntOption(
'num_games_per_thread',
('For offline mode, it is the number of concurrent games per '
'thread, used to increase diversity of games; for selfplay mode, '
'it is the number of games played at each thread, and after that '
'we need to call restartAllGames() to resume.'), -1)
spec.addIntOption('num_future_actions',
'TODO: fill this help message in', 1)
spec.addIntOption('move_cutoff', 'Cutoff ply in replay', -1)
spec.addStrOption('mode', 'TODO: fill this help message in', 'online')
spec.addBoolOption('black_use_policy_network_only',
'TODO: fill this help message in', False)
spec.addIntOption('ply_pass_enabled',
'TODO: fill this help message in', 0)
spec.addBoolOption('use_mcts', 'TODO: fill this help message in',
False)
spec.addBoolOption('use_df_feature', 'TODO: fill this help message in',
False)
spec.addStrOption('dump_record_prefix',
'TODO: fill this help message in', '')
spec.addFloatOption('resign_thres', 'TODO: fill this help message in',
0.0)
spec.addBoolOption('following_pass', 'TODO: fill this help message in',
False)
spec.addIntOption('gpu', 'TODO: fill this help message in', -1)
spec.addBoolOption('parameter_print',
'TODO: fill this help message in', True)
spec.addIntOption('batchsize', 'batch size', 128)
spec.addIntOption('batchsize2', 'batch size', -1)
spec.addFloatOption('eval_winrate_thres',
'Win rate threshold for evalution', 0.55)
spec.addIntOption(
'suicide_after_n_games',
'return after n games have finished, -1 means it never ends', -1)
spec.merge(PyOptionSpec.fromClasses((ContextArgs, MoreLabels)))
return spec
def get_option_spec(cls):
spec = PyOptionSpec()
spec.addFloatOption(
'grad_clip_norm',
'gradient norm clipping',
0.0)
spec.addStrOption(
'value_node',
'name of the value node',
'V')
return spec
def get_option_spec(cls):
spec = PyOptionSpec()
spec.addStrOption('value_node', 'name of the value node', 'V')
spec.addFloatOption('adv_clip',
'clip value of advantage. 0.0 means no clipping',
0.0)
spec.merge(
PyOptionSpec.fromClasses(
(PolicyGradient, DiscountedReward, ValueMatcher)))
return spec
def get_option_spec(cls):
spec = PyOptionSpec()
spec.addStrOption(
'sample_policy',
'choices of epsilon-greedy, multinomial, or uniform',
'epsilon-greedy')
spec.addBoolOption('store_greedy',
('if enabled, picks maximum-probability action; '
'otherwise, sample from distribution'), False)
spec.addFloatOption('epsilon', 'used in epsilon-greedy', 0.0)
spec.addStrListOption('sample_nodes', 'nodes to be sampled and saved',
['pi,a'])
return spec
def get_option_spec(cls):
spec = PyOptionSpec()
spec.addBoolOption(
'bn',
'toggles batch norm',
True)
spec.addBoolOption(
'leaky_relu',
'toggles leaky ReLU',
False)
spec.addFloatOption(
'bn_momentum',
'batch norm momentum (pytorch style)',
0.1)
spec.addIntOption(
'num_block',
'number of blocks',
20)
spec.addIntOption(
'dim',
'model dimension',
128)
spec.addBoolOption(
'use_data_parallel',
'TODO: fill this in',
False)
spec.addBoolOption(
'use_data_parallel_distributed',
'TODO: fill this in',
False)
spec.addIntOption(
'dist_rank',
'TODO: fill this in',
-1)
spec.addIntOption(
'dist_world_size',
'TODO: fill this in',
-1)
spec.addStrOption(
'dist_url',
'TODO: fill this in',
'')
spec.addIntOption(
'gpu',
'which gpu to use',
-1)
spec.merge(GoResNet.get_option_spec())
return spec
def get_option_spec(cls): spec = PyOptionSpec() spec.addBoolOption( 'bn', 'toggles batch norm', True) spec.addBoolOption( 'leaky_relu', 'toggles leaky ReLU', False) spec.addFloatOption( 'bn_momentum', 'batch norm momentum (pytorch style)', 0.1) spec.addIntOption( 'num_block', 'number of blocks', 20) spec.addIntOption( 'dim', 'model dimension', 128) spec.addBoolOption( 'use_data_parallel', 'TODO: fill this in', False) spec.addBoolOption( 'use_data_parallel_distributed', 'TODO: fill this in', False) spec.addIntOption( 'dist_rank', 'TODO: fill this in', -1) spec.addIntOption( 'dist_world_size', 'TODO: fill this in', -1) spec.addStrOption( 'dist_url', 'TODO: fill this in', '') spec.addIntOption( 'gpu', 'which gpu to use', -1) spec.merge(GoResNet.get_option_spec()) return spec
def get_option_spec(cls):
spec = PyOptionSpec()
spec.addStrOption('opt_method', 'optimization method (adam or sgd)',
'adam')
spec.addFloatOption('lr', 'learning rate', 1e-3)
spec.addFloatOption('adam_eps', 'Adam epsilon', 1e-3)
spec.addFloatOption('momentum', 'momentum parameter', 0.9)
spec.addFloatOption('weight_decay', 'weight decay rate', 0.0)
return spec
def get_option_spec(cls):
spec = PyOptionSpec()
spec.addBoolOption(
'bn',
'toggles batch norm',
True)
spec.addBoolOption(
'leaky_relu',
'toggles leaky ReLU',
False)
spec.addFloatOption(
'bn_momentum',
'batch norm momentum (pytorch style)',
0.1)
spec.addIntOption(
'dim',
'model dimension',
128)
return spec
def get_option_spec(cls):
spec = PyOptionSpec()
spec.addStrOption(
'sample_policy',
'choices of epsilon-greedy, multinomial, or uniform',
'epsilon-greedy')
spec.addBoolOption(
'store_greedy',
('if enabled, picks maximum-probability action; '
'otherwise, sample from distribution'),
False)
spec.addFloatOption(
'epsilon',
'used in epsilon-greedy',
0.0)
spec.addStrListOption(
'sample_nodes',
'nodes to be sampled and saved',
['pi,a'])
return spec
def get_option_spec(cls):
spec = PyOptionSpec()
spec.addFloatOption('entropy_ratio',
'the entropy ratio we put on PolicyGradient', 0.01)
spec.addFloatOption('grad_clip_norm', 'gradient norm clipping', 0.0)
spec.addFloatOption('min_prob', 'mininal probability used in training',
1e-6)
spec.addFloatOption('ratio_clamp', 'maximum importance sampling ratio',
10.0)
spec.addStrListOption('policy_action_nodes',
'the entropy ratio we put on PolicyGradient',
['pi,a'])
return spec
def get_option_spec(cls):
spec = PyOptionSpec()
spec.addStrOption(
'opt_method',
'optimization method (adam or sgd)',
'adam')
spec.addFloatOption(
'lr',
'learning rate',
1e-3)
spec.addFloatOption(
'adam_eps',
'Adam epsilon',
1e-3)
spec.addFloatOption(
'momentum',
'momentum parameter',
0.9)
spec.addFloatOption(
'weight_decay',
'weight decay rate',
0.0)
return spec
def get_option_spec(cls):
spec = PyOptionSpec()
spec.addFloatOption(
'entropy_ratio',
'the entropy ratio we put on PolicyGradient',
0.01)
spec.addFloatOption(
'grad_clip_norm',
'gradient norm clipping',
0.0)
spec.addFloatOption(
'min_prob',
'mininal probability used in training',
1e-6)
spec.addFloatOption(
'ratio_clamp',
'maximum importance sampling ratio',
10.0)
spec.addStrListOption(
'policy_action_nodes',
'the entropy ratio we put on PolicyGradient',
['pi,a'])
return spec
def get_option_spec(cls):
spec = PyOptionSpec()
spec.addIntOption('num_games', 'number of games', 1024)
spec.addIntOption('batchsize', 'batch size', 128)
spec.addIntOption('T', 'number of timesteps', 6)
spec.addIntOption('mcts_threads', 'number of MCTS threads', 0)
spec.addIntOption('mcts_rollout_per_batch',
'Batch size for mcts rollout', 1)
spec.addIntOption('mcts_rollout_per_thread',
'number of rollotus per MCTS thread', 1)
spec.addBoolOption('mcts_verbose', 'enables mcts verbosity', False)
spec.addBoolOption('mcts_verbose_time',
'enables mcts verbosity for time stats', False)
spec.addBoolOption('mcts_persistent_tree',
'use persistent tree in MCTS', False)
spec.addBoolOption('mcts_use_prior', 'use prior in MCTS', False)
spec.addIntOption('mcts_virtual_loss',
'"virtual" number of losses for MCTS edges', 0)
spec.addStrOption('mcts_pick_method',
'criterion for mcts node selection', 'most_visited')
spec.addFloatOption('mcts_puct', 'prior weight', 1.0)
spec.addFloatOption(
'mcts_epsilon',
'for exploration enhancement, weight of randomization', 0.0)
spec.addFloatOption(
'mcts_alpha',
'for exploration enhancement, alpha term in gamma distribution',
0.0)
spec.addBoolOption("mcts_unexplored_q_zero",
'set all unexplored node to have Q value zero',
False)
spec.addBoolOption(
"mcts_root_unexplored_q_zero",
'set unexplored child of root node to have Q value zero', False)
return spec
def get_option_spec(cls):
spec = PyOptionSpec()
spec.addIntOption(
'num_games',
'number of games',
1024)
spec.addIntOption(
'batchsize',
'batch size',
128)
spec.addIntOption(
'T',
'number of timesteps',
6)
spec.addBoolOption(
'verbose_comm',
'enables verbose comm',
False)
spec.addIntOption(
'mcts_threads',
'number of MCTS threads',
0)
spec.addIntOption(
'mcts_rollout_per_batch',
'Batch size for mcts rollout',
1)
spec.addIntOption(
'mcts_rollout_per_thread',
'number of rollotus per MCTS thread',
1)
spec.addBoolOption(
'mcts_verbose',
'enables mcts verbosity',
False)
spec.addBoolOption(
'mcts_verbose_time',
'enables mcts verbosity for time stats',
False)
spec.addBoolOption(
'mcts_persistent_tree',
'use persistent tree in MCTS',
False)
spec.addBoolOption(
'mcts_use_prior',
'use prior in MCTS',
False)
spec.addIntOption(
'mcts_virtual_loss',
'"virtual" number of losses for MCTS edges',
0)
spec.addStrOption(
'mcts_pick_method',
'criterion for mcts node selection',
'most_visited')
spec.addFloatOption(
'mcts_puct',
'prior weight',
1.0)
spec.addFloatOption(
'mcts_epsilon',
'for exploration enhancement, weight of randomization',
0.0)
spec.addFloatOption(
'mcts_alpha',
'for exploration enhancement, alpha term in gamma distribution',
0.0)
spec.addBoolOption(
"mcts_unexplored_q_zero",
'set all unexplored node to have Q value zero',
False)
spec.addBoolOption(
"mcts_root_unexplored_q_zero",
'set unexplored child of root node to have Q value zero',
False)
return spec
def get_option_spec(cls):
spec = PyOptionSpec()
spec.addFloatOption('entropy_ratio', '', 0.01)
spec.addFloatOption('grad_clip_norm', '', 0.0)
spec.addFloatOption('discount', '', 0.99)
spec.addFloatOption('min_prob', '', 1e-6)
spec.addFloatOption('adv_clip', '', 0.0)
spec.addFloatOption('ratio_clamp', '', 10.0)
spec.addFloatOption('ppo_eps', '', 0.2)
spec.addBoolOption("use_a2c", "", False)
return spec
def get_option_spec(cls):
spec = PyOptionSpec()
spec.addFloatOption('discount', 'exponential discount rate', 0.99)
return spec
def get_option_spec(cls):
spec = PyOptionSpec()
spec.addFloatOption(
'entropy_ratio',
'',
0.01)
spec.addFloatOption(
'grad_clip_norm',
'',
0.0)
spec.addFloatOption(
'discount',
'',
0.99
)
spec.addFloatOption(
'min_prob',
'',
1e-6)
spec.addFloatOption(
'adv_clip',
'',
0.0)
spec.addFloatOption(
'ratio_clamp',
'',
10.0)
return spec
def get_option_spec(cls):
spec = PyOptionSpec()
spec.addFloatOption('grad_clip_norm', 'gradient norm clipping', 0.0)
spec.addStrOption('value_node', 'name of the value node', 'V')
return spec
def get_option_spec(cls):
spec = PyOptionSpec()
spec.addStrOption(
'preload_sgf',
'TODO: fill this help message in',
'')
spec.addIntOption(
'preload_sgf_move_to',
'TODO: fill this help message in',
-1)
spec.addBoolOption(
'actor_only',
'TODO: fill this help message in',
False)
spec.addStrListOption(
'list_files',
'Provide a list of json files for offline training',
[])
spec.addIntOption(
'port',
'TODO: fill this help message in',
5556)
spec.addStrOption(
'server_addr',
'TODO: fill this help message in',
'')
spec.addStrOption(
'server_id',
'TODO: fill this help message in',
'')
spec.addIntOption(
'q_min_size',
'TODO: fill this help message in',
10)
spec.addIntOption(
'q_max_size',
'TODO: fill this help message in',
1000)
spec.addIntOption(
'num_reader',
'TODO: fill this help message in',
50)
spec.addIntOption(
'num_reset_ranking',
'TODO: fill this help message in',
5000)
spec.addIntOption(
'client_max_delay_sec',
'Maximum amount of allowed delays in sec. If the client '
'didn\'t respond after that, we think it is dead.',
1200)
spec.addBoolOption(
'verbose',
'TODO: fill this help message in',
False)
spec.addBoolOption(
'keep_prev_selfplay',
'TODO: fill this help message in',
False)
spec.addBoolOption(
'print_result',
'TODO: fill this help message in',
False)
spec.addIntOption(
'data_aug',
'specify data augumentation, 0-7, -1 mean random',
-1)
spec.addIntOption(
'ratio_pre_moves',
('how many moves to perform in each thread, before we use the '
'data to train the model'),
0)
spec.addFloatOption(
'start_ratio_pre_moves',
('how many moves to perform in each thread, before we use the '
'first sgf file to train the model'),
0.5)
spec.addIntOption(
'num_games_per_thread',
('For offline mode, it is the number of concurrent games per '
'thread, used to increase diversity of games; for selfplay mode, '
'it is the number of games played at each thread, and after that '
'we need to call restartAllGames() to resume.'),
-1)
spec.addIntOption(
'expected_num_clients',
'Expected number of clients',
-1
)
spec.addIntOption(
'num_future_actions',
'TODO: fill this help message in',
1)
spec.addIntOption(
'move_cutoff',
'Cutoff ply in replay',
-1)
spec.addStrOption(
'mode',
'TODO: fill this help message in',
'online')
spec.addBoolOption(
'black_use_policy_network_only',
'TODO: fill this help message in',
False)
spec.addBoolOption(
'white_use_policy_network_only',
'TODO: fill this help message in',
False)
spec.addIntOption(
'ply_pass_enabled',
'TODO: fill this help message in',
0)
spec.addBoolOption(
'use_mcts',
'TODO: fill this help message in',
False)
spec.addBoolOption(
'use_mcts_ai2',
'TODO: fill this help message in',
False)
spec.addFloatOption(
'white_puct',
'PUCT for white when it is > 0.0. If it is -1 then we use'
'the same puct for both side (specified by mcts_options).'
'A HACK to use different puct for different model. Should'
'be replaced by a more systematic approach.',
-1.0)
spec.addIntOption(
'white_mcts_rollout_per_batch',
'white mcts rollout per batch',
-1)
spec.addIntOption(
'white_mcts_rollout_per_thread',
'white mcts rollout per thread',
-1)
spec.addBoolOption(
'use_df_feature',
'TODO: fill this help message in',
False)
spec.addStrOption(
'dump_record_prefix',
'TODO: fill this help message in',
'')
spec.addIntOption(
'policy_distri_cutoff',
'TODO: fill this help message in',
0)
spec.addFloatOption(
'resign_thres',
'TODO: fill this help message in',
0.0)
spec.addBoolOption(
'following_pass',
'TODO: fill this help message in',
False)
spec.addIntOption(
'selfplay_timeout_usec',
'TODO: fill this help message in',
0)
spec.addIntOption(
'gpu',
'TODO: fill this help message in',
-1)
spec.addBoolOption(
'policy_distri_training_for_all',
'TODO: fill this help message in',
False)
spec.addBoolOption(
'parameter_print',
'TODO: fill this help message in',
True)
spec.addIntOption(
'batchsize',
'batch size',
128)
spec.addIntOption(
'batchsize2',
'batch size',
-1)
spec.addIntOption(
'T',
'number of timesteps',
6)
spec.addIntOption(
'selfplay_init_num',
('Initial number of selfplay games to generate before training a '
'new model'),
2000)
spec.addIntOption(
'selfplay_update_num',
('Additional number of selfplay games to generate after a model '
'is updated'),
1000)
spec.addBoolOption(
'selfplay_async',
('Whether to use async mode in selfplay'),
False)
spec.addIntOption(
'eval_num_games',
('number of evaluation to be performed to decide whether a model '
'is better than the other'),
400)
spec.addFloatOption(
'eval_winrate_thres',
'Win rate threshold for evalution',
0.55)
spec.addIntOption(
'eval_old_model',
('If specified, then we directly switch to evaluation mode '
'between the loaded model and the old model specified by this '
'switch'),
-1)
spec.addStrOption(
'eval_model_pair',
('If specified for df_selfplay.py, then the two models will be '
'evaluated on this client'),
'')
spec.addStrOption(
'comment',
'Comment for this run',
'')
spec.addBoolOption(
'cheat_eval_new_model_wins_half',
'When enabled, in evaluation mode, when the game '
'finishes, the player with the most recent model gets 100% '
'win rate half of the time.'
'This is used to test the framework',
False)
spec.addBoolOption(
'cheat_selfplay_random_result',
'When enabled, in selfplay mode the result of the game is random'
'This is used to test the framework',
False)
spec.addIntOption(
'suicide_after_n_games',
'return after n games have finished, -1 means it never ends',
-1)
spec.merge(PyOptionSpec.fromClasses((ContextArgs, MoreLabels)))
return spec
def get_option_spec(cls):
spec = PyOptionSpec()
spec.addBoolOption('actor_only', 'TODO: fill this help message in',
False)
spec.addStrListOption(
'list_files', 'Provide a list of json files for offline training',
[])
spec.addIntOption('port', 'TODO: fill this help message in', 5556)
spec.addStrOption('server_addr', 'TODO: fill this help message in', '')
spec.addStrOption('server_id', 'TODO: fill this help message in', '')
spec.addIntOption('q_min_size', 'TODO: fill this help message in', 10)
spec.addIntOption('q_max_size', 'TODO: fill this help message in',
1000)
spec.addIntOption('num_reader', 'TODO: fill this help message in', 50)
spec.addIntOption('num_reset_ranking',
'TODO: fill this help message in', 5000)
spec.addIntOption(
'client_max_delay_sec',
'Maximum amount of allowed delays in sec. If the client '
'didn\'t respond after that, we think it is dead.', 1200)
spec.addBoolOption('verbose', 'TODO: fill this help message in', False)
spec.addBoolOption('keep_prev_selfplay',
'TODO: fill this help message in', False)
spec.addIntOption(
'num_games_per_thread',
('For offline mode, it is the number of concurrent games per '
'thread, used to increase diversity of games; for selfplay mode, '
'it is the number of games played at each thread, and after that '
'we need to call restartAllGames() to resume.'), -1)
spec.addIntOption('expected_num_clients', 'Expected number of clients',
-1)
spec.addIntOption('checkers_num_future_actions',
'TODO: fill this help message in', 1)
spec.addStrOption('mode', 'TODO: fill this help message in', 'play')
spec.addBoolOption('black_use_policy_network_only',
'TODO: fill this help message in', False)
spec.addBoolOption('white_use_policy_network_only',
'TODO: fill this help message in', False)
spec.addBoolOption('use_mcts', 'TODO: fill this help message in',
False)
spec.addBoolOption('use_mcts_ai2', 'TODO: fill this help message in',
False)
spec.addFloatOption(
'white_puct',
'PUCT for white when it is > 0.0. If it is -1 then we use'
'the same puct for both side (specified by mcts_options).'
'A HACK to use different puct for different model. Should'
'be replaced by a more systematic approach.', -1.0)
spec.addIntOption('white_mcts_rollout_per_batch',
'white mcts rollout per batch', -1)
spec.addIntOption('white_mcts_rollout_per_thread',
'white mcts rollout per thread', -1)
spec.addStrOption('dump_record_prefix',
'TODO: fill this help message in', '')
spec.addStrOption('selfplay_records_directory',
'TODO: fill this help message in', '')
spec.addStrOption('eval_records_directory',
'TODO: fill this help message in', '')
spec.addStrOption('records_buffer_directory',
'TODO: fill this help message in', '')
spec.addIntOption('policy_distri_cutoff',
'first N moves will be randomly', 0)
spec.addIntOption('selfplay_timeout_usec',
'TODO: fill this help message in', 0)
spec.addIntOption('gpu', 'TODO: fill this help message in', -1)
spec.addBoolOption('policy_distri_training_for_all',
'TODO: fill this help message in', False)
spec.addBoolOption('parameter_print',
'TODO: fill this help message in', True)
spec.addIntOption('batchsize', 'batch size', 128)
spec.addIntOption('batchsize2', 'batch size', -1)
spec.addIntOption('T', 'number of timesteps', 6)
spec.addIntOption(
'selfplay_init_num',
('Initial number of selfplay games to generate before training a '
'new model'), 2000)
spec.addIntOption(
'selfplay_update_num',
('Additional number of selfplay games to generate after a model '
'is updated'), 1000)
spec.addBoolOption('selfplay_async',
('Whether to use async mode in selfplay'), False)
spec.addIntOption(
'eval_num_games',
('number of evaluation to be performed to decide whether a model '
'is better than the other'), 400)
spec.addFloatOption('eval_winrate_thres',
'Win rate threshold for evalution', 0.55)
spec.addIntOption(
'eval_old_model',
('If specified, then we directly switch to evaluation mode '
'between the loaded model and the old model specified by this '
'switch'), -1)
spec.addStrOption(
'eval_model_pair',
('If specified for df_selfplay.py, then the two models will be '
'evaluated on this client'), '')
spec.addBoolOption(
'cheat_eval_new_model_wins_half',
'When enabled, in evaluation mode, when the game '
'finishes, the player with the most recent model gets 100% '
'win rate half of the time.'
'This is used to test the framework', False)
spec.addBoolOption(
'cheat_selfplay_random_result',
'When enabled, in selfplay mode the result of the game is random'
'This is used to test the framework', False)
spec.addBoolOption('human_plays_for_black', '', False)
spec.addIntOption(
'suicide_after_n_games',
'return after n games have finished, -1 means it never ends', -1)
spec.merge(PyOptionSpec.fromClasses((ContextArgs, MoreLabels)))
return spec
def get_option_spec(cls):
spec = PyOptionSpec()
spec.addIntOption(
'anneal_entropy',
'',
0)
spec.addFloatOption(
'entropy_ratio',
'',
0.01)
spec.addFloatOption(
'init_entropy_ratio',
'the entropy ratio we put on PolicyGradient',
0.01)
spec.addFloatOption(
'final_entropy_ratio',
'',
0.0)
spec.addIntOption(
'entropy_decay_epoch',
'decay the entropy linearly during the first k epochs',
100)
spec.addFloatOption(
'min_prob',
'mininal probability used in training',
1e-6)
spec.addFloatOption(
'ratio_clamp',
'maximum importance sampling ratio',
0.1)
spec.addFloatOption(
'max_grad_norm',
'maximum norm of gradient',
0.5)
spec.addFloatOption(
'discount',
'exponential discount rate',
0.99)
return spec