def main(args):
np.random.seed(2019)
if not args.show_graphs:
matplotlib.use("pdf")
if args.gpus is not None:
os.environ["CUDA_VISIBLE_DEVICES"] = args.gpus
saver = Saver(None)
if args.save:
dir_variables = [
SAVE_PREFIX, args.game, args.num_blocks, args.encoder_learning_rate, args.encoder_optimizer, args.num_steps
]
dir_switches = [
"no_sample", "only_one_q_value" "disable_batch_norm"
]
saver.create_dir_name(SAVE_TEMPLATE, dir_variables, dir_switches, args)
saver.create_dir(add_run_subdir=True)
saver.save_by_print(args, "settings")
logger = LearnExpectationContinuousLogger(save_file=saver.get_save_file("main", "log"), print_logs=True)
logger.silence_tensorflow()
if args.game == constants.GAME_BREAKOUT:
num_actions = 6
else:
raise ValueError("Unknown game.")
runner = QRunnerMinAtar(
args.load_path, num_actions, logger, saver, args.num_blocks,
args.encoder_learning_rate, args.weight_decay, args.encoder_optimizer, args.num_steps,
disable_batch_norm=args.disable_batch_norm, disable_softplus=args.disable_softplus,
no_sample=args.no_sample, only_one_q_value=args.only_one_q_value,
validation_freq=args.validation_freq, validation_fraction=args.validation_fraction, summaries=args.summaries,
load_model_path=args.load_model_path, zero_sd_after_training=args.zero_sd_after_training
)
runner.setup()
runner.main_training_loop()
if args.save_model:
runner.save_model()
runner.evaluate_and_visualize()
runner.close_model_session()
def main(args):
np.random.seed(2019)
if not args.show_graphs:
matplotlib.use("pdf")
if args.gpus is not None:
os.environ["CUDA_VISIBLE_DEVICES"] = args.gpus
saver = Saver(None)
if args.save:
dir_variables = [
SAVE_PREFIX.format(args.base_dir), args.num_pucks, args.grid_size, args.grid_size, args.num_blocks, args.num_components,
args.beta0, args.beta1, args.beta2, args.encoder_learning_rate, args.encoder_optimizer, args.num_steps
]
dir_switches = ["no_sample", "only_one_q_value", "gt_q_values", "disable_batch_norm"]
saver.create_dir_name(SAVE_TEMPLATE, dir_variables, dir_switches, args)
saver.create_dir(add_run_subdir=True)
saver.save_by_print(args, "settings")
logger = LearnExpectationContinuousLogger(save_file=saver.get_save_file("main", "log"), print_logs=True)
logger.silence_tensorflow()
runner = QGMMPriorRunner(
args.load_path, args.grid_size, args.num_pucks, logger, saver, args.num_blocks, args.num_components,
args.hiddens, args.encoder_learning_rate, args.beta0, args.beta1, args.beta2, args.weight_decay,
args.encoder_optimizer, args.num_steps, disable_batch_norm=args.disable_batch_norm,
disable_softplus=args.disable_softplus, no_sample=args.no_sample, only_one_q_value=args.only_one_q_value,
gt_q_values=args.gt_q_values, disable_resize=args.disable_resize, oversample=args.oversample,
validation_freq=args.validation_freq, validation_fraction=args.validation_fraction, summaries=args.summaries,
load_model_path=args.load_model_path, include_goal_states=args.include_goal_states,
q_values_noise_sd=args.q_values_noise_sd, new_dones=args.new_dones
)
runner.setup()
runner.main_training_loop()
if args.save_model:
runner.save_model()
runner.evaluate_and_visualize()
runner.close_model_session()
def main(args):
assert args.active_indices is not None
assert args.eval_indices is not None
assert args.dones_index is not None
np.random.seed(2019)
if not args.show_graphs and not args.show_qs:
matplotlib.use("pdf")
if args.gpus is not None:
os.environ["CUDA_VISIBLE_DEVICES"] = args.gpus
saver = Saver(None)
if args.save:
dir_variables = [
SAVE_PREFIX.format(args.base_dir), args.num_pucks, args.grid_size, args.grid_size, args.num_blocks, args.num_components,
args.beta0, args.beta1, args.beta2, args.beta3, args.encoder_learning_rate, args.encoder_optimizer,
args.num_steps
]
dir_switches = [
"no_sample", "only_one_q_value", "disable_batch_norm", "train_prior",
"post_train_prior", "post_train_t_and_prior", "post_train_hmm"
]
saver.create_dir_name(SAVE_TEMPLATE, dir_variables, dir_switches, args)
saver.create_dir(add_run_subdir=True)
saver.save_by_print(args, "settings")
logger = LearnExpectationContinuousLogger(save_file=saver.get_save_file("main", "log"), print_logs=True)
logger.silence_tensorflow()
runner = QHMMPriorMultiTask(
args.load_path, args.grid_size, args.num_pucks, logger, saver, args.num_blocks, args.num_components,
args.hiddens, args.encoder_learning_rate, args.beta0, args.beta1, args.beta2, args.weight_decay,
args.encoder_optimizer, args.num_steps, args.active_indices, args.eval_indices, args.dones_index,
disable_batch_norm=args.disable_batch_norm, disable_softplus=args.disable_softplus, no_sample=args.no_sample,
only_one_q_value=args.only_one_q_value, oversample=args.oversample,
validation_freq=args.validation_freq, validation_fraction=args.validation_fraction, summaries=args.summaries,
load_model_path=args.load_model_path, beta3=args.beta3, post_train_prior=args.post_train_prior,
post_train_hmm=args.post_train_hmm, post_train_t_and_prior=args.post_train_t_and_prior,
save_gifs=args.save_gifs, hard_abstract_state=args.hard_abstract_state,
zero_sd_after_training=args.zero_sd_after_training, prune_abstraction=args.prune_abstraction,
prune_threshold=args.prune_threshold, old_bn_settings=args.old_bn_settings,
include_goal_states=args.include_goal_states, shift_q_values=args.shift_q_values,
soft_picture_goals=args.soft_picture_goals, goal_rewards_threshold=args.goal_rewards_threshold,
q_values_indices=args.q_values_indices, fast_eval=args.fast_eval,
sample_abstract_state=args.sample_abstract_state, softmax_policy=args.softmax_policy,
softmax_policy_temp=args.softmax_policy_temp, show_qs=args.show_qs,
fix_prior_training=args.fix_prior_training, model_learning_rate=args.model_learning_rate,
random_shape=args.random_shape
)
runner.setup()
runner.main_training_loop()
if args.post_train_hmm:
runner.post_hmm_training_loop(args.post_train_hmm_steps)
if args.save_model:
runner.save_model()
runner.evaluate_and_visualize()
runner.close_model_session()
def main(args):
np.random.seed(2019)
if not args.show_graphs:
matplotlib.use("pdf")
if args.gpus is not None:
os.environ["CUDA_VISIBLE_DEVICES"] = args.gpus
saver = Saver(None)
if args.save:
dir_variables = [
SAVE_PREFIX.format(args.base_dir), args.game, args.num_blocks,
args.num_components, args.beta0, args.beta1, args.beta2,
args.beta3, args.encoder_learning_rate, args.encoder_optimizer,
args.num_steps
]
dir_switches = [
"no_sample", "only_one_q_value", "disable_batch_norm",
"train_prior", "post_train_prior", "post_train_t_and_prior",
"post_train_hmm"
]
saver.create_dir_name(SAVE_TEMPLATE, dir_variables, dir_switches, args)
saver.create_dir(add_run_subdir=True)
saver.save_by_print(args, "settings")
logger = LearnExpectationContinuousLogger(save_file=saver.get_save_file(
"main", "log"),
print_logs=True)
logger.silence_tensorflow()
num_actions = 6
model_learning_rate = args.encoder_learning_rate
if args.model_learning_rate is not None:
model_learning_rate = args.model_learning_rate
runner = QHMMPriorRunnerMinAtar(
args.load_path,
args.game,
num_actions,
logger,
saver,
args.num_blocks,
args.num_components,
args.encoder_learning_rate,
args.beta0,
args.beta1,
args.beta2,
args.weight_decay,
args.encoder_optimizer,
args.num_steps,
disable_batch_norm=args.disable_batch_norm,
disable_softplus=args.disable_softplus,
no_sample=args.no_sample,
only_one_q_value=args.only_one_q_value,
validation_freq=args.validation_freq,
validation_fraction=args.validation_fraction,
summaries=args.summaries,
load_model_path=args.load_model_path,
beta3=args.beta3,
post_train_prior=args.post_train_prior,
post_train_hmm=args.post_train_hmm,
post_train_t_and_prior=args.post_train_t_and_prior,
save_gifs=args.save_gifs,
zero_sd_after_training=args.zero_sd_after_training,
hard_abstract_state=args.hard_abstract_state,
freeze_hmm_no_entropy_at=args.freeze_hmm_no_entropy_at,
cluster_predict_qs=args.cluster_predict_qs,
cluster_predict_qs_weight=args.cluster_predict_qs_weight,
prune_threshold=args.prune_threshold,
prune_abstraction=args.prune_abstraction,
prune_abstraction_new_means=args.prune_abstraction_new_means,
sample_abstract_state=args.sample_abstract_state,
softmax_policy=args.softmax_policy,
softmax_policy_temp=args.softmax_policy_temp,
discount=args.discount,
fix_prior_training=args.fix_prior_training,
model_learning_rate=model_learning_rate,
q_scaling_factor=args.q_scaling_factor,
eval_episodes=args.eval_episodes)
runner.setup()
runner.main_training_loop()
if args.post_train_hmm:
runner.post_hmm_training_loop(args.post_train_hmm_steps)
if args.save_model:
runner.save_model()
runner.evaluate_and_visualize()
runner.close_model_session()