def __init__(self, algo, args, exp_name): self.args = args self.algo = algo env = algo.env baseline = algo.baseline # Logger default_log_dir = config.LOG_DIR if args.log_dir is None: log_dir = osp.join(default_log_dir, exp_name) else: log_dir = args.log_dir tabular_log_file = osp.join(log_dir, args.tabular_log_file) text_log_file = osp.join(log_dir, args.text_log_file) params_log_file = osp.join(log_dir, args.params_log_file) logger.log_parameters_lite(params_log_file, args) logger.add_text_output(text_log_file) logger.add_tabular_output(tabular_log_file) logger.set_snapshot_dir(log_dir) logger.set_snapshot_mode(args.snapshot_mode) logger.set_log_tabular_only(args.log_tabular_only) logger.push_prefix("[%s] " % exp_name) prev_snapshot_dir = logger.get_snapshot_dir() prev_mode = logger.get_snapshot_mode()
def main(): now = datetime.datetime.now(dateutil.tz.tzlocal()) rand_id = str(uuid.uuid4())[:5] timestamp = now.strftime('%Y_%m_%d_%H_%M_%S_%f_%Z') default_exp_name = 'experiment_%s_%s' % (timestamp, rand_id) parser = argparse.ArgumentParser() parser.add_argument('--exp_name', type=str, default=default_exp_name, help='Name of the experiment.') parser.add_argument('--discount', type=float, default=0.95) parser.add_argument('--gae_lambda', type=float, default=0.99) parser.add_argument('--reward_scale', type=float, default=1.0) parser.add_argument('--enable_obsnorm', action='store_true', default=False) parser.add_argument('--chunked', action='store_true', default=False) parser.add_argument('--n_iter', type=int, default=250) parser.add_argument('--sampler_workers', type=int, default=1) parser.add_argument('--max_traj_len', type=int, default=250) parser.add_argument('--update_curriculum', action='store_true', default=False) parser.add_argument('--anneal_step_size', type=int, default=0) parser.add_argument('--n_timesteps', type=int, default=8000) parser.add_argument('--control', type=str, default='centralized') parser.add_argument('--buffer_size', type=int, default=1) parser.add_argument('--radius', type=float, default=0.015) parser.add_argument('--n_evaders', type=int, default=10) parser.add_argument('--n_pursuers', type=int, default=8) parser.add_argument('--n_poison', type=int, default=10) parser.add_argument('--n_coop', type=int, default=4) parser.add_argument('--n_sensors', type=int, default=30) parser.add_argument('--sensor_range', type=str, default='0.2') parser.add_argument('--food_reward', type=float, default=5) parser.add_argument('--poison_reward', type=float, default=-1) parser.add_argument('--encounter_reward', type=float, default=0.05) parser.add_argument('--reward_mech', type=str, default='local') parser.add_argument('--recurrent', type=str, default=None) parser.add_argument('--baseline_type', type=str, default='linear') parser.add_argument('--policy_hidden_sizes', type=str, default='128,128') parser.add_argument('--baseline_hidden_sizes', type=str, default='128,128') parser.add_argument('--max_kl', type=float, default=0.01) parser.add_argument('--log_dir', type=str, required=False) parser.add_argument('--tabular_log_file', type=str, default='progress.csv', help='Name of the tabular log file (in csv).') parser.add_argument('--text_log_file', type=str, default='debug.log', help='Name of the text log file (in pure text).') parser.add_argument('--params_log_file', type=str, default='params.json', help='Name of the parameter log file (in json).') parser.add_argument('--seed', type=int, help='Random seed for numpy') parser.add_argument('--args_data', type=str, help='Pickled data for stub objects') parser.add_argument('--snapshot_mode', type=str, default='all', help='Mode to save the snapshot. Can be either "all" ' '(all iterations will be saved), "last" (only ' 'the last iteration will be saved), or "none" ' '(do not save snapshots)') parser.add_argument( '--log_tabular_only', type=ast.literal_eval, default=False, help='Whether to only print the tabular log information (in a horizontal format)') args = parser.parse_args() parallel_sampler.initialize(n_parallel=args.sampler_workers) if args.seed is not None: set_seed(args.seed) parallel_sampler.set_seed(args.seed) args.hidden_sizes = tuple(map(int, args.policy_hidden_sizes.split(','))) centralized = True if args.control == 'centralized' else False sensor_range = np.array(map(float, args.sensor_range.split(','))) if len(sensor_range) == 1: sensor_range = sensor_range[0] else: assert sensor_range.shape == (args.n_pursuers,) env = MAWaterWorld(args.n_pursuers, args.n_evaders, args.n_coop, args.n_poison, radius=args.radius, n_sensors=args.n_sensors, food_reward=args.food_reward, poison_reward=args.poison_reward, encounter_reward=args.encounter_reward, reward_mech=args.reward_mech, sensor_range=sensor_range, obstacle_loc=None) env = TfEnv( RLLabEnv( StandardizedEnv(env, scale_reward=args.reward_scale, enable_obsnorm=args.enable_obsnorm), mode=args.control)) if args.buffer_size > 1: env = ObservationBuffer(env, args.buffer_size) if args.recurrent: feature_network = MLP( name='feature_net', input_shape=(env.spec.observation_space.flat_dim + env.spec.action_space.flat_dim,), output_dim=16, hidden_sizes=(128, 64, 32), hidden_nonlinearity=tf.nn.tanh, output_nonlinearity=None) if args.recurrent == 'gru': policy = GaussianGRUPolicy(env_spec=env.spec, feature_network=feature_network, hidden_dim=int(args.policy_hidden_sizes), name='policy') elif args.recurrent == 'lstm': policy = GaussianLSTMPolicy(env_spec=env.spec, feature_network=feature_network, hidden_dim=int(args.policy_hidden_sizes), name='policy') else: policy = GaussianMLPPolicy( name='policy', env_spec=env.spec, hidden_sizes=tuple(map(int, args.policy_hidden_sizes.split(','))), min_std=10e-5) if args.baseline_type == 'linear': baseline = LinearFeatureBaseline(env_spec=env.spec) elif args.baseline_type == 'mlp': raise NotImplementedError() # baseline = GaussianMLPBaseline( # env_spec=env.spec, hidden_sizes=tuple(map(int, args.baseline_hidden_sizes.split(',')))) else: baseline = ZeroBaseline(env_spec=env.spec) # logger default_log_dir = config.LOG_DIR if args.log_dir is None: log_dir = osp.join(default_log_dir, args.exp_name) else: log_dir = args.log_dir tabular_log_file = osp.join(log_dir, args.tabular_log_file) text_log_file = osp.join(log_dir, args.text_log_file) params_log_file = osp.join(log_dir, args.params_log_file) logger.log_parameters_lite(params_log_file, args) logger.add_text_output(text_log_file) logger.add_tabular_output(tabular_log_file) prev_snapshot_dir = logger.get_snapshot_dir() prev_mode = logger.get_snapshot_mode() logger.set_snapshot_dir(log_dir) logger.set_snapshot_mode(args.snapshot_mode) logger.set_log_tabular_only(args.log_tabular_only) logger.push_prefix("[%s] " % args.exp_name) algo = TRPO( env=env, policy=policy, baseline=baseline, batch_size=args.n_timesteps, max_path_length=args.max_traj_len, #max_path_length_limit=args.max_path_length_limit, update_max_path_length=args.update_curriculum, anneal_step_size=args.anneal_step_size, n_itr=args.n_iter, discount=args.discount, gae_lambda=args.gae_lambda, step_size=args.max_kl, optimizer=ConjugateGradientOptimizer(hvp_approach=FiniteDifferenceHvp(base_eps=1e-5)) if args.recurrent else None, mode=args.control if not args.chunked else 'chunk_{}'.format(args.control),) algo.train()
def run_experiment(argv): default_log_dir = config.LOG_DIR now = datetime.datetime.now(dateutil.tz.tzlocal()) # avoid name clashes when running distributed jobs rand_id = str(uuid.uuid4())[:5] timestamp = now.strftime('%Y_%m_%d_%H_%M_%S_%f_%Z') default_exp_name = 'experiment_%s_%s' % (timestamp, rand_id) parser = argparse.ArgumentParser() parser.add_argument('--n_parallel', type=int, default=1, help='Number of parallel workers to perform rollouts. 0 => don\'t start any workers') parser.add_argument( '--exp_name', type=str, default=default_exp_name, help='Name of the experiment.') parser.add_argument('--log_dir', type=str, default=None, help='Path to save the log and iteration snapshot.') parser.add_argument('--snapshot_mode', type=str, default='all', help='Mode to save the snapshot. Can be either "all" ' '(all iterations will be saved), "last" (only ' 'the last iteration will be saved), "gap" (every' '`snapshot_gap` iterations are saved), or "none" ' '(do not save snapshots)') parser.add_argument('--snapshot_gap', type=int, default=1, help='Gap between snapshot iterations.') parser.add_argument('--tabular_log_file', type=str, default='progress.csv', help='Name of the tabular log file (in csv).') parser.add_argument('--text_log_file', type=str, default='debug.log', help='Name of the text log file (in pure text).') parser.add_argument('--params_log_file', type=str, default='params.json', help='Name of the parameter log file (in json).') parser.add_argument('--variant_log_file', type=str, default='variant.json', help='Name of the variant log file (in json).') parser.add_argument('--resume_from', type=str, default=None, help='Name of the pickle file to resume experiment from.') parser.add_argument('--plot', type=ast.literal_eval, default=False, help='Whether to plot the iteration results') parser.add_argument('--log_tabular_only', type=ast.literal_eval, default=False, help='Whether to only print the tabular log information (in a horizontal format)') parser.add_argument('--seed', type=int, help='Random seed for numpy') parser.add_argument('--args_data', type=str, help='Pickled data for stub objects') parser.add_argument('--variant_data', type=str, help='Pickled data for variant configuration') parser.add_argument('--use_cloudpickle', type=ast.literal_eval, default=False) args = parser.parse_args(argv[1:]) if args.seed is not None: set_seed(args.seed) if args.n_parallel > 0: from rllab.sampler import parallel_sampler parallel_sampler.initialize(n_parallel=args.n_parallel) if args.seed is not None: parallel_sampler.set_seed(args.seed) if args.plot: from rllab.plotter import plotter plotter.init_worker() if args.log_dir is None: log_dir = osp.join(default_log_dir, args.exp_name) else: log_dir = args.log_dir tabular_log_file = osp.join(log_dir, args.tabular_log_file) text_log_file = osp.join(log_dir, args.text_log_file) params_log_file = osp.join(log_dir, args.params_log_file) if args.variant_data is not None: variant_data = pickle.loads(base64.b64decode(args.variant_data)) variant_log_file = osp.join(log_dir, args.variant_log_file) logger.log_variant(variant_log_file, variant_data) else: variant_data = None if not args.use_cloudpickle: logger.log_parameters_lite(params_log_file, args) logger.add_text_output(text_log_file) logger.add_tabular_output(tabular_log_file) prev_snapshot_dir = logger.get_snapshot_dir() prev_mode = logger.get_snapshot_mode() logger.set_snapshot_dir(log_dir) logger.set_snapshot_mode(args.snapshot_mode) logger.set_snapshot_gap(args.snapshot_gap) logger.set_log_tabular_only(args.log_tabular_only) logger.push_prefix("[%s] " % args.exp_name) if args.resume_from is not None: data = joblib.load(args.resume_from) assert 'algo' in data algo = data['algo'] algo.train() else: # read from stdin if args.use_cloudpickle: import cloudpickle method_call = cloudpickle.loads(base64.b64decode(args.args_data)) method_call(variant_data) else: data = pickle.loads(base64.b64decode(args.args_data)) maybe_iter = concretize(data) if is_iterable(maybe_iter): for _ in maybe_iter: pass logger.set_snapshot_mode(prev_mode) logger.set_snapshot_dir(prev_snapshot_dir) logger.remove_tabular_output(tabular_log_file) logger.remove_text_output(text_log_file) logger.pop_prefix()
def main(): now = datetime.datetime.now(dateutil.tz.tzlocal()) rand_id = str(uuid.uuid4())[:5] timestamp = now.strftime('%Y_%m_%d_%H_%M_%S_%f_%Z') default_exp_name = 'experiment_%s_%s' % (timestamp, rand_id) parser = argparse.ArgumentParser() parser.add_argument( '--exp_name', type=str, default=default_exp_name, help='Name of the experiment.') parser.add_argument('--discount', type=float, default=0.95) parser.add_argument('--gae_lambda', type=float, default=0.99) parser.add_argument('--n_iter', type=int, default=250) parser.add_argument('--sampler_workers', type=int, default=1) parser.add_argument('--max_traj_len', type=int, default=250) parser.add_argument('--update_curriculum', action='store_true', default=False) parser.add_argument('--n_timesteps', type=int, default=8000) parser.add_argument('--control', type=str, default='centralized') parser.add_argument('--control', type=str, default='centralized') parser.add_argument('--buffer_size', type=int, default=1) parser.add_argument('--n_good', type=int, default=3) parser.add_argument('--n_hostage', type=int, default=5) parser.add_argument('--n_bad', type=int, default=5) parser.add_argument('--n_coop_save', type=int, default=2) parser.add_argument('--n_coop_avoid', type=int, default=2) parser.add_argument('--n_sensors', type=int, default=20) parser.add_argument('--sensor_range', type=float, default=0.2) parser.add_argument('--save_reward', type=float, default=3) parser.add_argument('--hit_reward', type=float, default=-1) parser.add_argument('--encounter_reward', type=float, default=0.01) parser.add_argument('--bomb_reward', type=float, default=-10.) parser.add_argument('--recurrent', action='store_true', default=False) parser.add_argument('--baseline_type', type=str, default='linear') parser.add_argument('--policy_hidden_sizes', type=str, default='128,128') parser.add_argument('--baselin_hidden_sizes', type=str, default='128,128') parser.add_argument('--max_kl', type=float, default=0.01) parser.add_argument('--log_dir', type=str, required=False) parser.add_argument('--tabular_log_file', type=str, default='progress.csv', help='Name of the tabular log file (in csv).') parser.add_argument('--text_log_file', type=str, default='debug.log', help='Name of the text log file (in pure text).') parser.add_argument('--params_log_file', type=str, default='params.json', help='Name of the parameter log file (in json).') parser.add_argument('--seed', type=int, help='Random seed for numpy') parser.add_argument('--args_data', type=str, help='Pickled data for stub objects') parser.add_argument('--snapshot_mode', type=str, default='all', help='Mode to save the snapshot. Can be either "all" ' '(all iterations will be saved), "last" (only ' 'the last iteration will be saved), or "none" ' '(do not save snapshots)') parser.add_argument('--log_tabular_only', type=ast.literal_eval, default=False, help='Whether to only print the tabular log information (in a horizontal format)') args = parser.parse_args() parallel_sampler.initialize(n_parallel=args.sampler_workers) if args.seed is not None: set_seed(args.seed) parallel_sampler.set_seed(args.seed) args.hidden_sizes = tuple(map(int, args.policy_hidden_sizes.split(','))) centralized = True if args.control == 'centralized' else False sensor_range = np.array(map(float, args.sensor_range.split(','))) assert sensor_range.shape == (args.n_pursuers,) env = ContinuousHostageWorld(args.n_good, args.n_hostage, args.n_bad, args.n_coop_save, args.n_coop_avoid, n_sensors=args.n_sensors, sensor_range=args.sensor_range, save_reward=args.save_reward, hit_reward=args.hit_reward, encounter_reward=args.encounter_reward, bomb_reward=args.bomb_reward) env = RLLabEnv(StandardizedEnv(env), mode=args.control) if args.buffer_size > 1: env = ObservationBuffer(env, args.buffer_size) if args.recurrent: policy = GaussianGRUPolicy(env_spec=env.spec, hidden_sizes=args.hidden_sizes) else: policy = GaussianMLPPolicy(env_spec=env.spec, hidden_sizes=args.hidden_sizes) if args.baseline_type == 'linear': baseline = LinearFeatureBaseline(env_spec=env.spec) else: baseline = ZeroBaseline(obsfeat_space) # logger default_log_dir = config.LOG_DIR if args.log_dir is None: log_dir = osp.join(default_log_dir, args.exp_name) else: log_dir = args.log_dir tabular_log_file = osp.join(log_dir, args.tabular_log_file) text_log_file = osp.join(log_dir, args.text_log_file) params_log_file = osp.join(log_dir, args.params_log_file) logger.log_parameters_lite(params_log_file, args) logger.add_text_output(text_log_file) logger.add_tabular_output(tabular_log_file) prev_snapshot_dir = logger.get_snapshot_dir() prev_mode = logger.get_snapshot_mode() logger.set_snapshot_dir(log_dir) logger.set_snapshot_mode(args.snapshot_mode) logger.set_log_tabular_only(args.log_tabular_only) logger.push_prefix("[%s] " % args.exp_name) algo = TRPO(env=env, policy=policy, baseline=baseline, batch_size=args.n_timesteps, max_path_length=args.max_traj_len, n_itr=args.n_iter, discount=args.discount, step_size=args.max_kl, mode=args.control,) algo.train()
def setup(self, env, policy, start_itr): if not self.args.algo == 'thddpg': # Baseline if self.args.baseline_type == 'linear': baseline = LinearFeatureBaseline(env_spec=env.spec) elif self.args.baseline_type == 'zero': baseline = ZeroBaseline(env_spec=env.spec) else: raise NotImplementedError(self.args.baseline_type) if self.args.control == 'concurrent': baseline = [baseline for _ in range(len(env.agents))] # Logger default_log_dir = config.LOG_DIR if self.args.log_dir is None: log_dir = osp.join(default_log_dir, self.args.exp_name) else: log_dir = self.args.log_dir tabular_log_file = osp.join(log_dir, self.args.tabular_log_file) text_log_file = osp.join(log_dir, self.args.text_log_file) params_log_file = osp.join(log_dir, self.args.params_log_file) logger.log_parameters_lite(params_log_file, self.args) logger.add_text_output(text_log_file) logger.add_tabular_output(tabular_log_file) prev_snapshot_dir = logger.get_snapshot_dir() prev_mode = logger.get_snapshot_mode() logger.set_snapshot_dir(log_dir) logger.set_snapshot_mode(self.args.snapshot_mode) logger.set_log_tabular_only(self.args.log_tabular_only) logger.push_prefix("[%s] " % self.args.exp_name) if self.args.algo == 'tftrpo': algo = MATRPO(env=env, policy_or_policies=policy, baseline_or_baselines=baseline, batch_size=self.args.batch_size, start_itr=start_itr, max_path_length=self.args.max_path_length, n_itr=self.args.n_iter, discount=self.args.discount, gae_lambda=self.args.gae_lambda, step_size=self.args.step_size, optimizer=ConjugateGradientOptimizer( hvp_approach=FiniteDifferenceHvp(base_eps=1e-5)) if self.args.recurrent else None, ma_mode=self.args.control) elif self.args.algo == 'thddpg': qfunc = thContinuousMLPQFunction(env_spec=env.spec) if self.args.exp_strategy == 'ou': es = OUStrategy(env_spec=env.spec) elif self.args.exp_strategy == 'gauss': es = GaussianStrategy(env_spec=env.spec) else: raise NotImplementedError() algo = thDDPG(env=env, policy=policy, qf=qfunc, es=es, batch_size=self.args.batch_size, max_path_length=self.args.max_path_length, epoch_length=self.args.epoch_length, min_pool_size=self.args.min_pool_size, replay_pool_size=self.args.replay_pool_size, n_epochs=self.args.n_iter, discount=self.args.discount, scale_reward=0.01, qf_learning_rate=self.args.qfunc_lr, policy_learning_rate=self.args.policy_lr, eval_samples=self.args.eval_samples, mode=self.args.control) return algo
def __init__(self, env, args): self.args = args # Parallel setup parallel_sampler.initialize(n_parallel=args.n_parallel) if args.seed is not None: set_seed(args.seed) parallel_sampler.set_seed(args.seed) env, policy = rllab_envpolicy_parser(env, args) if not args.algo == 'thddpg': # Baseline if args.baseline_type == 'linear': baseline = LinearFeatureBaseline(env_spec=env.spec) elif args.baseline_type == 'zero': baseline = ZeroBaseline(env_spec=env.spec) else: raise NotImplementedError(args.baseline_type) # Logger default_log_dir = config.LOG_DIR if args.log_dir is None: log_dir = osp.join(default_log_dir, args.exp_name) else: log_dir = args.log_dir tabular_log_file = osp.join(log_dir, args.tabular_log_file) text_log_file = osp.join(log_dir, args.text_log_file) params_log_file = osp.join(log_dir, args.params_log_file) logger.log_parameters_lite(params_log_file, args) logger.add_text_output(text_log_file) logger.add_tabular_output(tabular_log_file) prev_snapshot_dir = logger.get_snapshot_dir() prev_mode = logger.get_snapshot_mode() logger.set_snapshot_dir(log_dir) logger.set_snapshot_mode(args.snapshot_mode) logger.set_log_tabular_only(args.log_tabular_only) logger.push_prefix("[%s] " % args.exp_name) if args.algo == 'tftrpo': self.algo = TRPO( env=env, policy=policy, baseline=baseline, batch_size=args.batch_size, max_path_length=args.max_path_length, n_itr=args.n_iter, discount=args.discount, gae_lambda=args.gae_lambda, step_size=args.step_size, optimizer=ConjugateGradientOptimizer( hvp_approach=FiniteDifferenceHvp( base_eps=1e-5)) if args.recurrent else None, mode=args.control) elif args.algo == 'thddpg': qfunc = thContinuousMLPQFunction(env_spec=env.spec) if args.exp_strategy == 'ou': es = OUStrategy(env_spec=env.spec) elif args.exp_strategy == 'gauss': es = GaussianStrategy(env_spec=env.spec) else: raise NotImplementedError() self.algo = thDDPG(env=env, policy=policy, qf=qfunc, es=es, batch_size=args.batch_size, max_path_length=args.max_path_length, epoch_length=args.epoch_length, min_pool_size=args.min_pool_size, replay_pool_size=args.replay_pool_size, n_epochs=args.n_iter, discount=args.discount, scale_reward=0.01, qf_learning_rate=args.qfunc_lr, policy_learning_rate=args.policy_lr, eval_samples=args.eval_samples, mode=args.control)
def run_experiment(argv): default_log_dir = config.LOG_DIR now = datetime.datetime.now(dateutil.tz.tzlocal()) # avoid name clashes when running distributed jobs rand_id = str(uuid.uuid4())[:5] timestamp = now.strftime('%Y_%m_%d_%H_%M_%S_%f_%Z') default_exp_name = 'experiment_%s_%s' % (timestamp, rand_id) parser = argparse.ArgumentParser() parser.add_argument('--n_parallel', type=int, default=1, help='Number of parallel workers to perform rollouts.') parser.add_argument( '--exp_name', type=str, default=default_exp_name, help='Name of the experiment.') parser.add_argument('--log_dir', type=str, default=default_log_dir, help='Path to save the log and iteration snapshot.') parser.add_argument('--snapshot_mode', type=str, default='all', help='Mode to save the snapshot. Can be either "all" ' '(all iterations will be saved), "last" (only ' 'the last iteration will be saved), or "none" ' '(do not save snapshots)') parser.add_argument('--tabular_log_file', type=str, default='progress.csv', help='Name of the tabular log file (in csv).') parser.add_argument('--text_log_file', type=str, default='debug.log', help='Name of the text log file (in pure text).') parser.add_argument('--params_log_file', type=str, default='params.json', help='Name of the parameter log file (in json).') parser.add_argument('--plot', type=ast.literal_eval, default=False, help='Whether to plot the iteration results') parser.add_argument('--log_tabular_only', type=ast.literal_eval, default=False, help='Whether to only print the tabular log information (in a horizontal format)') parser.add_argument('--seed', type=int, help='Random seed for numpy') parser.add_argument('--args_data', type=str, help='Pickled data for stub objects') args = parser.parse_args(argv[1:]) from sandbox.vime.sampler import parallel_sampler_expl as parallel_sampler parallel_sampler.initialize(n_parallel=args.n_parallel) if args.seed is not None: set_seed(args.seed) parallel_sampler.set_seed(args.seed) if args.plot: from rllab.plotter import plotter plotter.init_worker() # read from stdin data = pickle.loads(base64.b64decode(args.args_data)) log_dir = args.log_dir # exp_dir = osp.join(log_dir, args.exp_name) tabular_log_file = osp.join(log_dir, args.tabular_log_file) text_log_file = osp.join(log_dir, args.text_log_file) params_log_file = osp.join(log_dir, args.params_log_file) logger.log_parameters_lite(params_log_file, args) logger.add_text_output(text_log_file) logger.add_tabular_output(tabular_log_file) prev_snapshot_dir = logger.get_snapshot_dir() prev_mode = logger.get_snapshot_mode() logger.set_snapshot_dir(log_dir) logger.set_snapshot_mode(args.snapshot_mode) logger.set_log_tabular_only(args.log_tabular_only) logger.push_prefix("[%s] " % args.exp_name) maybe_iter = concretize(data) if is_iterable(maybe_iter): for _ in maybe_iter: pass logger.set_snapshot_mode(prev_mode) logger.set_snapshot_dir(prev_snapshot_dir) logger.remove_tabular_output(tabular_log_file) logger.remove_text_output(text_log_file) logger.pop_prefix()
def setup(self, env, policy, start_itr): if not self.args.algo == 'thddpg': # Baseline if self.args.baseline_type == 'linear': baseline = LinearFeatureBaseline(env_spec=env.spec) elif self.args.baseline_type == 'zero': baseline = ZeroBaseline(env_spec=env.spec) else: raise NotImplementedError(self.args.baseline_type) if self.args.control == 'concurrent': baseline = [baseline for _ in range(len(env.agents))] # Logger default_log_dir = config.LOG_DIR if self.args.log_dir is None: log_dir = osp.join(default_log_dir, self.args.exp_name) else: log_dir = self.args.log_dir tabular_log_file = osp.join(log_dir, self.args.tabular_log_file) text_log_file = osp.join(log_dir, self.args.text_log_file) params_log_file = osp.join(log_dir, self.args.params_log_file) logger.log_parameters_lite(params_log_file, self.args) logger.add_text_output(text_log_file) logger.add_tabular_output(tabular_log_file) prev_snapshot_dir = logger.get_snapshot_dir() prev_mode = logger.get_snapshot_mode() logger.set_snapshot_dir(log_dir) logger.set_snapshot_mode(self.args.snapshot_mode) logger.set_log_tabular_only(self.args.log_tabular_only) logger.push_prefix("[%s] " % self.args.exp_name) if self.args.algo == 'tftrpo': algo = MATRPO( env=env, policy_or_policies=policy, baseline_or_baselines=baseline, batch_size=self.args.batch_size, start_itr=start_itr, max_path_length=self.args.max_path_length, n_itr=self.args.n_iter, discount=self.args.discount, gae_lambda=self.args.gae_lambda, step_size=self.args.step_size, optimizer=ConjugateGradientOptimizer( hvp_approach=FiniteDifferenceHvp( base_eps=1e-5)) if self.args.recurrent else None, ma_mode=self.args.control) elif self.args.algo == 'thddpg': qfunc = thContinuousMLPQFunction(env_spec=env.spec) if self.args.exp_strategy == 'ou': es = OUStrategy(env_spec=env.spec) elif self.args.exp_strategy == 'gauss': es = GaussianStrategy(env_spec=env.spec) else: raise NotImplementedError() algo = thDDPG(env=env, policy=policy, qf=qfunc, es=es, batch_size=self.args.batch_size, max_path_length=self.args.max_path_length, epoch_length=self.args.epoch_length, min_pool_size=self.args.min_pool_size, replay_pool_size=self.args.replay_pool_size, n_epochs=self.args.n_iter, discount=self.args.discount, scale_reward=0.01, qf_learning_rate=self.args.qfunc_lr, policy_learning_rate=self.args.policy_lr, eval_samples=self.args.eval_samples, mode=self.args.control) return algo
def main(): now = datetime.datetime.now(dateutil.tz.tzlocal()) rand_id = str(uuid.uuid4())[:5] timestamp = now.strftime('%Y_%m_%d_%H_%M_%S_%f_%Z') default_exp_name = 'experiment_%s_%s' % (timestamp, rand_id) parser = argparse.ArgumentParser() parser.add_argument('--exp_name', type=str, default=default_exp_name, help='Name of the experiment.') parser.add_argument('--discount', type=float, default=0.95) parser.add_argument('--gae_lambda', type=float, default=0.99) parser.add_argument('--reward_scale', type=float, default=1.0) parser.add_argument('--enable_obsnorm', action='store_true', default=False) parser.add_argument('--chunked', action='store_true', default=False) parser.add_argument('--n_iter', type=int, default=250) parser.add_argument('--sampler_workers', type=int, default=1) parser.add_argument('--max_traj_len', type=int, default=250) parser.add_argument('--update_curriculum', action='store_true', default=False) parser.add_argument('--anneal_step_size', type=int, default=0) parser.add_argument('--n_timesteps', type=int, default=8000) parser.add_argument('--control', type=str, default='centralized') parser.add_argument('--buffer_size', type=int, default=1) parser.add_argument('--radius', type=float, default=0.015) parser.add_argument('--n_evaders', type=int, default=10) parser.add_argument('--n_pursuers', type=int, default=8) parser.add_argument('--n_poison', type=int, default=10) parser.add_argument('--n_coop', type=int, default=4) parser.add_argument('--n_sensors', type=int, default=30) parser.add_argument('--sensor_range', type=str, default='0.2') parser.add_argument('--food_reward', type=float, default=5) parser.add_argument('--poison_reward', type=float, default=-1) parser.add_argument('--encounter_reward', type=float, default=0.05) parser.add_argument('--reward_mech', type=str, default='local') parser.add_argument('--recurrent', type=str, default=None) parser.add_argument('--baseline_type', type=str, default='linear') parser.add_argument('--policy_hidden_sizes', type=str, default='128,128') parser.add_argument('--baseline_hidden_sizes', type=str, default='128,128') parser.add_argument('--max_kl', type=float, default=0.01) parser.add_argument('--log_dir', type=str, required=False) parser.add_argument('--tabular_log_file', type=str, default='progress.csv', help='Name of the tabular log file (in csv).') parser.add_argument('--text_log_file', type=str, default='debug.log', help='Name of the text log file (in pure text).') parser.add_argument('--params_log_file', type=str, default='params.json', help='Name of the parameter log file (in json).') parser.add_argument('--seed', type=int, help='Random seed for numpy') parser.add_argument('--args_data', type=str, help='Pickled data for stub objects') parser.add_argument('--snapshot_mode', type=str, default='all', help='Mode to save the snapshot. Can be either "all" ' '(all iterations will be saved), "last" (only ' 'the last iteration will be saved), or "none" ' '(do not save snapshots)') parser.add_argument( '--log_tabular_only', type=ast.literal_eval, default=False, help= 'Whether to only print the tabular log information (in a horizontal format)' ) args = parser.parse_args() parallel_sampler.initialize(n_parallel=args.sampler_workers) if args.seed is not None: set_seed(args.seed) parallel_sampler.set_seed(args.seed) args.hidden_sizes = tuple(map(int, args.policy_hidden_sizes.split(','))) centralized = True if args.control == 'centralized' else False sensor_range = np.array(map(float, args.sensor_range.split(','))) if len(sensor_range) == 1: sensor_range = sensor_range[0] else: assert sensor_range.shape == (args.n_pursuers, ) env = MAWaterWorld(args.n_pursuers, args.n_evaders, args.n_coop, args.n_poison, radius=args.radius, n_sensors=args.n_sensors, food_reward=args.food_reward, poison_reward=args.poison_reward, encounter_reward=args.encounter_reward, reward_mech=args.reward_mech, sensor_range=sensor_range, obstacle_loc=None) env = TfEnv( RLLabEnv(StandardizedEnv(env, scale_reward=args.reward_scale, enable_obsnorm=args.enable_obsnorm), mode=args.control)) if args.buffer_size > 1: env = ObservationBuffer(env, args.buffer_size) if args.recurrent: feature_network = MLP( name='feature_net', input_shape=(env.spec.observation_space.flat_dim + env.spec.action_space.flat_dim, ), output_dim=16, hidden_sizes=(128, 64, 32), hidden_nonlinearity=tf.nn.tanh, output_nonlinearity=None) if args.recurrent == 'gru': policy = GaussianGRUPolicy(env_spec=env.spec, feature_network=feature_network, hidden_dim=int( args.policy_hidden_sizes), name='policy') elif args.recurrent == 'lstm': policy = GaussianLSTMPolicy(env_spec=env.spec, feature_network=feature_network, hidden_dim=int( args.policy_hidden_sizes), name='policy') else: policy = GaussianMLPPolicy( name='policy', env_spec=env.spec, hidden_sizes=tuple(map(int, args.policy_hidden_sizes.split(','))), min_std=10e-5) if args.baseline_type == 'linear': baseline = LinearFeatureBaseline(env_spec=env.spec) elif args.baseline_type == 'mlp': raise NotImplementedError() # baseline = GaussianMLPBaseline( # env_spec=env.spec, hidden_sizes=tuple(map(int, args.baseline_hidden_sizes.split(',')))) else: baseline = ZeroBaseline(env_spec=env.spec) # logger default_log_dir = config.LOG_DIR if args.log_dir is None: log_dir = osp.join(default_log_dir, args.exp_name) else: log_dir = args.log_dir tabular_log_file = osp.join(log_dir, args.tabular_log_file) text_log_file = osp.join(log_dir, args.text_log_file) params_log_file = osp.join(log_dir, args.params_log_file) logger.log_parameters_lite(params_log_file, args) logger.add_text_output(text_log_file) logger.add_tabular_output(tabular_log_file) prev_snapshot_dir = logger.get_snapshot_dir() prev_mode = logger.get_snapshot_mode() logger.set_snapshot_dir(log_dir) logger.set_snapshot_mode(args.snapshot_mode) logger.set_log_tabular_only(args.log_tabular_only) logger.push_prefix("[%s] " % args.exp_name) algo = TRPO( env=env, policy=policy, baseline=baseline, batch_size=args.n_timesteps, max_path_length=args.max_traj_len, #max_path_length_limit=args.max_path_length_limit, update_max_path_length=args.update_curriculum, anneal_step_size=args.anneal_step_size, n_itr=args.n_iter, discount=args.discount, gae_lambda=args.gae_lambda, step_size=args.max_kl, optimizer=ConjugateGradientOptimizer(hvp_approach=FiniteDifferenceHvp( base_eps=1e-5)) if args.recurrent else None, mode=args.control if not args.chunked else 'chunk_{}'.format(args.control), ) algo.train()
def run_experiment(argv): default_log_dir = config.LOG_DIR now = datetime.datetime.now(dateutil.tz.tzlocal()) # avoid name clashes when running distributed jobs rand_id = str(uuid.uuid4())[:5] timestamp = now.strftime('%Y_%m_%d_%H_%M_%S_%f_%Z') default_exp_name = 'experiment_%s_%s' % (timestamp, rand_id) parser = argparse.ArgumentParser() parser.add_argument( '--n_parallel', type=int, default=1, help= 'Number of parallel workers to perform rollouts. 0 => don\'t start any workers' ) parser.add_argument('--exp_name', type=str, default=default_exp_name, help='Name of the experiment.') parser.add_argument('--log_dir', type=str, default=None, help='Path to save the log and iteration snapshot.') parser.add_argument('--snapshot_mode', type=str, default='all', help='Mode to save the snapshot. Can be either "all" ' '(all iterations will be saved), "last" (only ' 'the last iteration will be saved), "gap" (every' '`snapshot_gap` iterations are saved), or "none" ' '(do not save snapshots)') parser.add_argument('--snapshot_gap', type=int, default=1, help='Gap between snapshot iterations.') parser.add_argument('--tabular_log_file', type=str, default='progress.csv', help='Name of the tabular log file (in csv).') parser.add_argument('--text_log_file', type=str, default='debug.log', help='Name of the text log file (in pure text).') parser.add_argument('--tensorboard_log_dir', type=str, default='tb', help='Name of the folder for tensorboard_summary.') parser.add_argument( '--tensorboard_step_key', type=str, default=None, help= 'Name of the step key in log data which shows the step in tensorboard_summary.' ) parser.add_argument('--params_log_file', type=str, default='params.json', help='Name of the parameter log file (in json).') parser.add_argument('--variant_log_file', type=str, default='variant.json', help='Name of the variant log file (in json).') parser.add_argument( '--resume_from', type=str, default=None, help='Name of the pickle file to resume experiment from.') parser.add_argument('--plot', type=ast.literal_eval, default=False, help='Whether to plot the iteration results') parser.add_argument( '--log_tabular_only', type=ast.literal_eval, default=False, help= 'Whether to only print the tabular log information (in a horizontal format)' ) parser.add_argument('--seed', type=int, help='Random seed for numpy') parser.add_argument('--args_data', type=str, help='Pickled data for stub objects') parser.add_argument('--variant_data', type=str, help='Pickled data for variant configuration') parser.add_argument('--use_cloudpickle', type=ast.literal_eval, default=False) parser.add_argument('--checkpoint_dir', type=str, default='checkpoint', help='Name of the folder for checkpoints.') parser.add_argument('--obs_dir', type=str, default='obs', help='Name of the folder for original observations.') args = parser.parse_args(argv[1:]) if args.seed is not None: set_seed(args.seed) if args.n_parallel > 0: from rllab.sampler import parallel_sampler parallel_sampler.initialize(n_parallel=args.n_parallel) if args.seed is not None: parallel_sampler.set_seed(args.seed) if args.plot: from rllab.plotter import plotter plotter.init_worker() if args.log_dir is None: log_dir = osp.join(default_log_dir, args.exp_name) else: log_dir = args.log_dir tabular_log_file = osp.join(log_dir, args.tabular_log_file) text_log_file = osp.join(log_dir, args.text_log_file) params_log_file = osp.join(log_dir, args.params_log_file) tensorboard_log_dir = osp.join(log_dir, args.tensorboard_log_dir) checkpoint_dir = osp.join(log_dir, args.checkpoint_dir) obs_dir = osp.join(log_dir, args.obs_dir) if args.variant_data is not None: variant_data = pickle.loads(base64.b64decode(args.variant_data)) variant_log_file = osp.join(log_dir, args.variant_log_file) logger.log_variant(variant_log_file, variant_data) else: variant_data = None if not args.use_cloudpickle: logger.log_parameters_lite(params_log_file, args) logger.add_text_output(text_log_file) logger.add_tabular_output(tabular_log_file) logger.set_tensorboard_dir(tensorboard_log_dir) logger.set_checkpoint_dir(checkpoint_dir) logger.set_obs_dir(obs_dir) prev_snapshot_dir = logger.get_snapshot_dir() prev_mode = logger.get_snapshot_mode() logger.set_snapshot_dir(log_dir) logger.set_snapshot_mode(args.snapshot_mode) logger.set_snapshot_gap(args.snapshot_gap) logger.set_log_tabular_only(args.log_tabular_only) logger.set_tensorboard_step_key(args.tensorboard_step_key) logger.push_prefix("[%s] " % args.exp_name) git_commit = get_git_commit_hash() logger.log('Git commit: {}'.format(git_commit)) git_diff_file_path = osp.join(log_dir, 'git_diff_{}.patch'.format(git_commit)) save_git_diff_to_file(git_diff_file_path) logger.log('hostname: {}, pid: {}, tmux session: {}'.format( socket.gethostname(), os.getpid(), get_tmux_session_name())) if args.resume_from is not None: data = joblib.load(args.resume_from) assert 'algo' in data algo = data['algo'] algo.train() else: # read from stdin if args.use_cloudpickle: import cloudpickle method_call = cloudpickle.loads(base64.b64decode(args.args_data)) method_call(variant_data) else: data = pickle.loads(base64.b64decode(args.args_data)) maybe_iter = concretize(data) if is_iterable(maybe_iter): for _ in maybe_iter: pass logger.set_snapshot_mode(prev_mode) logger.set_snapshot_dir(prev_snapshot_dir) logger.remove_tabular_output(tabular_log_file) logger.remove_text_output(text_log_file) logger.pop_prefix()
def main(): now = datetime.datetime.now(dateutil.tz.tzlocal()) rand_id = str(uuid.uuid4())[:5] timestamp = now.strftime('%Y_%m_%d_%H_%M_%S_%f_%Z') default_exp_name = 'experiment_%s_%s' % (timestamp, rand_id) parser = argparse.ArgumentParser() parser.add_argument('--exp_name', type=str, default=default_exp_name, help='Name of the experiment.') parser.add_argument('--discount', type=float, default=0.99) parser.add_argument('--gae_lambda', type=float, default=1.0) parser.add_argument('--reward_scale', type=float, default=1.0) parser.add_argument('--n_iter', type=int, default=250) parser.add_argument('--sampler_workers', type=int, default=1) parser.add_argument('--max_traj_len', type=int, default=250) parser.add_argument('--update_curriculum', action='store_true', default=False) parser.add_argument('--n_timesteps', type=int, default=8000) parser.add_argument('--control', type=str, default='centralized') parser.add_argument('--rectangle', type=str, default='10,10') parser.add_argument('--map_type', type=str, default='rectangle') parser.add_argument('--n_evaders', type=int, default=5) parser.add_argument('--n_pursuers', type=int, default=2) parser.add_argument('--obs_range', type=int, default=3) parser.add_argument('--n_catch', type=int, default=2) parser.add_argument('--urgency', type=float, default=0.0) parser.add_argument('--pursuit', dest='train_pursuit', action='store_true') parser.add_argument('--evade', dest='train_pursuit', action='store_false') parser.set_defaults(train_pursuit=True) parser.add_argument('--surround', action='store_true', default=False) parser.add_argument('--constraint_window', type=float, default=1.0) parser.add_argument('--sample_maps', action='store_true', default=False) parser.add_argument('--map_file', type=str, default='../maps/map_pool.npy') parser.add_argument('--flatten', action='store_true', default=False) parser.add_argument('--reward_mech', type=str, default='global') parser.add_argument('--catchr', type=float, default=0.1) parser.add_argument('--term_pursuit', type=float, default=5.0) parser.add_argument('--recurrent', type=str, default=None) parser.add_argument('--policy_hidden_sizes', type=str, default='128,128') parser.add_argument('--baselin_hidden_sizes', type=str, default='128,128') parser.add_argument('--baseline_type', type=str, default='linear') parser.add_argument('--conv', action='store_true', default=False) parser.add_argument('--max_kl', type=float, default=0.01) parser.add_argument('--checkpoint', type=str, default=None) parser.add_argument('--log_dir', type=str, required=False) parser.add_argument('--tabular_log_file', type=str, default='progress.csv', help='Name of the tabular log file (in csv).') parser.add_argument('--text_log_file', type=str, default='debug.log', help='Name of the text log file (in pure text).') parser.add_argument('--params_log_file', type=str, default='params.json', help='Name of the parameter log file (in json).') parser.add_argument('--seed', type=int, help='Random seed for numpy') parser.add_argument('--args_data', type=str, help='Pickled data for stub objects') parser.add_argument('--snapshot_mode', type=str, default='all', help='Mode to save the snapshot. Can be either "all" ' '(all iterations will be saved), "last" (only ' 'the last iteration will be saved), or "none" ' '(do not save snapshots)') parser.add_argument( '--log_tabular_only', type=ast.literal_eval, default=False, help= 'Whether to only print the tabular log information (in a horizontal format)' ) args = parser.parse_args() parallel_sampler.initialize(n_parallel=args.sampler_workers) if args.seed is not None: set_seed(args.seed) parallel_sampler.set_seed(args.seed) args.hidden_sizes = tuple(map(int, args.policy_hidden_sizes.split(','))) if args.checkpoint: with tf.Session() as sess: data = joblib.load(args.checkpoint) policy = data['policy'] env = data['env'] else: if args.sample_maps: map_pool = np.load(args.map_file) else: if args.map_type == 'rectangle': env_map = TwoDMaps.rectangle_map( *map(int, args.rectangle.split(','))) elif args.map_type == 'complex': env_map = TwoDMaps.complex_map( *map(int, args.rectangle.split(','))) else: raise NotImplementedError() map_pool = [env_map] env = PursuitEvade(map_pool, n_evaders=args.n_evaders, n_pursuers=args.n_pursuers, obs_range=args.obs_range, n_catch=args.n_catch, train_pursuit=args.train_pursuit, urgency_reward=args.urgency, surround=args.surround, sample_maps=args.sample_maps, constraint_window=args.constraint_window, flatten=args.flatten, reward_mech=args.reward_mech, catchr=args.catchr, term_pursuit=args.term_pursuit) env = TfEnv( RLLabEnv(StandardizedEnv(env, scale_reward=args.reward_scale, enable_obsnorm=False), mode=args.control)) if args.recurrent: if args.conv: feature_network = ConvNetwork( name='feature_net', input_shape=emv.spec.observation_space.shape, output_dim=5, conv_filters=(16, 32, 32), conv_filter_sizes=(3, 3, 3), conv_strides=(1, 1, 1), conv_pads=('VALID', 'VALID', 'VALID'), hidden_sizes=(64, ), hidden_nonlinearity=tf.nn.relu, output_nonlinearity=tf.nn.softmax) else: feature_network = MLP( name='feature_net', input_shape=(env.spec.observation_space.flat_dim + env.spec.action_space.flat_dim, ), output_dim=5, hidden_sizes=(256, 128, 64), hidden_nonlinearity=tf.nn.tanh, output_nonlinearity=None) if args.recurrent == 'gru': policy = CategoricalGRUPolicy(env_spec=env.spec, feature_network=feature_network, hidden_dim=int( args.policy_hidden_sizes), name='policy') elif args.recurrent == 'lstm': policy = CategoricalLSTMPolicy(env_spec=env.spec, feature_network=feature_network, hidden_dim=int( args.policy_hidden_sizes), name='policy') elif args.conv: feature_network = ConvNetwork( name='feature_net', input_shape=env.spec.observation_space.shape, output_dim=5, conv_filters=(8, 16), conv_filter_sizes=(3, 3), conv_strides=(2, 1), conv_pads=('VALID', 'VALID'), hidden_sizes=(32, ), hidden_nonlinearity=tf.nn.relu, output_nonlinearity=tf.nn.softmax) policy = CategoricalMLPPolicy(name='policy', env_spec=env.spec, prob_network=feature_network) else: policy = CategoricalMLPPolicy(name='policy', env_spec=env.spec, hidden_sizes=args.hidden_sizes) if args.baseline_type == 'linear': baseline = LinearFeatureBaseline(env_spec=env.spec) else: baseline = ZeroBaseline(env_spec=env.spec) # logger default_log_dir = config.LOG_DIR if args.log_dir is None: log_dir = osp.join(default_log_dir, args.exp_name) else: log_dir = args.log_dir tabular_log_file = osp.join(log_dir, args.tabular_log_file) text_log_file = osp.join(log_dir, args.text_log_file) params_log_file = osp.join(log_dir, args.params_log_file) logger.log_parameters_lite(params_log_file, args) logger.add_text_output(text_log_file) logger.add_tabular_output(tabular_log_file) prev_snapshot_dir = logger.get_snapshot_dir() prev_mode = logger.get_snapshot_mode() logger.set_snapshot_dir(log_dir) logger.set_snapshot_mode(args.snapshot_mode) logger.set_log_tabular_only(args.log_tabular_only) logger.push_prefix("[%s] " % args.exp_name) algo = TRPO( env=env, policy=policy, baseline=baseline, batch_size=args.n_timesteps, max_path_length=args.max_traj_len, n_itr=args.n_iter, discount=args.discount, gae_lambda=args.gae_lambda, step_size=args.max_kl, optimizer=ConjugateGradientOptimizer(hvp_approach=FiniteDifferenceHvp( base_eps=1e-5)) if args.recurrent else None, mode=args.control, ) algo.train()
def setup(self, env, policy, start_itr): # Baseline if self.args.baseline_type == 'linear': baseline = LinearFeatureBaseline(env_spec=env.spec) elif self.args.baseline_type == 'zero': baseline = ZeroBaseline(env_spec=env.spec) else: raise NotImplementedError(self.args.baseline_type) # Logger default_log_dir = config.LOG_DIR if self.args.log_dir is None: log_dir = osp.join(default_log_dir, self.args.exp_name) else: log_dir = self.args.log_dir tabular_log_file = osp.join(log_dir, self.args.tabular_log_file) text_log_file = osp.join(log_dir, self.args.text_log_file) params_log_file = osp.join(log_dir, self.args.params_log_file) logger.log_parameters_lite(params_log_file, self.args) logger.add_text_output(text_log_file) logger.add_tabular_output(tabular_log_file) # prev_snapshot_dir = logger.get_snapshot_dir() # prev_mode = logger.get_snapshot_mode() logger.set_snapshot_dir(log_dir) logger.set_snapshot_mode(self.args.snapshot_mode) logger.set_log_tabular_only(self.args.log_tabular_only) logger.push_prefix("[%s] " % self.args.exp_name) if self.args.algo == 'reinforce': algo = MAReinforce(env=env, policy_or_policies=policy, plot=False, baseline_or_baselines=baseline, batch_size=self.args.batch_size, pause_for_plot=True, start_itr=start_itr, max_path_length=self.args.max_path_length, n_itr=self.args.n_iter, discount=self.args.discount, gae_lambda=self.args.gae_lambda, step_size=self.args.step_size, ma_mode=self.args.control, save_param_update=self.args.save_param_update) elif self.args.algo == 'dqn': algo = MADQN(env=env, networks=policy, plot=False, batch_size=self.args.batch_size, pause_for_plot=True, start_itr=start_itr, max_path_length=self.args.max_path_length, n_itr=self.args.n_iter, discount=self.args.discount, ma_mode=self.args.control, pre_trained_size=self.args.replay_pre_trained_size, target_network_update=self.args.target_network_update, save_param_update=self.args.save_param_update) elif self.args.algo == 'a2c': algo = MAA2C(env=env, policy_or_policies=policy, plot=False, baseline_or_baselines=baseline, batch_size=self.args.batch_size, pause_for_plot=True, start_itr=start_itr, max_path_length=self.args.max_path_length, n_itr=self.args.n_iter, discount=self.args.discount, ma_mode=self.args.control, actor_learning_rate=self.args.policy_lr, critic_learning_rate=self.args.qfunc_lr, value_coefficient=0.5, entropy_coefficient=0.01, clip_grads=0.5, save_param_update=self.args.save_param_update) return algo
def main(): now = datetime.datetime.now(dateutil.tz.tzlocal()) rand_id = str(uuid.uuid4())[:5] timestamp = now.strftime('%Y_%m_%d_%H_%M_%S_%f_%Z') default_exp_name = 'experiment_%s_%s' % (timestamp, rand_id) parser = argparse.ArgumentParser() parser.add_argument('--exp_name', type=str, default=default_exp_name, help='Name of the experiment.') parser.add_argument('--discount', type=float, default=0.95) parser.add_argument('--gae_lambda', type=float, default=0.99) parser.add_argument('--n_iter', type=int, default=250) parser.add_argument('--sampler_workers', type=int, default=1) parser.add_argument('--max_traj_len', type=int, default=250) parser.add_argument('--update_curriculum', action='store_true', default=False) parser.add_argument('--n_timesteps', type=int, default=8000) parser.add_argument('--control', type=str, default='centralized') parser.add_argument('--control', type=str, default='centralized') parser.add_argument('--buffer_size', type=int, default=1) parser.add_argument('--n_good', type=int, default=3) parser.add_argument('--n_hostage', type=int, default=5) parser.add_argument('--n_bad', type=int, default=5) parser.add_argument('--n_coop_save', type=int, default=2) parser.add_argument('--n_coop_avoid', type=int, default=2) parser.add_argument('--n_sensors', type=int, default=20) parser.add_argument('--sensor_range', type=float, default=0.2) parser.add_argument('--save_reward', type=float, default=3) parser.add_argument('--hit_reward', type=float, default=-1) parser.add_argument('--encounter_reward', type=float, default=0.01) parser.add_argument('--bomb_reward', type=float, default=-10.) parser.add_argument('--recurrent', action='store_true', default=False) parser.add_argument('--baseline_type', type=str, default='linear') parser.add_argument('--policy_hidden_sizes', type=str, default='128,128') parser.add_argument('--baselin_hidden_sizes', type=str, default='128,128') parser.add_argument('--max_kl', type=float, default=0.01) parser.add_argument('--log_dir', type=str, required=False) parser.add_argument('--tabular_log_file', type=str, default='progress.csv', help='Name of the tabular log file (in csv).') parser.add_argument('--text_log_file', type=str, default='debug.log', help='Name of the text log file (in pure text).') parser.add_argument('--params_log_file', type=str, default='params.json', help='Name of the parameter log file (in json).') parser.add_argument('--seed', type=int, help='Random seed for numpy') parser.add_argument('--args_data', type=str, help='Pickled data for stub objects') parser.add_argument('--snapshot_mode', type=str, default='all', help='Mode to save the snapshot. Can be either "all" ' '(all iterations will be saved), "last" (only ' 'the last iteration will be saved), or "none" ' '(do not save snapshots)') parser.add_argument( '--log_tabular_only', type=ast.literal_eval, default=False, help= 'Whether to only print the tabular log information (in a horizontal format)' ) args = parser.parse_args() parallel_sampler.initialize(n_parallel=args.sampler_workers) if args.seed is not None: set_seed(args.seed) parallel_sampler.set_seed(args.seed) args.hidden_sizes = tuple(map(int, args.policy_hidden_sizes.split(','))) centralized = True if args.control == 'centralized' else False sensor_range = np.array(map(float, args.sensor_range.split(','))) assert sensor_range.shape == (args.n_pursuers, ) env = ContinuousHostageWorld(args.n_good, args.n_hostage, args.n_bad, args.n_coop_save, args.n_coop_avoid, n_sensors=args.n_sensors, sensor_range=args.sensor_range, save_reward=args.save_reward, hit_reward=args.hit_reward, encounter_reward=args.encounter_reward, bomb_reward=args.bomb_reward) env = RLLabEnv(StandardizedEnv(env), mode=args.control) if args.buffer_size > 1: env = ObservationBuffer(env, args.buffer_size) if args.recurrent: policy = GaussianGRUPolicy(env_spec=env.spec, hidden_sizes=args.hidden_sizes) else: policy = GaussianMLPPolicy(env_spec=env.spec, hidden_sizes=args.hidden_sizes) if args.baseline_type == 'linear': baseline = LinearFeatureBaseline(env_spec=env.spec) else: baseline = ZeroBaseline(obsfeat_space) # logger default_log_dir = config.LOG_DIR if args.log_dir is None: log_dir = osp.join(default_log_dir, args.exp_name) else: log_dir = args.log_dir tabular_log_file = osp.join(log_dir, args.tabular_log_file) text_log_file = osp.join(log_dir, args.text_log_file) params_log_file = osp.join(log_dir, args.params_log_file) logger.log_parameters_lite(params_log_file, args) logger.add_text_output(text_log_file) logger.add_tabular_output(tabular_log_file) prev_snapshot_dir = logger.get_snapshot_dir() prev_mode = logger.get_snapshot_mode() logger.set_snapshot_dir(log_dir) logger.set_snapshot_mode(args.snapshot_mode) logger.set_log_tabular_only(args.log_tabular_only) logger.push_prefix("[%s] " % args.exp_name) algo = TRPO( env=env, policy=policy, baseline=baseline, batch_size=args.n_timesteps, max_path_length=args.max_traj_len, n_itr=args.n_iter, discount=args.discount, step_size=args.max_kl, mode=args.control, ) algo.train()
def run_experiment(argv): # e2crawfo: These imports, in this order, were necessary for fixing issues on cedar. import rllab.mujoco_py.mjlib import tensorflow default_log_dir = config.LOG_DIR now = datetime.datetime.now(dateutil.tz.tzlocal()) # avoid name clashes when running distributed jobs rand_id = str(uuid.uuid4())[:5] timestamp = now.strftime('%Y_%m_%d_%H_%M_%S_%f_%Z') default_exp_name = 'experiment_%s_%s' % (timestamp, rand_id) parser = argparse.ArgumentParser() parser.add_argument( '--n_parallel', type=int, default=1, help= 'Number of parallel workers to perform rollouts. 0 => don\'t start any workers' ) parser.add_argument('--exp_name', type=str, default=default_exp_name, help='Name of the experiment.') parser.add_argument('--log_dir', type=str, default=None, help='Path to save the log and iteration snapshot.') parser.add_argument('--snapshot_mode', type=str, default='all', help='Mode to save the snapshot. Can be either "all" ' '(all iterations will be saved), "last" (only ' 'the last iteration will be saved), or "none" ' '(do not save snapshots)') parser.add_argument('--snapshot_gap', type=int, default=1, help='Gap between snapshot iterations.') parser.add_argument('--tabular_log_file', type=str, default='progress.csv', help='Name of the tabular log file (in csv).') parser.add_argument('--text_log_file', type=str, default='debug.log', help='Name of the text log file (in pure text).') parser.add_argument('--params_log_file', type=str, default='params.json', help='Name of the parameter log file (in json).') parser.add_argument('--variant_log_file', type=str, default='variant.json', help='Name of the variant log file (in json).') parser.add_argument( '--resume_from', type=str, default=None, help='Name of the pickle file to resume experiment from.') parser.add_argument('--plot', type=ast.literal_eval, default=False, help='Whether to plot the iteration results') parser.add_argument( '--log_tabular_only', type=ast.literal_eval, default=False, help= 'Whether to only print the tabular log information (in a horizontal format)' ) parser.add_argument('--seed', type=int, help='Random seed for numpy') parser.add_argument('--args_data', type=str, help='Pickled data for stub objects') parser.add_argument('--variant_data', type=str, help='Pickled data for variant configuration') parser.add_argument('--use_cloudpickle', type=ast.literal_eval, default=False) args = parser.parse_args(argv[1:]) if args.seed is not None: set_seed(args.seed) if args.n_parallel > 0: from rllab.sampler import parallel_sampler parallel_sampler.initialize(n_parallel=args.n_parallel) if args.seed is not None: parallel_sampler.set_seed(args.seed) if args.plot: from rllab.plotter import plotter plotter.init_worker() if args.log_dir is None: log_dir = osp.join(default_log_dir, args.exp_name) else: log_dir = args.log_dir tabular_log_file = osp.join(log_dir, args.tabular_log_file) text_log_file = osp.join(log_dir, args.text_log_file) params_log_file = osp.join(log_dir, args.params_log_file) if args.variant_data is not None: variant_data = pickle.loads(base64.b64decode(args.variant_data)) variant_log_file = osp.join(log_dir, args.variant_log_file) logger.log_variant(variant_log_file, variant_data) else: variant_data = None if not args.use_cloudpickle: logger.log_parameters_lite(params_log_file, args) logger.add_text_output(text_log_file) logger.add_tabular_output(tabular_log_file) prev_snapshot_dir = logger.get_snapshot_dir() prev_mode = logger.get_snapshot_mode() logger.set_snapshot_dir(log_dir) logger.set_tf_summary_dir(osp.join(log_dir, "tf_summary")) logger.set_snapshot_mode(args.snapshot_mode) logger.set_snapshot_gap(args.snapshot_gap) logger.set_log_tabular_only(args.log_tabular_only) logger.push_prefix("[%s] " % args.exp_name) if args.resume_from is not None: data = joblib.load(args.resume_from) assert 'algo' in data algo = data['algo'] maybe_iter = algo.train() if is_iterable(maybe_iter): for _ in maybe_iter: pass else: # read from stdin if args.use_cloudpickle: import cloudpickle method_call = cloudpickle.loads(base64.b64decode(args.args_data)) method_call(variant_data) else: data = pickle.loads(base64.b64decode(args.args_data)) maybe_iter = concretize(data) if is_iterable(maybe_iter): for _ in maybe_iter: pass logger.set_snapshot_mode(prev_mode) logger.set_snapshot_dir(prev_snapshot_dir) logger.remove_tabular_output(tabular_log_file) logger.remove_text_output(text_log_file) logger.pop_prefix()
def run_experiment(argv): default_log_dir = config.LOG_DIR now = datetime.datetime.now(dateutil.tz.tzlocal()) # avoid name clashes when running distributed jobs rand_id = str(uuid.uuid4())[:5] timestamp = now.strftime('%Y_%m_%d_%H_%M_%S_%f_%Z') default_exp_name = 'experiment_%s_%s' % (timestamp, rand_id) parser = argparse.ArgumentParser() parser.add_argument( '--n_parallel', type=int, default=1, help= 'Number of parallel workers to perform rollouts. 0 => don\'t start any workers' ) parser.add_argument('--exp_name', type=str, default=default_exp_name, help='Name of the experiment.') parser.add_argument('--log_dir', type=str, default=None, help='Path to save the log and iteration snapshot.') parser.add_argument('--snapshot_mode', type=str, default='all', help='Mode to save the snapshot. Can be either "all" ' '(all iterations will be saved), "last" (only ' 'the last iteration will be saved), "gap" (every' '`snapshot_gap` iterations are saved), or "none" ' '(do not save snapshots)') parser.add_argument('--snapshot_gap', type=int, default=1, help='Gap between snapshot iterations.') parser.add_argument('--tabular_log_file', type=str, default='progress.csv', help='Name of the tabular log file (in csv).') parser.add_argument('--text_log_file', type=str, default='debug.log', help='Name of the text log file (in pure text).') parser.add_argument('--params_log_file', type=str, default='params.json', help='Name of the parameter log file (in json).') parser.add_argument('--variant_log_file', type=str, default='variant.json', help='Name of the variant log file (in json).') parser.add_argument( '--resume_from', type=str, default=None, help='Name of the pickle file to resume experiment from.') parser.add_argument('--plot', type=ast.literal_eval, default=False, help='Whether to plot the iteration results') parser.add_argument( '--log_tabular_only', type=ast.literal_eval, default=False, help= 'Whether to only print the tabular log information (in a horizontal format)' ) parser.add_argument('--seed', type=int, help='Random seed for numpy') parser.add_argument('--args_data', type=str, help='Pickled data for stub objects') parser.add_argument('--variant_data', type=str, help='Pickled data for variant configuration') parser.add_argument('--use_cloudpickle', type=ast.literal_eval, default=False) args = parser.parse_args(argv[1:]) if args.seed is not None: set_seed(args.seed) if args.n_parallel > 0: from rllab.sampler import parallel_sampler parallel_sampler.initialize(n_parallel=args.n_parallel) if args.seed is not None: parallel_sampler.set_seed(args.seed) if args.plot: from rllab.plotter import plotter plotter.init_worker() if args.log_dir is None: log_dir = osp.join(default_log_dir, args.exp_name) else: log_dir = args.log_dir tabular_log_file = osp.join(log_dir, args.tabular_log_file) text_log_file = osp.join(log_dir, args.text_log_file) params_log_file = osp.join(log_dir, args.params_log_file) if args.variant_data is not None: variant_data = pickle.loads(base64.b64decode(args.variant_data)) variant_log_file = osp.join(log_dir, args.variant_log_file) logger.log_variant(variant_log_file, variant_data) else: variant_data = None if not args.use_cloudpickle: logger.log_parameters_lite(params_log_file, args) logger.add_text_output(text_log_file) logger.add_tabular_output(tabular_log_file) prev_snapshot_dir = logger.get_snapshot_dir() prev_mode = logger.get_snapshot_mode() logger.set_snapshot_dir(log_dir) logger.set_snapshot_mode(args.snapshot_mode) logger.set_snapshot_gap(args.snapshot_gap) logger.set_log_tabular_only(args.log_tabular_only) logger.push_prefix("[%s] " % args.exp_name) #variant_data is the variant dictionary sent from trpoTests_ExpLite if (args.resume_from is not None) and ( '&|&' in args.resume_from ): #separate string on &|& to get iters and file location vals = args.resume_from.split( '&|&') #dirRes | numItrs to go | new batchSize dirRes = vals[0] numItrs = int(vals[1]) if (len(vals) > 2): batchSize = int(vals[2]) print("resuming from :{}".format(dirRes)) data = joblib.load(dirRes) #data is dict : 'baseline', 'algo', 'itr', 'policy', 'env' assert 'algo' in data algo = data['algo'] assert 'policy' in data pol = data['policy'] bl = data['baseline'] oldBatchSize = algo.batch_size algo.n_itr = numItrs if (len(vals) > 2): algo.batch_size = batchSize print( 'algo iters : {} cur iter :{} oldBatchSize : {} newBatchSize : {}' .format(algo.n_itr, algo.current_itr, oldBatchSize, algo.batch_size)) else: print('algo iters : {} cur iter :{} '.format( algo.n_itr, algo.current_itr)) algo.train() else: print('Not resuming - building new exp') # read from stdin if args.use_cloudpickle: #set to use cloudpickle import cloudpickle method_call = cloudpickle.loads(base64.b64decode(args.args_data)) method_call(variant_data) else: print('not use cloud pickle') data = pickle.loads(base64.b64decode(args.args_data)) maybe_iter = concretize(data) if is_iterable(maybe_iter): for _ in maybe_iter: pass logger.set_snapshot_mode(prev_mode) logger.set_snapshot_dir(prev_snapshot_dir) logger.remove_tabular_output(tabular_log_file) logger.remove_text_output(text_log_file) logger.pop_prefix()
def run_experiment(argv): default_log_dir = config.LOG_DIR now = datetime.datetime.now(dateutil.tz.tzlocal()) # avoid name clashes when running distributed jobs rand_id = str(uuid.uuid4())[:5] timestamp = now.strftime('%Y_%m_%d_%H_%M_%S_%f_%Z') default_exp_name = 'experiment_%s_%s' % (timestamp, rand_id) parser = argparse.ArgumentParser() parser.add_argument('--n_parallel', type=int, default=1, help='Number of parallel workers to perform rollouts.') parser.add_argument('--exp_name', type=str, default=default_exp_name, help='Name of the experiment.') parser.add_argument('--log_dir', type=str, default=default_log_dir, help='Path to save the log and iteration snapshot.') parser.add_argument('--snapshot_mode', type=str, default='all', help='Mode to save the snapshot. Can be either "all" ' '(all iterations will be saved), "last" (only ' 'the last iteration will be saved), or "none" ' '(do not save snapshots)') parser.add_argument('--snapshot_gap', type=int, default=1, help='Gap between snapshot iterations.') parser.add_argument('--tabular_log_file', type=str, default='progress.csv', help='Name of the tabular log file (in csv).') parser.add_argument('--text_log_file', type=str, default='debug.log', help='Name of the text log file (in pure text).') parser.add_argument('--params_log_file', type=str, default='params.json', help='Name of the parameter log file (in json).') parser.add_argument('--plot', type=ast.literal_eval, default=False, help='Whether to plot the iteration results') parser.add_argument( '--log_tabular_only', type=ast.literal_eval, default=False, help= 'Whether to only print the tabular log information (in a horizontal format)' ) parser.add_argument('--seed', type=int, help='Random seed for numpy') parser.add_argument('--args_data', type=str, help='Pickled data for stub objects') parser.add_argument('--use_cloudpickle', type=ast.literal_eval, default=False, help='Whether to plot the iteration results') args = parser.parse_args(argv[1:]) if args.seed is not None: set_seed(args.seed) if args.n_parallel > 0: from sandbox.vase.sampler import parallel_sampler_expl as parallel_sampler parallel_sampler.initialize(n_parallel=args.n_parallel) if args.seed is not None: set_seed(args.seed) parallel_sampler.set_seed(args.seed) if args.plot: from rllab.plotter import plotter plotter.init_worker() # read from stdin data = pickle.loads(base64.b64decode(args.args_data)) log_dir = args.log_dir # exp_dir = osp.join(log_dir, args.exp_name) tabular_log_file = osp.join(log_dir, args.tabular_log_file) text_log_file = osp.join(log_dir, args.text_log_file) params_log_file = osp.join(log_dir, args.params_log_file) logger.log_parameters_lite(params_log_file, args) logger.add_text_output(text_log_file) logger.add_tabular_output(tabular_log_file) prev_snapshot_dir = logger.get_snapshot_dir() prev_mode = logger.get_snapshot_mode() logger.set_snapshot_gap(args.snapshot_gap) logger.set_snapshot_dir(log_dir) logger.set_snapshot_mode(args.snapshot_mode) logger.set_log_tabular_only(args.log_tabular_only) logger.push_prefix("[%s] " % args.exp_name) maybe_iter = concretize(data) if is_iterable(maybe_iter): for _ in maybe_iter: pass logger.set_snapshot_mode(prev_mode) logger.set_snapshot_dir(prev_snapshot_dir) logger.remove_tabular_output(tabular_log_file) logger.remove_text_output(text_log_file) logger.pop_prefix()
def main(): now = datetime.datetime.now(dateutil.tz.tzlocal()) rand_id = str(uuid.uuid4())[:5] timestamp = now.strftime('%Y_%m_%d_%H_%M_%S_%f_%Z') default_exp_name = 'experiment_%s_%s' % (timestamp, rand_id) parser = argparse.ArgumentParser() parser.add_argument( '--exp_name', type=str, default=default_exp_name, help='Name of the experiment.') parser.add_argument('--discount', type=float, default=0.99) parser.add_argument('--gae_lambda', type=float, default=1.0) parser.add_argument('--reward_scale', type=float, default=1.0) parser.add_argument('--n_iter', type=int, default=250) parser.add_argument('--sampler_workers', type=int, default=1) parser.add_argument('--max_traj_len', type=int, default=250) parser.add_argument('--update_curriculum', action='store_true', default=False) parser.add_argument('--n_timesteps', type=int, default=8000) parser.add_argument('--control', type=str, default='centralized') parser.add_argument('--rectangle', type=str, default='10,10') parser.add_argument('--map_type', type=str, default='rectangle') parser.add_argument('--n_evaders', type=int, default=5) parser.add_argument('--n_pursuers', type=int, default=2) parser.add_argument('--obs_range', type=int, default=3) parser.add_argument('--n_catch', type=int, default=2) parser.add_argument('--urgency', type=float, default=0.0) parser.add_argument('--pursuit', dest='train_pursuit', action='store_true') parser.add_argument('--evade', dest='train_pursuit', action='store_false') parser.set_defaults(train_pursuit=True) parser.add_argument('--surround', action='store_true', default=False) parser.add_argument('--constraint_window', type=float, default=1.0) parser.add_argument('--sample_maps', action='store_true', default=False) parser.add_argument('--map_file', type=str, default='../maps/map_pool.npy') parser.add_argument('--flatten', action='store_true', default=False) parser.add_argument('--reward_mech', type=str, default='global') parser.add_argument('--catchr', type=float, default=0.1) parser.add_argument('--term_pursuit', type=float, default=5.0) parser.add_argument('--recurrent', type=str, default=None) parser.add_argument('--policy_hidden_sizes', type=str, default='128,128') parser.add_argument('--baselin_hidden_sizes', type=str, default='128,128') parser.add_argument('--baseline_type', type=str, default='linear') parser.add_argument('--conv', action='store_true', default=False) parser.add_argument('--max_kl', type=float, default=0.01) parser.add_argument('--log_dir', type=str, required=False) parser.add_argument('--tabular_log_file', type=str, default='progress.csv', help='Name of the tabular log file (in csv).') parser.add_argument('--text_log_file', type=str, default='debug.log', help='Name of the text log file (in pure text).') parser.add_argument('--params_log_file', type=str, default='params.json', help='Name of the parameter log file (in json).') parser.add_argument('--seed', type=int, help='Random seed for numpy') parser.add_argument('--args_data', type=str, help='Pickled data for stub objects') parser.add_argument('--snapshot_mode', type=str, default='all', help='Mode to save the snapshot. Can be either "all" ' '(all iterations will be saved), "last" (only ' 'the last iteration will be saved), or "none" ' '(do not save snapshots)') parser.add_argument('--log_tabular_only', type=ast.literal_eval, default=False, help='Whether to only print the tabular log information (in a horizontal format)') args = parser.parse_args() parallel_sampler.initialize(n_parallel=args.sampler_workers) if args.seed is not None: set_seed(args.seed) parallel_sampler.set_seed(args.seed) args.hidden_sizes = tuple(map(int, args.policy_hidden_sizes.split(','))) if args.sample_maps: map_pool = np.load(args.map_file) else: if args.map_type == 'rectangle': env_map = TwoDMaps.rectangle_map(*map(int, args.rectangle.split(','))) elif args.map_type == 'complex': env_map = TwoDMaps.complex_map(*map(int, args.rectangle.split(','))) else: raise NotImplementedError() map_pool = [env_map] env = PursuitEvade(map_pool, n_evaders=args.n_evaders, n_pursuers=args.n_pursuers, obs_range=args.obs_range, n_catch=args.n_catch, train_pursuit=args.train_pursuit, urgency_reward=args.urgency, surround=args.surround, sample_maps=args.sample_maps, constraint_window=args.constraint_window, flatten=args.flatten, reward_mech=args.reward_mech, catchr=args.catchr, term_pursuit=args.term_pursuit) env = RLLabEnv( StandardizedEnv(env, scale_reward=args.reward_scale, enable_obsnorm=False), mode=args.control) if args.recurrent: if args.conv: feature_network = ConvNetwork( input_shape=emv.spec.observation_space.shape, output_dim=5, conv_filters=(8,16,16), conv_filter_sizes=(3,3,3), conv_strides=(1,1,1), conv_pads=('VALID','VALID','VALID'), hidden_sizes=(64,), hidden_nonlinearity=NL.rectify, output_nonlinearity=NL.softmax) else: feature_network = MLP( input_shape=(env.spec.observation_space.flat_dim + env.spec.action_space.flat_dim,), output_dim=5, hidden_sizes=(128,128,128), hidden_nonlinearity=NL.tanh, output_nonlinearity=None) if args.recurrent == 'gru': policy = CategoricalGRUPolicy(env_spec=env.spec, feature_network=feature_network, hidden_dim=int(args.policy_hidden_sizes)) elif args.conv: feature_network = ConvNetwork( input_shape=env.spec.observation_space.shape, output_dim=5, conv_filters=(8,16,16), conv_filter_sizes=(3,3,3), conv_strides=(1,1,1), conv_pads=('valid','valid','valid'), hidden_sizes=(64,), hidden_nonlinearity=NL.rectify, output_nonlinearity=NL.softmax) policy = CategoricalMLPPolicy(env_spec=env.spec, prob_network=feature_network) else: policy = CategoricalMLPPolicy(env_spec=env.spec, hidden_sizes=args.hidden_sizes) if args.baseline_type == 'linear': baseline = LinearFeatureBaseline(env_spec=env.spec) else: baseline = ZeroBaseline(obsfeat_space) # logger default_log_dir = config.LOG_DIR if args.log_dir is None: log_dir = osp.join(default_log_dir, args.exp_name) else: log_dir = args.log_dir tabular_log_file = osp.join(log_dir, args.tabular_log_file) text_log_file = osp.join(log_dir, args.text_log_file) params_log_file = osp.join(log_dir, args.params_log_file) logger.log_parameters_lite(params_log_file, args) logger.add_text_output(text_log_file) logger.add_tabular_output(tabular_log_file) prev_snapshot_dir = logger.get_snapshot_dir() prev_mode = logger.get_snapshot_mode() logger.set_snapshot_dir(log_dir) logger.set_snapshot_mode(args.snapshot_mode) logger.set_log_tabular_only(args.log_tabular_only) logger.push_prefix("[%s] " % args.exp_name) algo = TRPO( env=env, policy=policy, baseline=baseline, batch_size=args.n_timesteps, max_path_length=args.max_traj_len, n_itr=args.n_iter, discount=args.discount, gae_lambda=args.gae_lambda, step_size=args.max_kl, mode=args.control,) algo.train()