def _build(self): variant = copy.deepcopy(self._variant) env = self.env = get_environment_from_variant(variant) replay_pool = self.replay_pool = (get_replay_pool_from_variant( variant, env)) sampler = self.sampler = get_sampler_from_variant(variant) Qs = self.Qs = get_Q_function_from_variant(variant, env) policy = self.policy = get_policy_from_variant(variant, env, Qs) initial_exploration_policy = self.initial_exploration_policy = ( get_policy('UniformPolicy', env)) self.algorithm = get_algorithm_from_variant( variant=variant, env=env, policy=policy, initial_exploration_policy=initial_exploration_policy, Qs=Qs, pool=replay_pool, sampler=sampler, session=self._session, ) initialize_tf_variables(self._session, only_uninitialized=True) self._built = True
def _build(self): variant = copy.deepcopy(self._variant) print(variant) environment_params = variant['environment_params'] training_environment = self.training_environment = ( get_environment_from_params(environment_params['training'])) evaluation_environment = self.evaluation_environment = ( get_environment_from_params(environment_params['evaluation']) if 'evaluation' in environment_params else training_environment) replay_pool = self.replay_pool = (get_replay_pool_from_variant( variant, training_environment)) sampler = self.sampler = get_sampler_from_variant(variant) Qs = self.Qs = get_Q_function_from_variant(variant, training_environment) policy = self.policy = get_policy_from_variant(variant, training_environment, Qs) initial_exploration_policy = self.initial_exploration_policy = ( get_policy('UniformPolicy', training_environment)) self.algorithm = get_algorithm_from_variant( variant=self._variant, training_environment=training_environment, evaluation_environment=evaluation_environment, policy=policy, initial_exploration_policy=initial_exploration_policy, Qs=Qs, pool=replay_pool, sampler=sampler, session=self._session) initialize_tf_variables(self._session, only_uninitialized=True) self._built = True
def init_policy(): session = tf.keras.backend.get_session() checkpoint_path = CHECKPOINT_PATH.rstrip('/') experiment_path = os.path.dirname(checkpoint_path) variant_path = os.path.join(experiment_path, 'params.pkl') with open(variant_path, 'rb') as f: variant = pickle.load(f) with session.as_default(): pickle_path = os.path.join(checkpoint_path, 'checkpoint.pkl') with open(pickle_path, 'rb') as f: picklable = pickle.load(f) environment_params = (variant['environment_params']['evaluation'] if 'evaluation' in variant['environment_params'] else variant['environment_params']['training']) environment_params['n_parallel_envs'] = 1 evaluation_environment = get_environment_from_params(environment_params) policy = get_policy_from_variant(variant, evaluation_environment) policy.set_weights(picklable['policy_weights']) Qs = get_Q_function_from_variant(variant, evaluation_environment) for i, Q in enumerate(Qs): Qs[i].load_weights(os.path.join(checkpoint_path, 'Qs_{}'.format(i))) return policy, Qs
def _build(self): variant = copy.deepcopy(self._variant) env = self.env = get_environment_from_variant(variant) replay_pool = self.replay_pool = (get_replay_pool_from_variant( variant, env)) sampler = self.sampler = get_sampler_from_variant(variant) Qs = self.Qs = get_Q_function_from_variant(variant, env) #policy = self.policy = get_policy_from_variant(variant, env, Qs) policy = self.policy = get_policy_from_variant(variant, env) initial_exploration_policy = self.initial_exploration_policy = ( get_policy('UniformPolicy', env)) self.algorithm = get_algorithm_from_variant( variant=self._variant, env=self.env, policy=policy, initial_exploration_policy=initial_exploration_policy, Qs=Qs, pool=replay_pool, sampler=sampler, session=self._session) print([ x for x in tf.get_default_graph().get_operations() if x.type == "Placeholder" ]) initialize_tf_variables(self._session, only_uninitialized=True) self._built = True
def run_experiment(variant, reporter): env = get_environment('gym', 'MultiGoal', 'Default', { 'actuation_cost_coeff': 1, 'distance_cost_coeff': 0.1, 'goal_reward': 1, 'init_sigma': 0.1, }) pool = SimpleReplayPool( observation_space=env.observation_space, action_space=env.action_space, max_size=1e6) sampler = SimpleSampler( max_path_length=30, min_pool_size=100, batch_size=64) Qs = get_Q_function_from_variant(variant, env) policy = get_policy_from_variant(variant, env, Qs) plotter = QFPolicyPlotter( Q=Qs[0], policy=policy, obs_lst=np.array(((-2.5, 0.0), (0.0, 0.0), (2.5, 2.5), (-2.5, -2.5))), default_action=(np.nan, np.nan), n_samples=100) algorithm = SAC( sampler=sampler, reparameterize=True, epoch_length=100, n_epochs=1000, n_train_repeat=1, eval_render_mode=None, eval_n_episodes=10, eval_deterministic=False, env=env, policy=policy, initial_exploration_policy=None, pool=pool, Qs=Qs, plotter=plotter, lr=3e-4, target_entropy=-2.0, discount=0.99, tau=1e-4, save_full_state=True, ) initialize_tf_variables(algorithm._session, only_uninitialized=True) for train_result in algorithm.train(): reporter(**train_result)
def _restore(self, checkpoint_dir): assert isinstance(checkpoint_dir, str), checkpoint_dir checkpoint_dir = checkpoint_dir.rstrip('/') with self._session.as_default(): pickle_path = self._pickle_path(checkpoint_dir) with open(pickle_path, 'rb') as f: picklable = pickle.load(f) training_environment = self.training_environment = picklable[ 'training_environment'] evaluation_environment = self.evaluation_environment = picklable[ 'evaluation_environment'] replay_pool = self.replay_pool = (get_replay_pool_from_variant( self._variant, training_environment)) if self._variant['run_params'].get('checkpoint_replay_pool', False): self._restore_replay_pool(checkpoint_dir) sampler = self.sampler = picklable['sampler'] Qs = self.Qs = get_Q_function_from_variant(self._variant, training_environment) self._restore_value_functions(checkpoint_dir) policy = self.policy = (get_policy_from_variant( self._variant, training_environment)) self.policy.set_weights(picklable['policy_weights']) initial_exploration_policy = self.initial_exploration_policy = ( get_policy_from_params(self._variant['exploration_policy_params'], training_environment)) self.algorithm = get_algorithm_from_variant( variant=self._variant, training_environment=training_environment, evaluation_environment=evaluation_environment, policy=policy, initial_exploration_policy=initial_exploration_policy, Qs=Qs, pool=replay_pool, sampler=sampler, session=self._session) self.algorithm.__setstate__(picklable['algorithm'].__getstate__()) tf_checkpoint = self._get_tf_checkpoint() status = tf_checkpoint.restore( tf.train.latest_checkpoint( os.path.split(self._tf_checkpoint_prefix(checkpoint_dir))[0])) status.assert_consumed().run_restore_ops(self._session) initialize_tf_variables(self._session, only_uninitialized=True) # TODO(hartikainen): target Qs should either be checkpointed or pickled. for Q, Q_target in zip(self.algorithm._Qs, self.algorithm._Q_targets): Q_target.set_weights(Q.get_weights()) self._built = True
def _build(self): """ called by tune to build algorithm """ variant = copy.deepcopy(self._variant) environment_params = variant['environment_params'] training_environment = self.training_environment = ( get_environment_from_params(environment_params['training'])) mjc_model_environment = self.mjc_model_environment = ( get_environment_from_params(environment_params['training'])) evaluation_environment = self.evaluation_environment = ( get_environment_from_params(environment_params['evaluation']) if 'evaluation' in environment_params else training_environment) replay_pool = self.replay_pool = (get_replay_pool_from_variant( variant, training_environment)) sampler = self.sampler = get_sampler_from_variant(variant) Qs = self.Qs = get_Q_function_from_variant(variant, training_environment) policy = self.policy = get_policy_from_variant(variant, training_environment, Qs, self._session) initial_exploration_policy = self.initial_exploration_policy = ( get_policy('UniformPolicy', training_environment)) #### get termination function domain = environment_params['training']['domain'] static_fns = mbpo.static[domain.lower()] #### #### build algorithm self.algorithm = get_algorithm_from_variant( variant=self._variant, training_environment=training_environment, evaluation_environment=evaluation_environment, mjc_model_environment=mjc_model_environment, policy=policy, initial_exploration_policy=initial_exploration_policy, Qs=Qs, pool=replay_pool, static_fns=static_fns, sampler=sampler, session=self._session) initialize_tf_variables(self._session, only_uninitialized=True) # add graph since ray doesn't seem to automatically add that graph_writer = tf.summary.FileWriter(self.logdir, self._session.graph) graph_writer.flush() graph_writer.close() #### finalize graph # tf.get_default_graph().finalize() ### good for debugging, but interferes with Qs on SAC self._built = True
def _build(self): variant = copy.deepcopy(self._variant) training_environment = self.training_environment = ( get_goal_example_environment_from_variant(variant)) evaluation_environment = self.evaluation_environment = ( get_goal_example_environment_from_variant(variant)) replay_pool = self.replay_pool = (get_replay_pool_from_variant( variant, training_environment)) sampler = self.sampler = get_sampler_from_variant(variant) # 创建网络 Dense :inputs:[state,action] outputs:size=1 Qs = self.Qs = get_Q_function_from_variant(variant, training_environment) policy = self.policy = get_policy_from_variant(variant, training_environment, Qs) initial_exploration_policy = self.initial_exploration_policy = ( get_policy('UniformPolicy', training_environment)) algorithm_kwargs = { 'variant': self._variant, 'training_environment': self.training_environment, 'evaluation_environment': self.evaluation_environment, 'policy': policy, 'initial_exploration_policy': initial_exploration_policy, 'Qs': Qs, 'pool': replay_pool, 'sampler': sampler, 'session': self._session, } if self._variant['algorithm_params']['type'] in [ 'SACClassifier', 'RAQ', 'VICE', 'VICEGAN', 'VICERAQ' ]: reward_classifier = self.reward_classifier \ = get_reward_classifier_from_variant(self._variant, training_environment) algorithm_kwargs['classifier'] = reward_classifier goal_examples_train, goal_examples_validation = \ get_goal_example_from_variant(variant) algorithm_kwargs['goal_examples'] = goal_examples_train algorithm_kwargs['goal_examples_validation'] = \ goal_examples_validation self.algorithm = get_algorithm_from_variant(**algorithm_kwargs) initialize_tf_variables(self._session, only_uninitialized=True) self._built = True
def run_experiment(variant, reporter): training_environment = ( get_environment('gym', 'MultiGoal', 'Default-v0', { 'actuation_cost_coeff': 30, 'distance_cost_coeff': 1, 'goal_reward': 10, 'init_sigma': 0.1, })) evaluation_environment = training_environment.copy() pool = SimpleReplayPool( environment=training_environment, max_size=1e6) sampler = SimpleSampler(max_path_length=30) Qs = get_Q_function_from_variant(variant, training_environment) policy = get_policy_from_variant(variant, training_environment) plotter = QFPolicyPlotter( Q=Qs[0], policy=policy, obs_lst=np.array(((-2.5, 0.0), (0.0, 0.0), (2.5, 2.5), (-2.5, -2.5))), default_action=(np.nan, np.nan), n_samples=100) algorithm = get_algorithm_from_variant( variant=variant, training_environment=training_environment, evaluation_environment=evaluation_environment, policy=policy, Qs=Qs, pool=pool, sampler=sampler, min_pool_size=100, batch_size=46, plotter=plotter, ) initialize_tf_variables(algorithm._session, only_uninitialized=True) for train_result in algorithm.train(): reporter(**train_result)
def _build(self): variant = copy.deepcopy(self._variant) environment_params = variant['environment_params'] training_environment = self.training_environment = ( get_environment_from_params(environment_params['training'])) evaluation_environment = self.evaluation_environment = ( get_environment_from_params(environment_params['evaluation']) if 'evaluation' in environment_params else training_environment) seed = variant['run_params']['seed'] training_environment.seed(seed) # Set a different seed for the evaluation env # to ensure the policy is not just memorizing action sequences for seen initial states evaluation_environment.seed(seed + 10) replay_pool = self.replay_pool = ( get_replay_pool_from_variant(variant, training_environment)) sampler = self.sampler = get_sampler_from_variant(variant) Qs = self.Qs = get_Q_function_from_variant( variant, training_environment) policy = self.policy = get_policy_from_variant( variant, training_environment, Qs) initial_exploration_policy = self.initial_exploration_policy = ( get_policy('UniformPolicy', training_environment)) self.algorithm = get_algorithm_from_variant( variant=self._variant, training_environment=training_environment, evaluation_environment=evaluation_environment, policy=policy, initial_exploration_policy=initial_exploration_policy, Qs=Qs, pool=replay_pool, sampler=sampler, session=self._session) initialize_tf_variables(self._session, only_uninitialized=True) self._built = True
def build(self): environment_params = self.variant['environment_params'] training_environment = self.training_environment = ( get_environment_from_params(environment_params['training'])) evaluation_environment = self.evaluation_environment = ( get_environment_from_params(environment_params['evaluation']) if 'evaluation' in environment_params else training_environment) replay_pool = self.replay_pool = (get_replay_pool_from_variant( self.variant, training_environment)) sampler = self.sampler = get_sampler_from_variant(self.variant) Qs = self.Qs = get_Q_function_from_variant(self.variant, training_environment) policy = self.policy = get_policy_from_variant(self.variant, training_environment, Qs) initial_exploration_policy = self.initial_exploration_policy = ( get_policy('UniformPolicy', training_environment)) #### get termination function domain = environment_params['training']['domain'] static_fns = static[domain.lower()] #### log_path = './log/%s' % (self.variant['algorithm_params']['domain']) if (not os.path.exists(log_path)): os.makedirs(log_path) self.algorithm = get_algorithm_from_variant( variant=self.variant, training_environment=training_environment, evaluation_environment=evaluation_environment, policy=policy, initial_exploration_policy=initial_exploration_policy, Qs=Qs, pool=replay_pool, static_fns=static_fns, sampler=sampler, session=self._session, log_file='./log/%s/%d.log' % (self.variant['algorithm_params']['domain'], time.time())) initialize_tf_variables(self._session, only_uninitialized=True)
def _build(self): variant = copy.deepcopy(self._variant) print(variant.keys()) env = self.env = get_environment_from_params( variant['environment_params']['training']) replay_pool = self.replay_pool = (get_replay_pool_from_variant( variant, env)) sampler = self.sampler = get_sampler_from_variant(variant) Qs = self.Qs = get_Q_function_from_variant(variant, env) policy = self.policy = get_policy_from_variant(variant, env, Qs) initial_exploration_policy = self.initial_exploration_policy = ( get_policy('UniformPolicy', env)) algorithm_kwargs = { 'variant': self._variant, 'env': self.env, 'policy': policy, 'initial_exploration_policy': initial_exploration_policy, 'Qs': Qs, 'pool': replay_pool, 'sampler': sampler, 'session': self._session, } if self._variant['algorithm_params']['type'] in CLASSIFIER_RL_ALGS: reward_classifier = self.reward_classifier \ = get_reward_classifier_from_variant(self._variant, env) algorithm_kwargs['classifier'] = reward_classifier goal_examples_train, goal_examples_validation = \ get_goal_example_from_variant(variant) algorithm_kwargs['goal_examples'] = goal_examples_train algorithm_kwargs['goal_examples_validation'] = \ goal_examples_validation self.algorithm = get_algorithm_from_variant(**algorithm_kwargs) initialize_tf_variables(self._session, only_uninitialized=True) self._built = True
def main(): import sys example_args = get_parser().parse_args(sys.argv[1:]) variant_spec = get_variant_spec(example_args) command_line_args = example_args print('vriant spec: {}'.format(variant_spec)) params = variant_spec.get('algorithm_params') local_dir = os.path.join(params.get('log_dir'), params.get('domain')) resources_per_trial = _normalize_trial_resources( command_line_args.resources_per_trial, command_line_args.trial_cpus, command_line_args.trial_gpus, command_line_args.trial_extra_cpus, command_line_args.trial_extra_gpus) experiment_id = params.get('exp_name') #### add pool_load_max_size to experiment_id if 'pool_load_max_size' in variant_spec['algorithm_params']['kwargs']: max_size = variant_spec['algorithm_params']['kwargs'][ 'pool_load_max_size'] experiment_id = '{}_{}e3'.format(experiment_id, int(max_size / 1000)) #### variant_spec = add_command_line_args_to_variant_spec( variant_spec, command_line_args) if command_line_args.video_save_frequency is not None: assert 'algorithm_params' in variant_spec variant_spec['algorithm_params']['kwargs']['video_save_frequency'] = ( command_line_args.video_save_frequency) variant = variant_spec # init set_seed(variant['run_params']['seed']) gpu_options = tf.GPUOptions(allow_growth=True) session = tf.Session(config=tf.ConfigProto(gpu_options=gpu_options)) tf.keras.backend.set_session(session) # build variant = copy.deepcopy(variant) tester.set_hyper_param(**variant) tester.add_record_param(['run_params.seed', 'info']) tester.configure(task_name='policy_learn', private_config_path=os.path.join(get_package_path(), 'rla_config.yaml'), run_file='main.py', log_root=get_package_path()) tester.log_files_gen() tester.print_args() environment_params = variant['environment_params'] training_environment = (get_environment_from_params( environment_params['training'])) evaluation_environment = (get_environment_from_params( environment_params['evaluation'](variant)) if 'evaluation' in environment_params else training_environment) replay_pool = (get_replay_pool_from_variant(variant, training_environment)) sampler = get_sampler_from_variant(variant) Qs = get_Q_function_from_variant(variant, training_environment) policy = get_policy_from_variant(variant, training_environment, Qs) initial_exploration_policy = (get_policy('UniformPolicy', training_environment)) #### get termination function domain = environment_params['training']['domain'] static_fns = mopo.static[domain.lower()] #### print("[ DEBUG ] KWARGS: {}".format(variant['algorithm_params']['kwargs'])) algorithm = get_algorithm_from_variant( variant=variant, training_environment=training_environment, evaluation_environment=evaluation_environment, policy=policy, initial_exploration_policy=initial_exploration_policy, Qs=Qs, pool=replay_pool, static_fns=static_fns, sampler=sampler, session=session) print('[ DEBUG ] finish construct model, start training') # train list(algorithm.train())
def _build(self): variant = copy.deepcopy(self._variant) #training_environment = self.training_environment = ( # get_goal_example_environment_from_variant( # variant['task'], gym_adapter=False)) training_environment = self.training_environment = (GymAdapter( domain=variant['domain'], task=variant['task'], **variant['env_params'])) #evaluation_environment = self.evaluation_environment = ( # get_goal_example_environment_from_variant( # variant['task_evaluation'], gym_adapter=False)) evaluation_environment = self.evaluation_environment = (GymAdapter( domain=variant['domain'], task=variant['task_evaluation'], **variant['env_params'])) # training_environment = self.training_environment = ( # flatten_multiworld_env(self.training_environment)) # evaluation_environment = self.evaluation_environment = ( # flatten_multiworld_env(self.evaluation_environment)) #training_environment = self.training_environment = ( # GymAdapter(env=training_environment)) #evaluation_environment = self.evaluation_environment = ( # GymAdapter(env=evaluation_environment)) # make sure this is her replay pool replay_pool = self.replay_pool = (get_replay_pool_from_variant( variant, training_environment)) sampler = self.sampler = get_sampler_from_variant(variant) Qs = self.Qs = get_Q_function_from_variant(variant, training_environment) policy = self.policy = get_policy_from_variant(variant, training_environment) initial_exploration_policy = self.initial_exploration_policy = ( get_policy_from_params(variant['exploration_policy_params'], training_environment)) algorithm_kwargs = { 'variant': self._variant, 'training_environment': self.training_environment, 'evaluation_environment': self.evaluation_environment, 'policy': policy, 'initial_exploration_policy': initial_exploration_policy, 'Qs': Qs, 'pool': replay_pool, 'sampler': sampler, 'session': self._session, } if self._variant['algorithm_params']['type'] in [ 'VICEGoalConditioned', 'VICEGANGoalConditioned' ]: reward_classifier = self.reward_classifier = ( get_reward_classifier_from_variant(self._variant, training_environment)) algorithm_kwargs['classifier'] = reward_classifier # goal_examples_train, goal_examples_validation = \ # get_goal_example_from_variant(variant) algorithm_kwargs['goal_examples'] = np.empty((1, 1)) algorithm_kwargs['goal_examples_validation'] = np.empty((1, 1)) # RND if variant['algorithm_params']['rnd_params']: from softlearning.rnd.utils import get_rnd_networks_from_variant rnd_networks = get_rnd_networks_from_variant( variant, training_environment) else: rnd_networks = () algorithm_kwargs['rnd_networks'] = rnd_networks self.algorithm = get_algorithm_from_variant(**algorithm_kwargs) initialize_tf_variables(self._session, only_uninitialized=True) self._built = True