def test_trainer_maxq(self): env = Env(self.state_dims, self.action_dims) env.seed(42) maxq_parameters = DiscreteActionModelParameters( actions=env.actions, rl=RLParameters( gamma=0.99, target_update_rate=0.9, reward_burnin=100, maxq_learning=True, ), rainbow=RainbowDQNParameters(double_q_learning=True, dueling_architecture=False), training=TrainingParameters( layers=self.layers, activations=self.activations, minibatch_size=self.minibatch_size, learning_rate=0.25, optimizer="ADAM", ), ) maxq_trainer = DQNTrainer(maxq_parameters, env.normalization) logger.info("Generating constant_reward MDPs..") states, actions, rewards, next_states, next_actions, is_terminal, possible_actions, possible_next_actions = env.generate_samples_discrete( self.num_samples) logger.info("Preprocessing constant_reward MDPs..") for epoch in range(self.epochs): tdps = env.preprocess_samples_discrete( states, actions, rewards, next_states, next_actions, is_terminal, possible_actions, possible_next_actions, self.minibatch_size, ) logger.info("Training.. " + str(epoch)) for tdp in tdps: maxq_trainer.train(tdp) logger.info(" ".join([ "Training epoch", str(epoch), "average q values", str(torch.mean(maxq_trainer.all_action_scores)), ])) # Q value should converge to very close to 100 avg_q_value_after_training = torch.mean(maxq_trainer.all_action_scores) self.assertLess(avg_q_value_after_training, 101) self.assertGreater(avg_q_value_after_training, 99)
def train_network(params): logger.info("Running DQN workflow with params:") logger.info(params) action_names = np.array(params["actions"]) rl_parameters = RLParameters(**params["rl"]) training_parameters = TrainingParameters(**params["training"]) rainbow_parameters = RainbowDQNParameters(**params["rainbow"]) trainer_params = DiscreteActionModelParameters( actions=params["actions"], rl=rl_parameters, training=training_parameters, rainbow=rainbow_parameters, ) dataset = JSONDataset(params["training_data_path"], batch_size=training_parameters.minibatch_size) norm_data = JSONDataset(params["state_norm_data_path"]) state_normalization = read_norm_params(norm_data.read_all()) num_batches = int(len(dataset) / training_parameters.minibatch_size) logger.info("Read in batch data set {} of size {} examples. Data split " "into {} batches of size {}.".format( params["training_data_path"], len(dataset), num_batches, training_parameters.minibatch_size, )) trainer = DQNTrainer(trainer_params, state_normalization, params["use_gpu"]) for epoch in range(params["epochs"]): for batch_idx in range(num_batches): helpers.report_training_status(batch_idx, num_batches, epoch, params["epochs"]) batch = dataset.read_batch(batch_idx) tdp = preprocess_batch_for_training(action_names, batch, state_normalization) trainer.train(tdp) logger.info("Training finished. Saving PyTorch model to {}".format( params["pytorch_output_path"])) helpers.save_model_to_file(trainer, params["pytorch_output_path"])
def train_network(params): writer = None if params["model_output_path"] is not None: writer = SummaryWriter(log_dir=params["model_output_path"]) logger.info("Running DQN workflow with params:") logger.info(params) # Set minibatch size based on # of devices being used to train params["training"]["minibatch_size"] *= minibatch_size_multiplier( params["use_gpu"], params["use_all_avail_gpus"]) action_names = np.array(params["actions"]) rl_parameters = RLParameters(**params["rl"]) training_parameters = TrainingParameters(**params["training"]) rainbow_parameters = RainbowDQNParameters(**params["rainbow"]) trainer_params = DiscreteActionModelParameters( actions=params["actions"], rl=rl_parameters, training=training_parameters, rainbow=rainbow_parameters, ) dataset = JSONDataset(params["training_data_path"], batch_size=training_parameters.minibatch_size) eval_dataset = JSONDataset(params["eval_data_path"], batch_size=16) state_normalization = read_norm_file(params["state_norm_data_path"]) num_batches = int(len(dataset) / training_parameters.minibatch_size) logger.info("Read in batch data set {} of size {} examples. Data split " "into {} batches of size {}.".format( params["training_data_path"], len(dataset), num_batches, training_parameters.minibatch_size, )) trainer = DQNTrainer( trainer_params, state_normalization, use_gpu=params["use_gpu"], use_all_avail_gpus=params["use_all_avail_gpus"], ) trainer = update_model_for_warm_start(trainer) preprocessor = Preprocessor(state_normalization, False) evaluator = Evaluator( trainer_params.actions, trainer_params.rl.gamma, trainer, metrics_to_score=trainer.metrics_to_score, ) start_time = time.time() for epoch in range(int(params["epochs"])): dataset.reset_iterator() for batch_idx in range(num_batches): report_training_status(batch_idx, num_batches, epoch, int(params["epochs"])) batch = dataset.read_batch(batch_idx) tdp = preprocess_batch_for_training(preprocessor, batch, action_names) tdp.set_type(trainer.dtype) trainer.train(tdp) eval_dataset.reset_iterator() accumulated_edp = None while True: batch = eval_dataset.read_batch(batch_idx) if batch is None: break tdp = preprocess_batch_for_training(preprocessor, batch, action_names) edp = EvaluationDataPage.create_from_tdp(tdp, trainer) if accumulated_edp is None: accumulated_edp = edp else: accumulated_edp = accumulated_edp.append(edp) accumulated_edp = accumulated_edp.compute_values(trainer.gamma) cpe_start_time = time.time() details = evaluator.evaluate_post_training(accumulated_edp) details.log() logger.info("CPE evaluation took {} seconds.".format(time.time() - cpe_start_time)) through_put = (len(dataset) * int(params["epochs"])) / (time.time() - start_time) logger.info("Training finished. Processed ~{} examples / s.".format( round(through_put))) if writer is not None: writer.close() return export_trainer_and_predictor(trainer, params["model_output_path"])
def train_network(params): writer = None if params["model_output_path"] is not None: writer = SummaryWriter(log_dir=params["model_output_path"]) logger.info("Running DQN workflow with params:") logger.info(params) # Set minibatch size based on # of devices being used to train params["training"]["minibatch_size"] *= minibatch_size_multiplier( params["use_gpu"], params["use_all_avail_gpus"] ) action_names = np.array(params["actions"]) rl_parameters = RLParameters(**params["rl"]) training_parameters = TrainingParameters(**params["training"]) rainbow_parameters = RainbowDQNParameters(**params["rainbow"]) trainer_params = DiscreteActionModelParameters( actions=params["actions"], rl=rl_parameters, training=training_parameters, rainbow=rainbow_parameters, ) dataset = JSONDataset( params["training_data_path"], batch_size=training_parameters.minibatch_size ) eval_dataset = JSONDataset(params["eval_data_path"], batch_size=16) state_normalization = read_norm_file(params["state_norm_data_path"]) num_batches = int(len(dataset) / training_parameters.minibatch_size) logger.info( "Read in batch data set {} of size {} examples. Data split " "into {} batches of size {}.".format( params["training_data_path"], len(dataset), num_batches, training_parameters.minibatch_size, ) ) trainer = DQNTrainer( trainer_params, state_normalization, use_gpu=params["use_gpu"], use_all_avail_gpus=params["use_all_avail_gpus"], ) trainer = update_model_for_warm_start(trainer) preprocessor = Preprocessor(state_normalization, False) evaluator = Evaluator( trainer_params.actions, trainer_params.rl.gamma, trainer, metrics_to_score=trainer.metrics_to_score, ) start_time = time.time() for epoch in range(int(params["epochs"])): dataset.reset_iterator() batch_idx = -1 while True: batch_idx += 1 report_training_status(batch_idx, num_batches, epoch, int(params["epochs"])) batch = dataset.read_batch() if batch is None: break tdp = preprocess_batch_for_training(preprocessor, batch, action_names) tdp.set_type(trainer.dtype) trainer.train(tdp) eval_dataset.reset_iterator() accumulated_edp = None while True: batch = eval_dataset.read_batch() if batch is None: break tdp = preprocess_batch_for_training(preprocessor, batch, action_names) tdp.set_type(trainer.dtype) edp = EvaluationDataPage.create_from_tdp(tdp, trainer) if accumulated_edp is None: accumulated_edp = edp else: accumulated_edp = accumulated_edp.append(edp) accumulated_edp = accumulated_edp.compute_values(trainer.gamma) cpe_start_time = time.time() details = evaluator.evaluate_post_training(accumulated_edp) details.log() logger.info( "CPE evaluation took {} seconds.".format(time.time() - cpe_start_time) ) through_put = (len(dataset) * int(params["epochs"])) / (time.time() - start_time) logger.info( "Training finished. Processed ~{} examples / s.".format(round(through_put)) ) if writer is not None: writer.close() return export_trainer_and_predictor(trainer, params["model_output_path"])
def train_network(params): writer = None if params["model_output_path"] is not None: writer = SummaryWriter( log_dir=os.path.join( os.path.expanduser(params["model_output_path"]), "training_data" ) ) logger.info("Running DQN workflow with params:") logger.info(params) action_names = np.array(params["actions"]) rl_parameters = RLParameters(**params["rl"]) training_parameters = TrainingParameters(**params["training"]) rainbow_parameters = RainbowDQNParameters(**params["rainbow"]) if params["in_training_cpe"] is not None: in_training_cpe_parameters = InTrainingCPEParameters( **params["in_training_cpe"] ) else: in_training_cpe_parameters = None trainer_params = DiscreteActionModelParameters( actions=params["actions"], rl=rl_parameters, training=training_parameters, rainbow=rainbow_parameters, in_training_cpe=in_training_cpe_parameters, ) dataset = JSONDataset( params["training_data_path"], batch_size=training_parameters.minibatch_size ) state_normalization = read_norm_file(params["state_norm_data_path"]) num_batches = int(len(dataset) / training_parameters.minibatch_size) logger.info( "Read in batch data set {} of size {} examples. Data split " "into {} batches of size {}.".format( params["training_data_path"], len(dataset), num_batches, training_parameters.minibatch_size, ) ) trainer = DQNTrainer(trainer_params, state_normalization, params["use_gpu"]) trainer = update_model_for_warm_start(trainer) preprocessor = Preprocessor(state_normalization, params["use_gpu"]) if trainer_params.in_training_cpe is not None: evaluator = Evaluator( trainer_params.actions, 10, trainer_params.rl.gamma, trainer, trainer_params.in_training_cpe.mdp_sampled_rate, ) else: evaluator = Evaluator( trainer_params.actions, 10, trainer_params.rl.gamma, trainer, float(DEFAULT_NUM_SAMPLES_FOR_CPE) / len(dataset), ) start_time = time.time() for epoch in range(int(params["epochs"])): for batch_idx in range(num_batches): report_training_status(batch_idx, num_batches, epoch, int(params["epochs"])) batch = dataset.read_batch(batch_idx) tdp = preprocess_batch_for_training(preprocessor, batch, action_names) trainer.train(tdp) trainer.evaluate( evaluator, tdp.actions, None, tdp.rewards, tdp.episode_values ) evaluator.collect_discrete_action_samples( mdp_ids=tdp.mdp_ids, sequence_numbers=tdp.sequence_numbers.cpu().numpy(), states=tdp.states.cpu().numpy(), logged_actions=tdp.actions.cpu().numpy(), logged_rewards=tdp.rewards.cpu().numpy(), logged_propensities=tdp.propensities.cpu().numpy(), logged_terminals=np.invert( tdp.not_terminals.cpu().numpy().astype(np.bool) ), ) cpe_start_time = time.time() evaluator.recover_samples_to_be_unshuffled() evaluator.score_cpe() if writer is not None: evaluator.log_to_tensorboard(writer, epoch) evaluator.clear_collected_samples() logger.info( "CPE evaluation took {} seconds.".format(time.time() - cpe_start_time) ) through_put = (len(dataset) * int(params["epochs"])) / (time.time() - start_time) logger.info( "Training finished. Processed ~{} examples / s.".format(round(through_put)) ) if writer is not None: writer.close() return export_trainer_and_predictor(trainer, params["model_output_path"])
def test_trainer_maxq(self): env = Env(self.state_dims, self.action_dims) env.seed(42) maxq_parameters = DiscreteActionModelParameters( actions=env.actions, rl=RLParameters( gamma=0.99, target_update_rate=0.9, reward_burnin=100, maxq_learning=True, ), rainbow=RainbowDQNParameters( double_q_learning=True, dueling_architecture=False ), training=TrainingParameters( layers=self.layers, activations=self.activations, minibatch_size=self.minibatch_size, learning_rate=0.25, optimizer="ADAM", ), ) maxq_trainer = DQNTrainer(maxq_parameters, env.normalization) logger.info("Generating constant_reward MDPs..") states, actions, rewards, next_states, next_actions, is_terminal, possible_actions, possible_next_actions = env.generate_samples_discrete( self.num_samples ) logger.info("Preprocessing constant_reward MDPs..") for epoch in range(self.epochs): tdps = env.preprocess_samples_discrete( states, actions, rewards, next_states, next_actions, is_terminal, possible_actions, possible_next_actions, self.minibatch_size, ) logger.info("Training.. " + str(epoch)) for tdp in tdps: maxq_trainer.train(tdp) logger.info( " ".join( [ "Training epoch", str(epoch), "average q values", str(torch.mean(maxq_trainer.all_action_scores)), ] ) ) # Q value should converge to very close to 100 avg_q_value_after_training = torch.mean(maxq_trainer.all_action_scores) self.assertLess(avg_q_value_after_training, 101) self.assertGreater(avg_q_value_after_training, 99)