def offline_gym( env_name: str, pkl_path: str, num_train_transitions: int, max_steps: Optional[int], seed: Optional[int] = None, ): """ Generate samples from a DiscreteRandomPolicy on the Gym environment and saves results in a pandas df parquet. """ initialize_seed(seed) env = EnvFactory.make(env_name) replay_buffer = ReplayBuffer.create_from_env( env=env, replay_memory_size=num_train_transitions, batch_size=1) fill_replay_buffer(env, replay_buffer, num_train_transitions) if isinstance(env.action_space, gym.spaces.Discrete): is_discrete_action = True else: assert isinstance(env.action_space, gym.spaces.Box) is_discrete_action = False df = replay_buffer_to_pre_timeline_df(is_discrete_action, replay_buffer) logger.info(f"Saving dataset with {len(df)} samples to {pkl_path}") df.to_pickle(pkl_path)
def run_test_offline( env_name: str, max_steps: Optional[int], model: ModelManager__Union, replay_memory_size: int, num_batches_per_epoch: int, num_train_epochs: int, passing_score_bar: float, num_eval_episodes: int, use_gpu: bool, ): env = EnvFactory.make(env_name) env.seed(SEED) env.action_space.seed(SEED) normalization = build_normalizer(env) logger.info(f"Normalization is: \n{pprint.pformat(normalization)}") manager = model.value trainer = manager.initialize_trainer( use_gpu=use_gpu, reward_options=RewardOptions(), normalization_data_map=normalization, ) # first fill the replay buffer to burn_in replay_buffer = ReplayBuffer.create_from_env( env=env, replay_memory_size=replay_memory_size, batch_size=trainer.minibatch_size, ) # always fill full RB fill_replay_buffer( env=env, replay_buffer=replay_buffer, desired_size=replay_memory_size ) device = torch.device("cuda") if use_gpu else None # pyre-fixme[6]: Expected `device` for 2nd param but got `Optional[torch.device]`. trainer_preprocessor = make_replay_buffer_trainer_preprocessor(trainer, device, env) writer = SummaryWriter() with summary_writer_context(writer): for epoch in range(num_train_epochs): logger.info(f"Evaluating before epoch {epoch}: ") eval_rewards = evaluate_cem(env, manager, max_steps, 1) for _ in tqdm(range(num_batches_per_epoch)): train_batch = replay_buffer.sample_transition_batch_tensor() preprocessed_batch = trainer_preprocessor(train_batch) trainer.train(preprocessed_batch) logger.info(f"Evaluating after training for {num_train_epochs} epochs: ") eval_rewards = evaluate_cem(env, manager, max_steps, num_eval_episodes) mean_rewards = np.mean(eval_rewards) assert ( mean_rewards >= passing_score_bar ), f"{mean_rewards} doesn't pass the bar {passing_score_bar}."
def run_test( env: str, model: ModelManager__Union, replay_memory_size: int, train_every_ts: int, train_after_ts: int, num_episodes: int, max_steps: Optional[int], last_score_bar: float, ): env = EnvFactory.make(env) env.seed(SEED) env.action_space.seed(SEED) normalization = build_normalizer(env) logger.info(f"Normalization is {normalization}") manager = model.value trainer = manager.initialize_trainer( use_gpu=False, reward_options=RewardOptions(), normalization_data_map=normalization, ) policy = manager.create_policy() replay_buffer = ReplayBuffer.create_from_env( env=env, replay_memory_size=replay_memory_size, batch_size=trainer.minibatch_size, ) post_step = train_with_replay_buffer_post_step( replay_buffer=replay_buffer, trainer=trainer, training_freq=train_every_ts, batch_size=trainer.minibatch_size, replay_burnin=train_after_ts, ) agent = Agent.create_for_env(env, policy=policy, post_transition_callback=post_step) reward_history = [] for i in range(num_episodes): logger.info(f"running episode {i}") ep_reward = run_episode(env=env, agent=agent, max_steps=max_steps) reward_history.append(ep_reward) assert reward_history[-1] >= last_score_bar, ( f"reward after {len(reward_history)} episodes is {reward_history[-1]}," f"less than < {last_score_bar}...\n" f"Full reward history: {reward_history}") return reward_history
def train_seq2reward( env: gym.Env, trainer: Seq2RewardTrainer, trainer_preprocessor, num_train_transitions: int, seq_len: int, batch_size: int, num_train_epochs: int, # for optional validation test_replay_buffer=None, ): train_replay_buffer = ReplayBuffer.create_from_env( env=env, replay_memory_size=num_train_transitions, batch_size=batch_size, stack_size=seq_len, return_everything_as_stack=True, ) fill_replay_buffer(env, train_replay_buffer, num_train_transitions) num_batch_per_epoch = train_replay_buffer.size // batch_size logger.info("Made RBs, starting to train now!") state_dim = env.observation_space.shape[0] for epoch in range(num_train_epochs): for i in range(num_batch_per_epoch): batch = train_replay_buffer.sample_transition_batch(batch_size=batch_size) preprocessed_batch = trainer_preprocessor(batch) adhoc_action_padding(preprocessed_batch, state_dim=state_dim) losses = trainer.train(preprocessed_batch) print_seq2reward_losses(epoch, i, losses) # validation if test_replay_buffer is not None: with torch.no_grad(): trainer.seq2reward_network.eval() test_batch = test_replay_buffer.sample_transition_batch( batch_size=batch_size ) preprocessed_test_batch = trainer_preprocessor(test_batch) adhoc_action_padding(preprocessed_test_batch, state_dim=state_dim) valid_losses = trainer.get_loss(preprocessed_test_batch) print_seq2reward_losses(epoch, "validation", valid_losses) trainer.seq2reward_network.train() return trainer
def create_embed_rl_dataset( env: gym.Env, memory_network: MemoryNetwork, num_state_embed_transitions: int, batch_size: int, seq_len: int, hidden_dim: int, use_gpu: bool, ): assert isinstance(env.action_space, gym.spaces.Discrete) assert isinstance(env.observation_space, gym.spaces.Box) assert len(env.observation_space.shape) == 1 logger.info("Starting to create embedded RL Dataset!") # seqlen+1 because MDNRNN embeds the first seq_len steps and then # the embedded state will be concatenated with the last step # Ie.. (o1,o2,...,on) -> RNN -> h1,h2,...,hn # and we set s_{n+1} = [o_{n+1}, h_n] embed_env = StateEmbedEnvironment( gym_env=env, mdnrnn=memory_network, max_embed_seq_len=seq_len + 1 ) # now create a filled replay buffer of embeddings # new obs shape dim = state_dim + hidden_dim embed_rb = ReplayBuffer.create_from_env( env=embed_env, replay_memory_size=num_state_embed_transitions, batch_size=batch_size, stack_size=1, ) fill_replay_buffer( env=embed_env, replay_buffer=embed_rb, desired_size=num_state_embed_transitions ) batch = embed_rb.sample_transition_batch_tensor( batch_size=num_state_embed_transitions ) state_min = min(batch.state.min(), batch.next_state.min()).item() state_max = max(batch.state.max(), batch.next_state.max()).item() logger.info( f"Finished making embed dataset with size {embed_rb.size}, " f"min {state_min}, max {state_max}" ) return embed_rb, state_min, state_max
def train_mdnrnn( env: gym.Env, trainer: MDNRNNTrainer, trainer_preprocessor, num_train_transitions: int, seq_len: int, batch_size: int, num_train_epochs: int, # for optional validation test_replay_buffer=None, ): train_replay_buffer = ReplayBuffer.create_from_env( env=env, replay_memory_size=num_train_transitions, batch_size=batch_size, stack_size=seq_len, return_everything_as_stack=True, ) fill_replay_buffer(env, train_replay_buffer, num_train_transitions) num_batch_per_epoch = train_replay_buffer.size // batch_size logger.info("Made RBs, starting to train now!") for epoch in range(num_train_epochs): for i in range(num_batch_per_epoch): batch = train_replay_buffer.sample_transition_batch_tensor( batch_size=batch_size ) preprocessed_batch = trainer_preprocessor(batch) losses = trainer.train(preprocessed_batch) print_mdnrnn_losses(epoch, i, losses) # validation if test_replay_buffer is not None: with torch.no_grad(): trainer.memory_network.mdnrnn.eval() test_batch = test_replay_buffer.sample_transition_batch_tensor( batch_size=batch_size ) preprocessed_test_batch = trainer_preprocessor(test_batch) valid_losses = trainer.get_loss(preprocessed_test_batch) print_mdnrnn_losses(epoch, "validation", valid_losses) trainer.memory_network.mdnrnn.train() return trainer
def test_create_from_recsim_interest_evolution(self): env_config = { "num_candidates": 20, "slate_size": 3, "resample_documents": False, "seed": 1, } env = interest_evolution.create_environment(env_config) replay_buffer = ReplayBuffer.create_from_env(env, replay_memory_size=100, batch_size=10, store_log_prob=True) obs = env.reset() observation = obs["user"] action = env.action_space.sample() log_prob = -1.0 doc_features = np.stack(list(obs["doc"].values()), axis=0) next_obs, reward, terminal, _env = env.step(action) response = next_obs["response"] click = np.array([r["click"] for r in response]) response_quality = np.stack([r["quality"] for r in response], axis=0) repsonse_cluster_id = np.array([r["cluster_id"] for r in response]) response_watch_time = np.stack([r["watch_time"] for r in response], axis=0) response_liked = np.array([r["liked"] for r in response]) replay_buffer.add( observation, action, reward, terminal, mdp_id=0, sequence_number=0, doc=doc_features, response_click=click, response_cluster_id=repsonse_cluster_id, response_quality=response_quality, response_liked=response_liked, response_watch_time=response_watch_time, log_prob=log_prob, )
def run_test( env_name: str, model: ModelManager__Union, replay_memory_size: int, train_every_ts: int, train_after_ts: int, num_train_episodes: int, max_steps: Optional[int], passing_score_bar: float, num_eval_episodes: int, use_gpu: bool, ): env = EnvFactory.make(env_name) env.seed(SEED) env.action_space.seed(SEED) normalization = build_normalizer(env) logger.info(f"Normalization is: \n{pprint.pformat(normalization)}") manager = model.value try: # pyre-fixme[16]: `Env` has no attribute `state_feature_config_provider`. manager.state_feature_config_provider = env.state_feature_config_provider logger.info( f"Using environment's state_feature_config_provider.\n" f"{manager.state_feature_config_provider}" ) except AttributeError: logger.info("state_feature_config_provider override not applicable") trainer = manager.initialize_trainer( use_gpu=use_gpu, reward_options=RewardOptions(), normalization_data_map=normalization, ) training_policy = manager.create_policy(serving=False) replay_buffer = ReplayBuffer.create_from_env( env=env, replay_memory_size=replay_memory_size, batch_size=trainer.minibatch_size, ) device = torch.device("cuda") if use_gpu else torch.device("cpu") # first fill the replay buffer to burn_in train_after_ts = max(train_after_ts, trainer.minibatch_size) fill_replay_buffer( env=env, replay_buffer=replay_buffer, desired_size=train_after_ts ) post_step = train_with_replay_buffer_post_step( replay_buffer=replay_buffer, env=env, trainer=trainer, training_freq=train_every_ts, batch_size=trainer.minibatch_size, device=device, ) agent = Agent.create_for_env( env, policy=training_policy, post_transition_callback=post_step, device=device ) writer = SummaryWriter() with summary_writer_context(writer): train_rewards = [] for i in range(num_train_episodes): trajectory = run_episode( env=env, agent=agent, mdp_id=i, max_steps=max_steps ) ep_reward = trajectory.calculate_cumulative_reward() train_rewards.append(ep_reward) logger.info( f"Finished training episode {i} (len {len(trajectory)})" f" with reward {ep_reward}." ) logger.info("============Train rewards=============") logger.info(train_rewards) logger.info(f"average: {np.mean(train_rewards)};\tmax: {np.max(train_rewards)}") # Check whether the max score passed the score bar; we explore during training # the return could be bad (leading to flakiness in C51 and QRDQN). assert np.max(train_rewards) >= passing_score_bar, ( f"max reward ({np.max(train_rewards)})after training for " f"{len(train_rewards)} episodes is less than < {passing_score_bar}.\n" ) serving_policy = manager.create_policy(serving=True) agent = Agent.create_for_env_with_serving_policy(env, serving_policy) eval_rewards = evaluate_for_n_episodes( n=num_eval_episodes, env=env, agent=agent, max_steps=max_steps ).squeeze(1) logger.info("============Eval rewards==============") logger.info(eval_rewards) logger.info(f"average: {np.mean(eval_rewards)};\tmax: {np.max(eval_rewards)}") assert np.mean(eval_rewards) >= passing_score_bar, ( f"Predictor reward is {np.mean(eval_rewards)}," f"less than < {passing_score_bar}.\n" )
def run_test( env_name: str, model: ModelManager__Union, replay_memory_size: int, train_every_ts: int, train_after_ts: int, num_train_episodes: int, max_steps: Optional[int], passing_score_bar: float, num_eval_episodes: int, use_gpu: bool, ): env = EnvFactory.make(env_name) env.seed(SEED) env.action_space.seed(SEED) normalization = build_normalizer(env) logger.info(f"Normalization is: \n{pprint.pformat(normalization)}") manager = model.value trainer = manager.initialize_trainer( use_gpu=use_gpu, reward_options=RewardOptions(), normalization_data_map=normalization, ) training_policy = manager.create_policy(serving=False) replay_buffer = ReplayBuffer.create_from_env( env=env, replay_memory_size=replay_memory_size, batch_size=trainer.minibatch_size, ) device = torch.device("cuda") if use_gpu else None # first fill the replay buffer to burn_in train_after_ts = max(train_after_ts, trainer.minibatch_size) fill_replay_buffer(env=env, replay_buffer=replay_buffer, desired_size=train_after_ts) post_step = train_with_replay_buffer_post_step( replay_buffer=replay_buffer, env=env, trainer=trainer, training_freq=train_every_ts, batch_size=trainer.minibatch_size, device=device, ) agent = Agent.create_for_env( env, policy=training_policy, post_transition_callback=post_step, # pyre-fixme[6]: Expected `Union[str, torch.device]` for 4th param but got # `Optional[torch.device]`. device=device, ) writer = SummaryWriter() with summary_writer_context(writer): train_rewards = [] for i in range(num_train_episodes): trajectory = run_episode(env=env, agent=agent, mdp_id=i, max_steps=max_steps) ep_reward = trajectory.calculate_cumulative_reward() train_rewards.append(ep_reward) logger.info( f"Finished training episode {i} with reward {ep_reward}.") assert train_rewards[-1] >= passing_score_bar, ( f"reward after {len(train_rewards)} episodes is {train_rewards[-1]}," f"less than < {passing_score_bar}...\n" f"Full reward history: {train_rewards}") logger.info("============Train rewards=============") logger.info(train_rewards) serving_policy = manager.create_policy(serving=True) agent = Agent.create_for_env_with_serving_policy(env, serving_policy) eval_rewards = evaluate_for_n_episodes(n=num_eval_episodes, env=env, agent=agent, max_steps=max_steps).squeeze(1) assert np.mean(eval_rewards) >= passing_score_bar, ( f"Predictor reward is {np.mean(eval_rewards)}," f"less than < {passing_score_bar}...\n" f"Full eval rewards: {eval_rewards}.") logger.info("============Eval rewards==============") logger.info(eval_rewards)
def train_mdnrnn_and_compute_feature_stats( env_name: str, model: ModelManager__Union, num_train_transitions: int, num_test_transitions: int, seq_len: int, batch_size: int, num_train_epochs: int, use_gpu: bool, saved_mdnrnn_path: Optional[str] = None, ): """ Train MDNRNN Memory Network and compute feature importance/sensitivity. """ env: gym.Env = EnvFactory.make(env_name) env.seed(SEED) manager = model.value trainer = manager.initialize_trainer( use_gpu=use_gpu, reward_options=RewardOptions(), normalization_data_map=build_normalizer(env), ) device = "cuda" if use_gpu else "cpu" trainer_preprocessor = make_replay_buffer_trainer_preprocessor(trainer, device, env) test_replay_buffer = ReplayBuffer.create_from_env( env=env, replay_memory_size=num_test_transitions, batch_size=batch_size, stack_size=seq_len, return_everything_as_stack=True, ) fill_replay_buffer(env, test_replay_buffer, num_test_transitions) if saved_mdnrnn_path is None: # train from scratch trainer = train_mdnrnn( env=env, trainer=trainer, trainer_preprocessor=trainer_preprocessor, num_train_transitions=num_train_transitions, seq_len=seq_len, batch_size=batch_size, num_train_epochs=num_train_epochs, test_replay_buffer=test_replay_buffer, ) else: # load a pretrained model, and just evaluate it trainer.memory_network.mdnrnn.load_state_dict(torch.load(saved_mdnrnn_path)) with torch.no_grad(): trainer.memory_network.mdnrnn.eval() test_batch = test_replay_buffer.sample_transition_batch_tensor( batch_size=test_replay_buffer.size ) preprocessed_test_batch = trainer_preprocessor(test_batch) feature_importance = calculate_feature_importance( env=env, trainer=trainer, use_gpu=use_gpu, test_batch=preprocessed_test_batch, ) feature_sensitivity = calculate_feature_sensitivity( env=env, trainer=trainer, use_gpu=use_gpu, test_batch=preprocessed_test_batch, ) trainer.memory_network.mdnrnn.train() return feature_importance, feature_sensitivity
def run_test( env: str, model: ModelManager__Union, replay_memory_size: int, train_every_ts: int, train_after_ts: int, num_train_episodes: int, max_steps: Optional[int], passing_score_bar: float, num_eval_episodes: int, use_gpu: bool, ): env = EnvFactory.make(env) env.seed(SEED) env.action_space.seed(SEED) normalization = build_normalizer(env) logger.info(f"Normalization is {normalization}") manager = model.value trainer = manager.initialize_trainer( use_gpu=use_gpu, reward_options=RewardOptions(), normalization_data_map=normalization, ) replay_buffer = ReplayBuffer.create_from_env( env=env, replay_memory_size=replay_memory_size, batch_size=trainer.minibatch_size, ) device = torch.device("cuda") if use_gpu else None post_step = train_with_replay_buffer_post_step( replay_buffer=replay_buffer, trainer=trainer, training_freq=train_every_ts, batch_size=trainer.minibatch_size, replay_burnin=train_after_ts, device=device, ) training_policy = manager.create_policy(serving=False) agent = Agent.create_for_env(env, policy=training_policy, post_transition_callback=post_step, device=device) train_rewards = [] for i in range(num_train_episodes): ep_reward = run_episode(env=env, agent=agent, max_steps=max_steps) train_rewards.append(ep_reward) logger.info(f"Finished training episode {i} with reward {ep_reward}.") assert train_rewards[-1] >= passing_score_bar, ( f"reward after {len(train_rewards)} episodes is {train_rewards[-1]}," f"less than < {passing_score_bar}...\n" f"Full reward history: {train_rewards}") logger.info("============Train rewards=============") logger.info(train_rewards) def gym_to_reagent_serving( obs: np.array) -> Tuple[torch.Tensor, torch.Tensor]: obs_tensor = torch.tensor(obs).float().unsqueeze(0) presence_tensor = torch.ones_like(obs_tensor) return (obs_tensor, presence_tensor) serving_policy = manager.create_policy(serving=True) agent = Agent.create_for_env(env, policy=serving_policy, obs_preprocessor=gym_to_reagent_serving) eval_rewards = [] for i in range(num_eval_episodes): ep_reward = run_episode(env=env, agent=agent, max_steps=max_steps) eval_rewards.append(ep_reward) logger.info(f"Finished eval episode {i} with reward {ep_reward}.") assert np.mean(eval_rewards) >= passing_score_bar, ( f"Predictor reward is {np.mean(eval_rewards)}," f"less than < {passing_score_bar}...\n" f"Full eval rewards: {eval_rewards}.") logger.info("============Eval rewards==============") logger.info(eval_rewards)
def train_seq2reward_and_compute_reward_mse( env_name: str, model: ModelManager__Union, num_train_transitions: int, num_test_transitions: int, seq_len: int, batch_size: int, num_train_epochs: int, use_gpu: bool, saved_seq2reward_path: Optional[str] = None, ): """ Train Seq2Reward Network and compute reward mse. """ env: gym.Env = EnvFactory.make(env_name) env.seed(SEED) manager = model.value trainer = manager.initialize_trainer( use_gpu=use_gpu, reward_options=RewardOptions(), normalization_data_map=build_normalizer(env), ) device = "cuda" if use_gpu else "cpu" # pyre-fixme[6]: Expected `device` for 2nd param but got `str`. trainer_preprocessor = make_replay_buffer_trainer_preprocessor(trainer, device, env) test_replay_buffer = ReplayBuffer.create_from_env( env=env, replay_memory_size=num_test_transitions, batch_size=batch_size, stack_size=seq_len, return_everything_as_stack=True, ) fill_replay_buffer(env, test_replay_buffer, num_test_transitions) if saved_seq2reward_path is None: # train from scratch trainer = train_seq2reward( env=env, trainer=trainer, trainer_preprocessor=trainer_preprocessor, num_train_transitions=num_train_transitions, seq_len=seq_len, batch_size=batch_size, num_train_epochs=num_train_epochs, test_replay_buffer=test_replay_buffer, ) else: # load a pretrained model, and just evaluate it trainer.seq2reward_network.load_state_dict(torch.load(saved_seq2reward_path)) state_dim = env.observation_space.shape[0] with torch.no_grad(): trainer.seq2reward_network.eval() test_batch = test_replay_buffer.sample_transition_batch( batch_size=test_replay_buffer.size ) preprocessed_test_batch = trainer_preprocessor(test_batch) adhoc_action_padding(preprocessed_test_batch, state_dim=state_dim) losses = trainer.get_loss(preprocessed_test_batch) detached_losses = losses.cpu().detach().item() trainer.seq2reward_network.train() return detached_losses