def test_episode_limit_with_vectorized_env_dataset(batch_size): """ Test that when adding the EpisodeLimit wrapper on top of a vectorized environment, the episode limit is with respect to each individual env rather than the batched env. """ start = 0 target = 10 starting_values = [start for i in range(batch_size)] targets = [target for i in range(batch_size)] env = SyncVectorEnv([ partial(DummyEnvironment, start=start, target=target, max_value=10 * 2) for start, target in zip(starting_values, targets) ]) max_episodes = 2 # TODO: For some reason the reverse order doesn't work! env = EpisodeLimit(env, max_episodes=max_episodes * batch_size) env = EnvDataset(env) for i, obs in enumerate(env): print(i, obs) actions = np.ones(batch_size) reward = env.send(actions) assert i == max_episodes * target - 1 with pytest.raises(gym.error.ClosedEnvironmentError): env.reset() with pytest.raises(gym.error.ClosedEnvironmentError): for i, obs in enumerate(env): print(i, obs) actions = np.ones(batch_size) reward = env.send(actions)
def test_step_limit_with_vectorized_env(batch_size): start = 0 target = 10 starting_values = [start for i in range(batch_size)] targets = [target for i in range(batch_size)] env = SyncVectorEnv([ partial(DummyEnvironment, start=start, target=target, max_value=target * 2) for start, target in zip(starting_values, targets) ]) env = ObservationLimit(env, max_steps=3 * batch_size) obs = env.reset() obs, reward, done, info = env.step(env.action_space.sample()) # obs, reward, done, info = env.step(env.action_space.sample()) obs = env.reset() assert env.is_closed with pytest.raises(gym.error.ClosedEnvironmentError): env.reset() with pytest.raises(gym.error.ClosedEnvironmentError): _ = env.step(env.action_space.sample())
def test_reset_vectorenv_with_unfinished_episodes_raises_warning(batch_size): """ Test that when adding the EpisodeLimit wrapper on top of a vectorized environment, the episode limit is with respect to each individual env rather than the batched env. """ start = 0 target = 10 starting_values = [start for i in range(batch_size)] targets = [target for i in range(batch_size)] env = SyncVectorEnv([ partial(DummyEnvironment, start=start, target=target, max_value=10 * 2) for start, target in zip(starting_values, targets) ]) env = EpisodeLimit(env, max_episodes=3 * batch_size) obs = env.reset() _ = env.step(env.action_space.sample()) _ = env.step(env.action_space.sample()) with pytest.warns(UserWarning) as record: env.reset()
def test_step_limit_with_vectorized_env_partial_final_batch(batch_size): """ In the case where the batch size isn't a multiple of the max observations, the env returns ceil(max_obs / batch_size) * batch_size observations in total. TODO: If we ever get to few-shot learning or something like that, we might have to care about this. """ start = 0 target = 10 starting_values = [start for i in range(batch_size)] targets = [target for i in range(batch_size)] env = SyncVectorEnv([ partial(DummyEnvironment, start=start, target=target, max_value=target * 2) for start, target in zip(starting_values, targets) ]) env = ObservationLimit(env, max_steps=3 * batch_size + 1) obs = env.reset() assert not env.is_closed obs, reward, done, info = env.step(env.action_space.sample()) obs, reward, done, info = env.step(env.action_space.sample()) assert not env.is_closed # obs, reward, done, info = env.step(env.action_space.sample()) obs = env.reset() assert env.is_closed with pytest.raises(gym.error.ClosedEnvironmentError): env.reset() with pytest.raises(gym.error.ClosedEnvironmentError): _ = env.step(env.action_space.sample())
def test_episode_limit_with_vectorized_env(batch_size): """ Test that when adding the EpisodeLimit wrapper on top of a vectorized environment, the episode limit is with respect to each individual env rather than the batched env. """ starting_values = [0 for i in range(batch_size)] targets = [10 for i in range(batch_size)] env = SyncVectorEnv([ partial(DummyEnvironment, start=start, target=target, max_value=10 * 2) for start, target in zip(starting_values, targets) ]) env = EpisodeLimit(env, max_episodes=2 * batch_size) obs = env.reset() assert obs.tolist() == starting_values print("reset obs: ", obs) for i in range(10): print(i, obs) actions = np.ones(batch_size) obs, reward, done, info = env.step(actions) # all episodes end at step 10 assert all(done) # Because of how VectorEnvs work, the obs are the new 'reset' obs, rather # than the final obs in the episode. assert obs.tolist() == starting_values assert obs.tolist() == starting_values print("reset obs: ", obs) for i in range(10): print(i, obs) actions = np.ones(batch_size) obs, reward, done, info = env.step(actions) # all episodes end at step 10 assert all(done) assert env.is_closed assert obs.tolist() == starting_values with pytest.raises(gym.error.ClosedEnvironmentError): actions = np.ones(batch_size) obs, reward, done, info = env.step(actions)
def test_space_with_tuple_observations(batch_size: int, n_workers: Optional[int]): def make_env(): env = gym.make("Breakout-v0") env = MultiTaskEnvironment( env, add_task_id_to_obs=True, add_task_dict_to_info=True ) return env env_fn = make_env env_fns = [env_fn for _ in range(batch_size)] # from gym.vector.utils import batch_space # env = BatchedVectorEnv(env_fns, n_workers=n_workers) from gym.vector import SyncVectorEnv env = SyncVectorEnv(env_fns) # FIXME: debugging # env = AsyncVectorEnv(env_fns) env.seed(123) assert env.observation_space == spaces.Dict( x=spaces.Box(0, 255, (batch_size, 210, 160, 3), np.uint8), task_labels=spaces.MultiDiscrete(np.ones(batch_size)), ) assert env.single_observation_space == spaces.Dict( x=spaces.Box(0, 255, (210, 160, 3), np.uint8), task_labels=spaces.Discrete(1) ) obs = env.reset() assert obs["x"].shape == env.observation_space["x"].shape assert obs["task_labels"].shape == env.observation_space["task_labels"].shape assert obs in env.observation_space actions = env.action_space.sample() step_obs, rewards, done, info = env.step(actions) assert step_obs in env.observation_space assert len(rewards) == batch_size assert len(done) == batch_size assert all([isinstance(v, bool) for v in done.tolist()]), [type(v) for v in done] assert len(info) == batch_size
def main(cfg): random.seed(cfg.exp.seed) np.random.seed(cfg.exp.seed) torch.manual_seed(cfg.exp.seed) torch.backends.cudnn.deterministic = cfg.exp.torch_deterministic # so that the environment automatically resets env = SyncVectorEnv([ lambda: RecordEpisodeStatistics(gym.make('CartPole-v1')) ]) actor, critic = Actor(), Critic() actor_optim = Adam(actor.parameters(), eps=1e-5, lr=cfg.params.actor_lr) critic_optim = Adam(critic.parameters(), eps=1e-5, lr=cfg.params.critic_lr) memory = Memory(mini_batch_size=cfg.params.mini_batch_size, batch_size=cfg.params.batch_size) obs = env.reset() global_rewards = [] NUM_UPDATES = (cfg.params.total_timesteps // cfg.params.batch_size) * cfg.params.epochs cur_timestep = 0 def calc_factor(cur_timestep: int) -> float: """Calculates the factor to be multiplied with the learning rate to update it.""" update_number = cur_timestep // cfg.params.batch_size total_updates = cfg.params.total_timesteps // cfg.params.batch_size fraction = 1.0 - update_number / total_updates return fraction actor_scheduler = LambdaLR(actor_optim, lr_lambda=calc_factor, verbose=True) critic_scheduler = LambdaLR(critic_optim, lr_lambda=calc_factor, verbose=True) while cur_timestep < cfg.params.total_timesteps: # keep playing the game obs = torch.as_tensor(obs, dtype=torch.float32) with torch.no_grad(): dist = actor(obs) action = dist.sample() log_prob = dist.log_prob(action) value = critic(obs) action = action.cpu().numpy() value = value.cpu().numpy() log_prob = log_prob.cpu().numpy() obs_, reward, done, info = env.step(action) if done[0]: tqdm.write(f'Reward: {info[0]["episode"]["r"]}, Avg Reward: {np.mean(global_rewards[-10:]):.3f}') global_rewards.append(info[0]['episode']['r']) wandb.log({'Avg_Reward': np.mean(global_rewards[-10:]), 'Reward': info[0]['episode']['r']}) memory.remember(obs.squeeze(0).cpu().numpy(), action.item(), log_prob.item(), reward.item(), done.item(), value.item()) obs = obs_ cur_timestep += 1 # if the current timestep is a multiple of the batch size, then we need to update the model if cur_timestep % cfg.params.batch_size == 0: for epoch in tqdm(range(cfg.params.epochs), desc=f'Num updates: {cfg.params.epochs * (cur_timestep // cfg.params.batch_size)} / {NUM_UPDATES}'): # sample a batch from memory of experiences old_states, old_actions, old_log_probs, old_rewards, old_dones, old_values, batch_indices = memory.sample() old_log_probs = torch.tensor(old_log_probs, dtype=torch.float32) old_actions = torch.tensor(old_actions, dtype=torch.float32) advantage = calculate_advantage(old_rewards, old_values, old_dones, gae_gamma=cfg.params.gae_gamma, gae_lambda=cfg.params.gae_lambda) advantage = torch.tensor(advantage, dtype=torch.float32) old_rewards = torch.tensor(old_rewards, dtype=torch.float32) old_values = torch.tensor(old_values, dtype=torch.float32) # for each mini batch from batch, calculate advantage using GAE for mini_batch_index in batch_indices: # remember: Normalization of advantage is done on mini batch, not the entire batch advantage[mini_batch_index] = (advantage[mini_batch_index] - advantage[mini_batch_index].mean()) / (advantage[mini_batch_index].std() + 1e-8) dist = actor(torch.tensor(old_states[mini_batch_index], dtype=torch.float32).unsqueeze(0)) # actions = dist.sample() log_probs = dist.log_prob(old_actions[mini_batch_index]).squeeze(0) entropy = dist.entropy().squeeze(0) log_ratio = log_probs - old_log_probs[mini_batch_index] ratio = torch.exp(log_ratio) with torch.no_grad(): # approx_kl = ((ratio-1)-log_ratio).mean() approx_kl = ((old_log_probs[mini_batch_index] - log_probs)**2).mean() wandb.log({'Approx_KL': approx_kl}) actor_loss = -torch.min( ratio * advantage[mini_batch_index], torch.clamp(ratio, 1 - cfg.params.actor_loss_clip, 1 + cfg.params.actor_loss_clip) * advantage[mini_batch_index] ).mean() values = critic(torch.tensor(old_states[mini_batch_index], dtype=torch.float32).unsqueeze(0)).squeeze(-1) returns = old_values[mini_batch_index] + advantage[mini_batch_index] critic_loss = torch.max( (values - returns)**2, (old_values[mini_batch_index] + torch.clamp( values - old_values[mini_batch_index], -cfg.params.critic_loss_clip, cfg.params.critic_loss_clip ) - returns )**2 ).mean() # critic_loss = F.mse_loss(values, returns) wandb.log({'Actor_Loss': actor_loss.item(), 'Critic_Loss': critic_loss.item(), 'Entropy': entropy.mean().item()}) loss = actor_loss + 0.25 * critic_loss - 0.01 * entropy.mean() actor_optim.zero_grad() critic_optim.zero_grad() loss.backward() nn.utils.clip_grad_norm_(actor.parameters(), cfg.params.max_grad_norm) nn.utils.clip_grad_norm_(critic.parameters(), cfg.params.max_grad_norm) actor_optim.step() critic_optim.step() memory.reset() actor_scheduler.step(cur_timestep) critic_scheduler.step(cur_timestep) y_pred, y_true = old_values.cpu().numpy(), (old_values + advantage).cpu().numpy() var_y = np.var(y_true) explained_var = np.nan if var_y == 0 else 1 - np.var(y_true - y_pred) / var_y wandb.log({'Explained_Var': explained_var}) if cfg.exp.save_weights: torch.save(actor.state_dict(), Path(f'{hydra.utils.get_original_cwd()}/{cfg.exp.model_dir}/actor.pth')) torch.save(critic.state_dict(), Path(f'{hydra.utils.get_original_cwd()}/{cfg.exp.model_dir}/critic.pth'))