# docs and experiment results can be found at https://docs.cleanrl.dev/rl-algorithms/dqn/#dqn_ataripy import argparse import os import random import time from distutils.util import strtobool import gymnasium as gym import numpy as np import torch import torch.nn as nn import torch.nn.functional as F import torch.optim as optim from stable_baselines3.common.atari_wrappers import ( ClipRewardEnv, EpisodicLifeEnv, FireResetEnv, MaxAndSkipEnv, NoopResetEnv, ) from stable_baselines3.common.buffers import ReplayBuffer from torch.utils.tensorboard import SummaryWriter from typing import Callable from dotenv import load_dotenv import sys sys.path.append('.') from wrappers.rotation import RotateObservation def parse_args(): # fmt: off parser = argparse.ArgumentParser() parser.add_argument("--exp-name", type=str, default=os.path.basename(__file__).rstrip(".py"), help="the name of this experiment") parser.add_argument("--seed", type=int, default=1, help="seed of the experiment") parser.add_argument("--torch-deterministic", type=lambda x: bool(strtobool(x)), default=True, nargs="?", const=True, help="if toggled, `torch.backends.cudnn.deterministic=False`") parser.add_argument("--cuda", type=lambda x: bool(strtobool(x)), default=True, nargs="?", const=True, help="if toggled, cuda will be enabled by default") parser.add_argument("--track", type=lambda x: bool(strtobool(x)), default=False, nargs="?", const=True, help="if toggled, this experiment will be tracked with Weights and Biases") parser.add_argument("--wandb-project-name", type=str, default="lagomorph", help="the wandb's project name") parser.add_argument("--wandb-entity", type=str, default=None, help="the entity (team) of wandb's project") parser.add_argument("--capture-video", type=lambda x: bool(strtobool(x)), default=False, nargs="?", const=True, help="whether to capture videos of the agent performances (check out `videos` folder)") parser.add_argument("--save-model", type=lambda x: bool(strtobool(x)), default=False, nargs="?", const=True, help="whether to save model into the `runs/{run_name}` folder") parser.add_argument("--upload-model", type=lambda x: bool(strtobool(x)), default=False, nargs="?", const=True, help="whether to upload the saved model to huggingface") parser.add_argument("--hf-entity", type=str, default="", help="the user or org name of the model repository from the Hugging Face Hub") # Algorithm specific arguments parser.add_argument("--env-id", type=str, default="BreakoutNoFrameskip-v4", help="the id of the environment") parser.add_argument("--total-timesteps", type=int, default=10000000, help="total timesteps of the experiments") parser.add_argument("--learning-rate", type=float, default=1e-4, help="the learning rate of the optimizer") parser.add_argument("--num-envs", type=int, default=1, help="the number of parallel game environments") parser.add_argument("--buffer-size", type=int, default=1000000, help="the replay memory buffer size") parser.add_argument("--gamma", type=float, default=0.99, help="the discount factor gamma") parser.add_argument("--tau", type=float, default=1., help="the target network update rate") parser.add_argument("--target-network-frequency", type=int, default=1000, help="the timesteps it takes to update the target network") parser.add_argument("--batch-size", type=int, default=32, help="the batch size of sample from the reply memory") parser.add_argument("--start-e", type=float, default=1, help="the starting epsilon for exploration") parser.add_argument("--end-e", type=float, default=0.01, help="the ending epsilon for exploration") parser.add_argument("--exploration-fraction", type=float, default=0.10, help="the fraction of `total-timesteps` it takes from start-e to go end-e") parser.add_argument("--learning-starts", type=int, default=80000, help="timestep to start learning") parser.add_argument("--train-frequency", type=int, default=4, help="the frequency of training") parser.add_argument("--checkpoint", type=lambda x: bool(strtobool(x)), default=False, nargs="?", const=True, help="whether to checkpoint the model") parser.add_argument("--checkpoint-frequency", type=int, default=100000, help="the frequency of saving checkpoint") parser.add_argument("--resume", type=lambda x: bool(strtobool(x)), default=False, nargs="?", const=True, help="resume training from the last checkpoint") parser.add_argument("--initial-steps", type=int, default=0, help="the number of steps that were done in previous training runs") parser.add_argument("--rotate", type=lambda x: bool(strtobool(x)), default=False, nargs="?", const=True, help="whether to rotate the observation sometimes") args = parser.parse_args() # fmt: on assert args.num_envs == 1, "vectorized envs are not supported at the moment" return args def make_env(env_id, seed, idx, capture_video, run_name): def thunk(): if capture_video and idx == 0: env = gym.make(env_id, render_mode="rgb_array") env = gym.wrappers.RecordVideo(env, f"videos/{run_name}") else: env = gym.make(env_id) env = gym.wrappers.RecordEpisodeStatistics(env) env = NoopResetEnv(env, noop_max=30) env = MaxAndSkipEnv(env, skip=4) env = EpisodicLifeEnv(env) if "FIRE" in env.unwrapped.get_action_meanings(): env = FireResetEnv(env) env = ClipRewardEnv(env) env = gym.wrappers.ResizeObservation(env, (84, 84)) env = gym.wrappers.GrayScaleObservation(env) if args.rotate: env = RotateObservation(env) env = gym.wrappers.FrameStack(env, 4) env.action_space.seed(seed) return env return thunk # ALGO LOGIC: initialize agent here: class QNetwork(nn.Module): def __init__(self, env): super().__init__() self.network = nn.Sequential( nn.Conv2d(4, 32, 8, stride=4), nn.ReLU(), nn.Conv2d(32, 64, 4, stride=2), nn.ReLU(), nn.Conv2d(64, 64, 3, stride=1), nn.ReLU(), nn.Flatten(), nn.Linear(3136, 512), nn.ReLU(), nn.Linear(512, env.single_action_space.n), ) def forward(self, x): return self.network(x / 255.0) def linear_schedule(start_e: float, end_e: float, duration: int, t: int): slope = (end_e - start_e) / duration return max(slope * t + start_e, end_e) def evaluate( model_path: str, make_env: Callable, env_id: str, eval_episodes: int, run_name: str, Model: torch.nn.Module, device: torch.device = torch.device("cpu"), epsilon: float = 0.05, capture_video: bool = True, ): envs = gym.vector.SyncVectorEnv([make_env(env_id, 0, 0, capture_video, run_name)]) model = Model(envs).to(device) model.load_state_dict(torch.load(model_path, map_location=device)) model.eval() obs, _ = envs.reset() episodic_returns = [] while len(episodic_returns) < eval_episodes: if random.random() < epsilon: actions = np.array([envs.single_action_space.sample() for _ in range(envs.num_envs)]) else: q_values = model(torch.Tensor(obs).to(device)) actions = torch.argmax(q_values, dim=1).cpu().numpy() next_obs, _, _, _, infos = envs.step(actions) if "final_info" in infos: for info in infos["final_info"]: if "episode" in info: print(f"eval_episode={len(episodic_returns)}, episodic_return={info['episode']['r']}") episodic_returns += [info["episode"]["r"]] obs = next_obs return episodic_returns if __name__ == "__main__": import stable_baselines3 as sb3 if sb3.__version__ < "2.0": raise ValueError( """Ongoing migration: run the following command to install the new dependencies: poetry run pip install "stable_baselines3==2.0.0a1" "gymnasium[atari,accept-rom-license]==0.28.1" "ale-py==0.8.1" """ ) args = parse_args() load_dotenv() run_name = f"{args.env_id}__{args.exp_name}__{args.seed}__{int(time.time())}" if args.track: import wandb wandb.init( project=args.wandb_project_name, entity=args.wandb_entity, sync_tensorboard=True, config=vars(args), name=run_name, monitor_gym=True, save_code=True, ) writer = SummaryWriter(f"runs/{run_name}") writer.add_text( "hyperparameters", "|param|value|\n|-|-|\n%s" % ("\n".join([f"|{key}|{value}|" for key, value in vars(args).items()])), ) # TRY NOT TO MODIFY: seeding random.seed(args.seed) np.random.seed(args.seed) torch.manual_seed(args.seed) torch.backends.cudnn.deterministic = args.torch_deterministic device = torch.device("cuda" if torch.cuda.is_available() and args.cuda else "cpu") # env setup envs = gym.vector.SyncVectorEnv( [make_env(args.env_id, args.seed + i, i, args.capture_video, run_name) for i in range(args.num_envs)] ) assert isinstance(envs.single_action_space, gym.spaces.Discrete), "only discrete action space is supported" q_network = QNetwork(envs).to(device) optimizer = optim.Adam(q_network.parameters(), lr=args.learning_rate) target_network = QNetwork(envs).to(device) target_network.load_state_dict(q_network.state_dict()) rb = ReplayBuffer( args.buffer_size, envs.single_observation_space, envs.single_action_space, device, optimize_memory_usage=True, handle_timeout_termination=False, ) start_time = time.time() if args.resume: checkpoint_path = f"models/Pong.checkpoint.pth" q_network.load_state_dict(torch.load(checkpoint_path)) print(f"checkpoint loaded from {checkpoint_path}") # TRY NOT TO MODIFY: start the game obs, _ = envs.reset(seed=args.seed) for global_step in range(args.initial_steps, args.total_timesteps): # ALGO LOGIC: put action logic here epsilon = linear_schedule(args.start_e, args.end_e, args.exploration_fraction * args.total_timesteps, global_step) if random.random() < epsilon: actions = np.array([envs.single_action_space.sample() for _ in range(envs.num_envs)]) else: q_values = q_network(torch.Tensor(obs).to(device)) actions = torch.argmax(q_values, dim=1).cpu().numpy() # TRY NOT TO MODIFY: execute the game and log data. next_obs, rewards, terminated, truncated, infos = envs.step(actions) # TRY NOT TO MODIFY: record rewards for plotting purposes if "final_info" in infos: for info in infos["final_info"]: # Skip the envs that are not done if "episode" not in info: continue print(f"global_step={global_step}, episodic_return={info['episode']['r']}") writer.add_scalar("charts/episodic_return", info["episode"]["r"], global_step) writer.add_scalar("charts/episodic_length", info["episode"]["l"], global_step) writer.add_scalar("charts/epsilon", epsilon, global_step) # TRY NOT TO MODIFY: save data to reply buffer; handle `final_observation` real_next_obs = next_obs.copy() for idx, d in enumerate(truncated): if d: real_next_obs[idx] = infos["final_observation"][idx] rb.add(obs, real_next_obs, actions, rewards, terminated, infos) # TRY NOT TO MODIFY: CRUCIAL step easy to overlook obs = next_obs # ALGO LOGIC: training. if global_step > args.learning_starts: if global_step % args.train_frequency == 0: data = rb.sample(args.batch_size) with torch.no_grad(): target_max, _ = target_network(data.next_observations).max(dim=1) td_target = data.rewards.flatten() + args.gamma * target_max * (1 - data.dones.flatten()) old_val = q_network(data.observations).gather(1, data.actions).squeeze() loss = F.mse_loss(td_target, old_val) if global_step % 100 == 0: writer.add_scalar("losses/td_loss", loss, global_step) writer.add_scalar("losses/q_values", old_val.mean().item(), global_step) print("SPS:", int(global_step / (time.time() - start_time))) writer.add_scalar("charts/SPS", int(global_step / (time.time() - start_time)), global_step) # optimize the model optimizer.zero_grad() loss.backward() optimizer.step() # update target network if global_step % args.target_network_frequency == 0: for target_network_param, q_network_param in zip(target_network.parameters(), q_network.parameters()): target_network_param.data.copy_( args.tau * q_network_param.data + (1.0 - args.tau) * target_network_param.data ) # save checkpoint if (args.checkpoint): if global_step % args.checkpoint_frequency == 0: checkpoint_path = f"runs/{run_name}/{args.exp_name}.checkpoint.pth" torch.save(q_network.state_dict(), checkpoint_path) print(f"checkpoint saved to {checkpoint_path}") if args.track: artifact = wandb.Artifact(f"{args.exp_name}_checkpoint", "checkpoint") artifact.add_file(checkpoint_path) wandb.log_artifact(artifact) if args.save_model: model_path = f"runs/{run_name}/Pong-base.cleanrl_model" torch.save(q_network.state_dict(), model_path) print(f"model saved to {model_path}") if args.track: artifact = wandb.Artifact(f"Pong-base_model", "model") artifact.add_file(model_path) wandb.log_artifact(artifact) episodic_returns = evaluate( model_path, make_env, args.env_id, eval_episodes=10, run_name=f"{run_name}-eval", Model=QNetwork, device=device, epsilon=0.05, ) for idx, episodic_return in enumerate(episodic_returns): writer.add_scalar("eval/episodic_return", episodic_return, idx) if args.upload_model: #TODO: Currently gives an error. *Maybe* I'll get around to fixing it once everything else works correctly. from utils.huggingface import push_to_hub repo_name = f"{args.env_id}-{args.exp_name}-seed{args.seed}" repo_id = f"{args.hf_entity}/{repo_name}" if args.hf_entity else repo_name push_to_hub(args, episodic_returns, repo_id, "DQN", f"runs/{run_name}", f"videos/{run_name}-eval") envs.close() writer.close()