pushing model

README.md

@@ -0,0 +1,85 @@
+---
+tags:
+- Swimmer-v4
+- deep-reinforcement-learning
+- reinforcement-learning
+- custom-implementation
+library_name: cleanrl
+model-index:
+- name: PPO
+  results:
+  - task:
+      type: reinforcement-learning
+      name: reinforcement-learning
+    dataset:
+      name: Swimmer-v4
+      type: Swimmer-v4
+    metrics:
+    - type: mean_reward
+      value: -13.47 +/- 3.89
+      name: mean_reward
+      verified: false
+---
+
+# (CleanRL) **PPO** Agent Playing **Swimmer-v4**
+
+This is a trained model of a PPO agent playing Swimmer-v4.
+The model was trained by using [CleanRL](https://github.com/vwxyzjn/cleanrl) and the most up-to-date training code can be
+found [here](https://github.com/vwxyzjn/cleanrl/blob/master/cleanrl/ppo_continuous_action.py).
+
+## Get Started
+
+To use this model, please install the `cleanrl` package with the following command:
+
+```
+pip install "cleanrl[ppo_continuous_action]"
+python -m cleanrl_utils.enjoy --exp-name ppo_continuous_action --env-id Swimmer-v4
+```
+
+Please refer to the [documentation](https://docs.cleanrl.dev/get-started/zoo/) for more detail.
+
+
+## Command to reproduce the training
+
+```bash
+curl -OL https://huggingface.co/cleanrl/Swimmer-v4-ppo_continuous_action-seed1/raw/main/ppo_continuous_action.py
+curl -OL https://huggingface.co/cleanrl/Swimmer-v4-ppo_continuous_action-seed1/raw/main/pyproject.toml
+curl -OL https://huggingface.co/cleanrl/Swimmer-v4-ppo_continuous_action-seed1/raw/main/poetry.lock
+poetry install --all-extras
+python ppo_continuous_action.py --track --save-model --upload-model --hf-entity cleanrl --env-id Swimmer-v4 --seed 1
+```
+
+# Hyperparameters
+```python
+{'anneal_lr': True,
+ 'batch_size': 2048,
+ 'capture_video': False,
+ 'clip_coef': 0.2,
+ 'clip_vloss': True,
+ 'cuda': True,
+ 'ent_coef': 0.0,
+ 'env_id': 'Swimmer-v4',
+ 'exp_name': 'ppo_continuous_action',
+ 'gae_lambda': 0.95,
+ 'gamma': 0.99,
+ 'hf_entity': 'cleanrl',
+ 'learning_rate': 0.0003,
+ 'max_grad_norm': 0.5,
+ 'minibatch_size': 64,
+ 'norm_adv': True,
+ 'num_envs': 1,
+ 'num_minibatches': 32,
+ 'num_steps': 2048,
+ 'save_model': True,
+ 'seed': 1,
+ 'target_kl': None,
+ 'torch_deterministic': True,
+ 'total_timesteps': 1000000,
+ 'track': True,
+ 'update_epochs': 10,
+ 'upload_model': True,
+ 'vf_coef': 0.5,
+ 'wandb_entity': None,
+ 'wandb_project_name': 'cleanRL'}
+```
+    

events.out.tfevents.1697394304.3090-172.825155.0

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:30cd49e3840c900dd9935bae542283e03dbed114a11b7e042a0cdc7f488685da
+size 376387

ppo_continuous_action.py

@@ -0,0 +1,355 @@
+# docs and experiment results can be found at https://docs.cleanrl.dev/rl-algorithms/ppo/#ppo_continuous_actionpy
+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.optim as optim
+from torch.distributions.normal import Normal
+from torch.utils.tensorboard import SummaryWriter
+
+
+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="cleanRL",
+        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="HalfCheetah-v4",
+        help="the id of the environment")
+    parser.add_argument("--total-timesteps", type=int, default=1000000,
+        help="total timesteps of the experiments")
+    parser.add_argument("--learning-rate", type=float, default=3e-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("--num-steps", type=int, default=2048,
+        help="the number of steps to run in each environment per policy rollout")
+    parser.add_argument("--anneal-lr", type=lambda x: bool(strtobool(x)), default=True, nargs="?", const=True,
+        help="Toggle learning rate annealing for policy and value networks")
+    parser.add_argument("--gamma", type=float, default=0.99,
+        help="the discount factor gamma")
+    parser.add_argument("--gae-lambda", type=float, default=0.95,
+        help="the lambda for the general advantage estimation")
+    parser.add_argument("--num-minibatches", type=int, default=32,
+        help="the number of mini-batches")
+    parser.add_argument("--update-epochs", type=int, default=10,
+        help="the K epochs to update the policy")
+    parser.add_argument("--norm-adv", type=lambda x: bool(strtobool(x)), default=True, nargs="?", const=True,
+        help="Toggles advantages normalization")
+    parser.add_argument("--clip-coef", type=float, default=0.2,
+        help="the surrogate clipping coefficient")
+    parser.add_argument("--clip-vloss", type=lambda x: bool(strtobool(x)), default=True, nargs="?", const=True,
+        help="Toggles whether or not to use a clipped loss for the value function, as per the paper.")
+    parser.add_argument("--ent-coef", type=float, default=0.0,
+        help="coefficient of the entropy")
+    parser.add_argument("--vf-coef", type=float, default=0.5,
+        help="coefficient of the value function")
+    parser.add_argument("--max-grad-norm", type=float, default=0.5,
+        help="the maximum norm for the gradient clipping")
+    parser.add_argument("--target-kl", type=float, default=None,
+        help="the target KL divergence threshold")
+    args = parser.parse_args()
+    args.batch_size = int(args.num_envs * args.num_steps)
+    args.minibatch_size = int(args.batch_size // args.num_minibatches)
+    # fmt: on
+    return args
+
+
+def make_env(env_id, idx, capture_video, run_name, gamma):
+    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.FlattenObservation(env)  # deal with dm_control's Dict observation space
+        env = gym.wrappers.RecordEpisodeStatistics(env)
+        env = gym.wrappers.ClipAction(env)
+        env = gym.wrappers.NormalizeObservation(env)
+        env = gym.wrappers.TransformObservation(env, lambda obs: np.clip(obs, -10, 10))
+        env = gym.wrappers.NormalizeReward(env, gamma=gamma)
+        env = gym.wrappers.TransformReward(env, lambda reward: np.clip(reward, -10, 10))
+        return env
+
+    return thunk
+
+
+def layer_init(layer, std=np.sqrt(2), bias_const=0.0):
+    torch.nn.init.orthogonal_(layer.weight, std)
+    torch.nn.init.constant_(layer.bias, bias_const)
+    return layer
+
+
+class Agent(nn.Module):
+    def __init__(self, envs):
+        super().__init__()
+        self.critic = nn.Sequential(
+            layer_init(nn.Linear(np.array(envs.single_observation_space.shape).prod(), 64)),
+            nn.Tanh(),
+            layer_init(nn.Linear(64, 64)),
+            nn.Tanh(),
+            layer_init(nn.Linear(64, 1), std=1.0),
+        )
+        self.actor_mean = nn.Sequential(
+            layer_init(nn.Linear(np.array(envs.single_observation_space.shape).prod(), 64)),
+            nn.Tanh(),
+            layer_init(nn.Linear(64, 64)),
+            nn.Tanh(),
+            layer_init(nn.Linear(64, np.prod(envs.single_action_space.shape)), std=0.01),
+        )
+        self.actor_logstd = nn.Parameter(torch.zeros(1, np.prod(envs.single_action_space.shape)))
+
+    def get_value(self, x):
+        return self.critic(x)
+
+    def get_action_and_value(self, x, action=None):
+        action_mean = self.actor_mean(x)
+        action_logstd = self.actor_logstd.expand_as(action_mean)
+        action_std = torch.exp(action_logstd)
+        probs = Normal(action_mean, action_std)
+        if action is None:
+            action = probs.sample()
+        return action, probs.log_prob(action).sum(1), probs.entropy().sum(1), self.critic(x)
+
+
+if __name__ == "__main__":
+    args = parse_args()
+    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, i, args.capture_video, run_name, args.gamma) for i in range(args.num_envs)]
+    )
+    assert isinstance(envs.single_action_space, gym.spaces.Box), "only continuous action space is supported"
+
+    agent = Agent(envs).to(device)
+    optimizer = optim.Adam(agent.parameters(), lr=args.learning_rate, eps=1e-5)
+
+    # ALGO Logic: Storage setup
+    obs = torch.zeros((args.num_steps, args.num_envs) + envs.single_observation_space.shape).to(device)
+    actions = torch.zeros((args.num_steps, args.num_envs) + envs.single_action_space.shape).to(device)
+    logprobs = torch.zeros((args.num_steps, args.num_envs)).to(device)
+    rewards = torch.zeros((args.num_steps, args.num_envs)).to(device)
+    dones = torch.zeros((args.num_steps, args.num_envs)).to(device)
+    values = torch.zeros((args.num_steps, args.num_envs)).to(device)
+
+    # TRY NOT TO MODIFY: start the game
+    global_step = 0
+    start_time = time.time()
+    next_obs, _ = envs.reset(seed=args.seed)
+    next_obs = torch.Tensor(next_obs).to(device)
+    next_done = torch.zeros(args.num_envs).to(device)
+    num_updates = args.total_timesteps // args.batch_size
+
+    for update in range(1, num_updates + 1):
+        # Annealing the rate if instructed to do so.
+        if args.anneal_lr:
+            frac = 1.0 - (update - 1.0) / num_updates
+            lrnow = frac * args.learning_rate
+            optimizer.param_groups[0]["lr"] = lrnow
+
+        for step in range(0, args.num_steps):
+            global_step += 1 * args.num_envs
+            obs[step] = next_obs
+            dones[step] = next_done
+
+            # ALGO LOGIC: action logic
+            with torch.no_grad():
+                action, logprob, _, value = agent.get_action_and_value(next_obs)
+                values[step] = value.flatten()
+            actions[step] = action
+            logprobs[step] = logprob
+
+            # TRY NOT TO MODIFY: execute the game and log data.
+            next_obs, reward, terminations, truncations, infos = envs.step(action.cpu().numpy())
+            done = np.logical_or(terminations, truncations)
+            rewards[step] = torch.tensor(reward).to(device).view(-1)
+            next_obs, next_done = torch.Tensor(next_obs).to(device), torch.Tensor(done).to(device)
+
+            # Only print when at least 1 env is done
+            if "final_info" not in infos:
+                continue
+
+            for info in infos["final_info"]:
+                # Skip the envs that are not done
+                if info is None:
+                    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)
+
+        # bootstrap value if not done
+        with torch.no_grad():
+            next_value = agent.get_value(next_obs).reshape(1, -1)
+            advantages = torch.zeros_like(rewards).to(device)
+            lastgaelam = 0
+            for t in reversed(range(args.num_steps)):
+                if t == args.num_steps - 1:
+                    nextnonterminal = 1.0 - next_done
+                    nextvalues = next_value
+                else:
+                    nextnonterminal = 1.0 - dones[t + 1]
+                    nextvalues = values[t + 1]
+                delta = rewards[t] + args.gamma * nextvalues * nextnonterminal - values[t]
+                advantages[t] = lastgaelam = delta + args.gamma * args.gae_lambda * nextnonterminal * lastgaelam
+            returns = advantages + values
+
+        # flatten the batch
+        b_obs = obs.reshape((-1,) + envs.single_observation_space.shape)
+        b_logprobs = logprobs.reshape(-1)
+        b_actions = actions.reshape((-1,) + envs.single_action_space.shape)
+        b_advantages = advantages.reshape(-1)
+        b_returns = returns.reshape(-1)
+        b_values = values.reshape(-1)
+
+        # Optimizing the policy and value network
+        b_inds = np.arange(args.batch_size)
+        clipfracs = []
+        for epoch in range(args.update_epochs):
+            np.random.shuffle(b_inds)
+            for start in range(0, args.batch_size, args.minibatch_size):
+                end = start + args.minibatch_size
+                mb_inds = b_inds[start:end]
+
+                _, newlogprob, entropy, newvalue = agent.get_action_and_value(b_obs[mb_inds], b_actions[mb_inds])
+                logratio = newlogprob - b_logprobs[mb_inds]
+                ratio = logratio.exp()
+
+                with torch.no_grad():
+                    # calculate approx_kl http://joschu.net/blog/kl-approx.html
+                    old_approx_kl = (-logratio).mean()
+                    approx_kl = ((ratio - 1) - logratio).mean()
+                    clipfracs += [((ratio - 1.0).abs() > args.clip_coef).float().mean().item()]
+
+                mb_advantages = b_advantages[mb_inds]
+                if args.norm_adv:
+                    mb_advantages = (mb_advantages - mb_advantages.mean()) / (mb_advantages.std() + 1e-8)
+
+                # Policy loss
+                pg_loss1 = -mb_advantages * ratio
+                pg_loss2 = -mb_advantages * torch.clamp(ratio, 1 - args.clip_coef, 1 + args.clip_coef)
+                pg_loss = torch.max(pg_loss1, pg_loss2).mean()
+
+                # Value loss
+                newvalue = newvalue.view(-1)
+                if args.clip_vloss:
+                    v_loss_unclipped = (newvalue - b_returns[mb_inds]) ** 2
+                    v_clipped = b_values[mb_inds] + torch.clamp(
+                        newvalue - b_values[mb_inds],
+                        -args.clip_coef,
+                        args.clip_coef,
+                    )
+                    v_loss_clipped = (v_clipped - b_returns[mb_inds]) ** 2
+                    v_loss_max = torch.max(v_loss_unclipped, v_loss_clipped)
+                    v_loss = 0.5 * v_loss_max.mean()
+                else:
+                    v_loss = 0.5 * ((newvalue - b_returns[mb_inds]) ** 2).mean()
+
+                entropy_loss = entropy.mean()
+                loss = pg_loss - args.ent_coef * entropy_loss + v_loss * args.vf_coef
+
+                optimizer.zero_grad()
+                loss.backward()
+                nn.utils.clip_grad_norm_(agent.parameters(), args.max_grad_norm)
+                optimizer.step()
+
+            if args.target_kl is not None:
+                if approx_kl > args.target_kl:
+                    break
+
+        y_pred, y_true = b_values.cpu().numpy(), b_returns.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
+
+        # TRY NOT TO MODIFY: record rewards for plotting purposes
+        writer.add_scalar("charts/learning_rate", optimizer.param_groups[0]["lr"], global_step)
+        writer.add_scalar("losses/value_loss", v_loss.item(), global_step)
+        writer.add_scalar("losses/policy_loss", pg_loss.item(), global_step)
+        writer.add_scalar("losses/entropy", entropy_loss.item(), global_step)
+        writer.add_scalar("losses/old_approx_kl", old_approx_kl.item(), global_step)
+        writer.add_scalar("losses/approx_kl", approx_kl.item(), global_step)
+        writer.add_scalar("losses/clipfrac", np.mean(clipfracs), global_step)
+        writer.add_scalar("losses/explained_variance", explained_var, 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)
+
+    if args.save_model:
+        model_path = f"runs/{run_name}/{args.exp_name}.cleanrl_model"
+        torch.save(agent.state_dict(), model_path)
+        print(f"model saved to {model_path}")
+        from cleanrl_utils.evals.ppo_eval import evaluate
+
+        episodic_returns = evaluate(
+            model_path,
+            make_env,
+            args.env_id,
+            eval_episodes=10,
+            run_name=f"{run_name}-eval",
+            Model=Agent,
+            device=device,
+            gamma=args.gamma,
+        )
+        for idx, episodic_return in enumerate(episodic_returns):
+            writer.add_scalar("eval/episodic_return", episodic_return, idx)
+
+        if args.upload_model:
+            from cleanrl_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, "PPO", f"runs/{run_name}", f"videos/{run_name}-eval")
+
+    envs.close()
+    writer.close()

pyproject.toml

@@ -0,0 +1,108 @@
+[tool.poetry]
+name = "cleanrl"
+version = "1.1.0"
+description = "High-quality single file implementation of Deep Reinforcement Learning algorithms with research-friendly features"
+authors = ["Costa Huang <costa.huang@outlook.com>"]
+packages = [
+    { include = "cleanrl" },
+    { include = "cleanrl_utils" },
+]
+keywords = ["reinforcement", "machine", "learning", "research"]
+license="MIT"
+readme = "README.md"
+
+[tool.poetry.dependencies]
+python = ">=3.7.1,<3.11"
+tensorboard = "^2.10.0"
+wandb = "^0.13.11"
+gym = "0.23.1"
+torch = ">=1.12.1"
+stable-baselines3 = "1.2.0"
+gymnasium = ">=0.28.1"
+moviepy = "^1.0.3"
+pygame = "2.1.0"
+huggingface-hub = "^0.11.1"
+rich = "<12.0"
+tenacity = "^8.2.2"
+
+ale-py = {version = "0.7.4", optional = true}
+AutoROM = {extras = ["accept-rom-license"], version = "^0.4.2", optional = true}
+opencv-python = {version = "^4.6.0.66", optional = true}
+procgen = {version = "^0.10.7", optional = true}
+pytest = {version = "^7.1.3", optional = true}
+mujoco = {version = "<=2.3.3", optional = true}
+imageio = {version = "^2.14.1", optional = true}
+free-mujoco-py = {version = "^2.1.6", optional = true}
+mkdocs-material = {version = "^8.4.3", optional = true}
+markdown-include = {version = "^0.7.0", optional = true}
+openrlbenchmark = {version = "^0.1.1b4", optional = true}
+jax = {version = "^0.3.17", optional = true}
+jaxlib = {version = "^0.3.15", optional = true}
+flax = {version = "^0.6.0", optional = true}
+optuna = {version = "^3.0.1", optional = true}
+optuna-dashboard = {version = "^0.7.2", optional = true}
+envpool = {version = "^0.6.4", optional = true}
+PettingZoo = {version = "1.18.1", optional = true}
+SuperSuit = {version = "3.4.0", optional = true}
+multi-agent-ale-py = {version = "0.1.11", optional = true}
+boto3 = {version = "^1.24.70", optional = true}
+awscli = {version = "^1.25.71", optional = true}
+shimmy = {version = ">=1.0.0", extras = ["dm-control"], optional = true}
+
+[tool.poetry.group.dev.dependencies]
+pre-commit = "^2.20.0"
+
+
+[tool.poetry.group.isaacgym]
+optional = true
+[tool.poetry.group.isaacgym.dependencies]
+isaacgymenvs = {git = "https://github.com/vwxyzjn/IsaacGymEnvs.git", rev = "poetry", python = ">=3.7.1,<3.10"}
+isaacgym = {path = "cleanrl/ppo_continuous_action_isaacgym/isaacgym", develop = true}
+
+
+[build-system]
+requires = ["poetry-core"]
+build-backend = "poetry.core.masonry.api"
+
+[tool.poetry.extras]
+atari = ["ale-py", "AutoROM", "opencv-python"]
+procgen = ["procgen"]
+plot = ["pandas", "seaborn"]
+pytest = ["pytest"]
+mujoco = ["mujoco", "imageio"]
+mujoco_py = ["free-mujoco-py"]
+jax = ["jax", "jaxlib", "flax"]
+docs = ["mkdocs-material", "markdown-include", "openrlbenchmark"]
+envpool = ["envpool"]
+optuna = ["optuna", "optuna-dashboard"]
+pettingzoo = ["PettingZoo", "SuperSuit", "multi-agent-ale-py"]
+cloud = ["boto3", "awscli"]
+dm_control = ["shimmy", "mujoco"]
+
+# dependencies for algorithm variant (useful when you want to run a specific algorithm)
+dqn = []
+dqn_atari = ["ale-py", "AutoROM", "opencv-python"]
+dqn_jax = ["jax", "jaxlib", "flax"]
+dqn_atari_jax = [
+    "ale-py", "AutoROM", "opencv-python", # atari
+    "jax", "jaxlib", "flax" # jax
+]
+c51 = []
+c51_atari = ["ale-py", "AutoROM", "opencv-python"]
+c51_jax = ["jax", "jaxlib", "flax"]
+c51_atari_jax = [
+    "ale-py", "AutoROM", "opencv-python", # atari
+    "jax", "jaxlib", "flax" # jax
+]
+ppo_atari_envpool_xla_jax_scan = [
+    "ale-py", "AutoROM", "opencv-python", # atari
+    "jax", "jaxlib", "flax", # jax
+    "envpool", # envpool
+]
+qdagger_dqn_atari_impalacnn = [
+    "ale-py", "AutoROM", "opencv-python"
+]
+qdagger_dqn_atari_jax_impalacnn = [
+    "ale-py", "AutoROM", "opencv-python", # atari
+    "jax", "jaxlib", "flax", # jax
+]