pushing model

README.md

@@ -0,0 +1,63 @@
+---
+tags:
+- CartPole-v1
+- deep-reinforcement-learning
+- reinforcement-learning
+- custom-implementation
+model-index:
+- name: DQN
+  results:
+  - task:
+      type: reinforcement-learning
+      name: reinforcement-learning
+    dataset:
+      name: CartPole-v1
+      type: CartPole-v1
+    metrics:
+    - type: mean_reward
+      value: 500.00 +/- 0.00
+      name: mean_reward
+      verified: false
+---
+
+# (CleanRL) **DQN** Agent Playing **CartPole-v1**
+
+This is a trained model of a DQN agent playing CartPole-v1.
+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/dqn_jax.py).
+
+## Command to reproduce the training
+
+```bash
+curl -OL https://huggingface.co/cleanrl/CartPole-v1-dqn_jax-seed1/raw/main/dqn.py
+curl -OL https://huggingface.co/cleanrl/CartPole-v1-dqn_jax-seed1/raw/main/pyproject.toml
+curl -OL https://huggingface.co/cleanrl/CartPole-v1-dqn_jax-seed1/raw/main/poetry.lock
+poetry install --all-extras
+python dqn_jax.py --save-model --upload-model --hf-entity cleanrl --seed 1
+```
+
+# Hyperparameters
+```python
+{'batch_size': 128,
+ 'buffer_size': 10000,
+ 'capture_video': False,
+ 'end_e': 0.05,
+ 'env_id': 'CartPole-v1',
+ 'exp_name': 'dqn_jax',
+ 'exploration_fraction': 0.5,
+ 'gamma': 0.99,
+ 'hf_entity': 'cleanrl',
+ 'learning_rate': 0.00025,
+ 'learning_starts': 10000,
+ 'save_model': True,
+ 'seed': 1,
+ 'start_e': 1,
+ 'target_network_frequency': 500,
+ 'total_timesteps': 500000,
+ 'track': False,
+ 'train_frequency': 10,
+ 'upload_model': True,
+ 'wandb_entity': None,
+ 'wandb_project_name': 'cleanRL'}
+```
+    

dqn_jax.py

@@ -0,0 +1,265 @@
+# docs and experiment results can be found at https://docs.cleanrl.dev/rl-algorithms/dqn/#dqn_jaxpy
+import argparse
+import os
+import random
+import time
+from distutils.util import strtobool
+
+import flax
+import flax.linen as nn
+import gym
+import jax
+import jax.numpy as jnp
+import numpy as np
+import optax
+from flax.training.train_state import TrainState
+from stable_baselines3.common.buffers import ReplayBuffer
+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("--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="CartPole-v1",
+        help="the id of the environment")
+    parser.add_argument("--total-timesteps", type=int, default=500000,
+        help="total timesteps of the experiments")
+    parser.add_argument("--learning-rate", type=float, default=2.5e-4,
+        help="the learning rate of the optimizer")
+    parser.add_argument("--buffer-size", type=int, default=10000,
+        help="the replay memory buffer size")
+    parser.add_argument("--gamma", type=float, default=0.99,
+        help="the discount factor gamma")
+    parser.add_argument("--target-network-frequency", type=int, default=500,
+        help="the timesteps it takes to update the target network")
+    parser.add_argument("--batch-size", type=int, default=128,
+        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=0.05,
+        help="the ending epsilon for exploration")
+    parser.add_argument("--exploration-fraction", type=float, default=0.5,
+        help="the fraction of `total-timesteps` it takes from start-e to go end-e")
+    parser.add_argument("--learning-starts", type=int, default=10000,
+        help="timestep to start learning")
+    parser.add_argument("--train-frequency", type=int, default=10,
+        help="the frequency of training")
+    args = parser.parse_args()
+    # fmt: on
+    return args
+
+
+def make_env(env_id, seed, idx, capture_video, run_name):
+    def thunk():
+        env = gym.make(env_id)
+        env = gym.wrappers.RecordEpisodeStatistics(env)
+        if capture_video:
+            if idx == 0:
+                env = gym.wrappers.RecordVideo(env, f"videos/{run_name}")
+        env.seed(seed)
+        env.action_space.seed(seed)
+        env.observation_space.seed(seed)
+        return env
+
+    return thunk
+
+
+# ALGO LOGIC: initialize agent here:
+class QNetwork(nn.Module):
+    action_dim: int
+
+    @nn.compact
+    def __call__(self, x: jnp.ndarray):
+        x = nn.Dense(120)(x)
+        x = nn.relu(x)
+        x = nn.Dense(84)(x)
+        x = nn.relu(x)
+        x = nn.Dense(self.action_dim)(x)
+        return x
+
+
+class TrainState(TrainState):
+    target_params: flax.core.FrozenDict
+
+
+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)
+
+
+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)
+    key = jax.random.PRNGKey(args.seed)
+    key, q_key = jax.random.split(key, 2)
+
+    # env setup
+    envs = gym.vector.SyncVectorEnv([make_env(args.env_id, args.seed, 0, args.capture_video, run_name)])
+    assert isinstance(envs.single_action_space, gym.spaces.Discrete), "only discrete action space is supported"
+
+    obs = envs.reset()
+
+    q_network = QNetwork(action_dim=envs.single_action_space.n)
+
+    q_state = TrainState.create(
+        apply_fn=q_network.apply,
+        params=q_network.init(q_key, obs),
+        target_params=q_network.init(q_key, obs),
+        tx=optax.adam(learning_rate=args.learning_rate),
+    )
+
+    q_network.apply = jax.jit(q_network.apply)
+    # This step is not necessary as init called on same observation and key will always lead to same initializations
+    q_state = q_state.replace(target_params=optax.incremental_update(q_state.params, q_state.target_params, 1))
+
+    rb = ReplayBuffer(
+        args.buffer_size,
+        envs.single_observation_space,
+        envs.single_action_space,
+        "cpu",
+        handle_timeout_termination=True,
+    )
+
+    @jax.jit
+    def update(q_state, observations, actions, next_observations, rewards, dones):
+        q_next_target = q_network.apply(q_state.target_params, next_observations)  # (batch_size, num_actions)
+        q_next_target = jnp.max(q_next_target, axis=-1)  # (batch_size,)
+        next_q_value = rewards + (1 - dones) * args.gamma * q_next_target
+
+        def mse_loss(params):
+            q_pred = q_network.apply(params, observations)  # (batch_size, num_actions)
+            q_pred = q_pred[np.arange(q_pred.shape[0]), actions.squeeze()]  # (batch_size,)
+            return ((q_pred - next_q_value) ** 2).mean(), q_pred
+
+        (loss_value, q_pred), grads = jax.value_and_grad(mse_loss, has_aux=True)(q_state.params)
+        q_state = q_state.apply_gradients(grads=grads)
+        return loss_value, q_pred, q_state
+
+    start_time = time.time()
+
+    # TRY NOT TO MODIFY: start the game
+    obs = envs.reset()
+    for global_step in range(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.apply(q_state.params, obs)
+            actions = q_values.argmax(axis=-1)
+            actions = jax.device_get(actions)
+
+        # TRY NOT TO MODIFY: execute the game and log data.
+        next_obs, rewards, dones, infos = envs.step(actions)
+
+        # TRY NOT TO MODIFY: record rewards for plotting purposes
+        for info in infos:
+            if "episode" in info.keys():
+                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)
+                break
+
+        # TRY NOT TO MODIFY: save data to reply buffer; handle `terminal_observation`
+        real_next_obs = next_obs.copy()
+        for idx, d in enumerate(dones):
+            if d:
+                real_next_obs[idx] = infos[idx]["terminal_observation"]
+        rb.add(obs, real_next_obs, actions, rewards, dones, infos)
+
+        # TRY NOT TO MODIFY: CRUCIAL step easy to overlook
+        obs = next_obs
+
+        # ALGO LOGIC: training.
+        if global_step > args.learning_starts and global_step % args.train_frequency == 0:
+            data = rb.sample(args.batch_size)
+            # perform a gradient-descent step
+            loss, old_val, q_state = update(
+                q_state,
+                data.observations.numpy(),
+                data.actions.numpy(),
+                data.next_observations.numpy(),
+                data.rewards.flatten().numpy(),
+                data.dones.flatten().numpy(),
+            )
+
+            if global_step % 100 == 0:
+                writer.add_scalar("losses/td_loss", jax.device_get(loss), global_step)
+                writer.add_scalar("losses/q_values", jax.device_get(old_val).mean(), 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)
+
+            # update the target network
+            if global_step % args.target_network_frequency == 0:
+                q_state = q_state.replace(target_params=optax.incremental_update(q_state.params, q_state.target_params, 1))
+
+    if args.save_model:
+        model_path = f"runs/{run_name}/{args.exp_name}.cleanrl_model"
+        with open(model_path, "wb") as f:
+            f.write(flax.serialization.to_bytes(q_state.params))
+        print(f"model saved to {model_path}")
+        from cleanrl_utils.evals.dqn_jax_eval import evaluate
+
+        episodic_returns = evaluate(
+            model_path,
+            make_env,
+            args.env_id,
+            eval_episodes=10,
+            run_name=f"{run_name}-eval",
+            Model=QNetwork,
+            epsilon=0.05,
+        )
+        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, "DQN", f"runs/{run_name}", f"videos/{run_name}-eval")
+
+    envs.close()
+    writer.close()

events.out.tfevents.1668717427.pop-os.3662910.0

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4b64cefe331992e4359953c3c8e93dd8b08932f76211005261cff85d27ae84a2
+size 1511787

pyproject.toml

@@ -0,0 +1,96 @@
+[tool.poetry]
+name = "cleanrl"
+version = "1.0.0"
+description = "High-quality single file implementation of Deep Reinforcement Learning algorithms with research-friendly features"
+authors = ["Costa Huang <costa.huang@outlook.com>"]
+include = ["cleanrl_utils"]
+keywords = ["reinforcement", "machine", "learning", "research"]
+license="MIT"
+readme = "README.md"
+
+[tool.poetry.dependencies]
+python = ">=3.7.1,<3.10"
+tensorboard = "^2.10.0"
+wandb = "^0.13.3"
+gym = {version = "0.23.1", extras = ["classic_control"]}
+torch = "^1.12.1"
+stable-baselines3 = "1.2.0"
+
+[tool.poetry.group.dev.dependencies]
+pre-commit = "^2.20.0"
+
+[tool.poetry.group.atari]
+optional = true
+[tool.poetry.group.atari.dependencies]
+ale-py = "0.7.4"
+AutoROM = {extras = ["accept-rom-license"], version = "^0.4.2"}
+opencv-python = "^4.6.0.66"
+
+[tool.poetry.group.pybullet]
+optional = true
+[tool.poetry.group.pybullet.dependencies]
+pybullet = "3.1.8"
+
+[tool.poetry.group.procgen]
+optional = true
+[tool.poetry.group.procgen.dependencies]
+procgen = "^0.10.7"
+
+[tool.poetry.group.pytest]
+optional = true
+[tool.poetry.group.pytest.dependencies]
+pytest = "^7.1.3"
+
+[tool.poetry.group.mujoco]
+optional = true
+[tool.poetry.group.mujoco.dependencies]
+free-mujoco-py = "^2.1.6"
+
+[tool.poetry.group.docs]
+optional = true
+[tool.poetry.group.docs.dependencies]
+mkdocs-material = "^8.4.3"
+markdown-include = "^0.7.0"
+
+[tool.poetry.group.jax]
+optional = true
+[tool.poetry.group.jax.dependencies]
+jax = "^0.3.17"
+jaxlib = "^0.3.15"
+flax = "^0.6.0"
+
+[tool.poetry.group.optuna]
+optional = true
+[tool.poetry.group.optuna.dependencies]
+optuna = "^3.0.1"
+optuna-dashboard = "^0.7.2"
+rich = "<12.0"
+
+[tool.poetry.group.envpool]
+optional = true
+[tool.poetry.group.envpool.dependencies]
+envpool = "^0.6.4"
+
+[tool.poetry.group.pettingzoo]
+optional = true
+[tool.poetry.group.pettingzoo.dependencies]
+PettingZoo = "1.18.1"
+SuperSuit = "3.4.0"
+multi-agent-ale-py = "0.1.11"
+
+
+[tool.poetry.group.cloud]
+optional = true
+[tool.poetry.group.cloud.dependencies]
+boto3 = "^1.24.70"
+awscli = "^1.25.71"
+
+[tool.poetry.group.isaacgym]
+optional = true
+[tool.poetry.group.isaacgym.dependencies]
+isaacgymenvs = {git = "https://github.com/vwxyzjn/IsaacGymEnvs.git", rev = "poetry"}
+isaacgym = {path = "cleanrl/ppo_continuous_action_isaacgym/isaacgym", develop = true}
+
+[build-system]
+requires = ["poetry-core"]
+build-backend = "poetry.core.masonry.api"