pushing model

README.md

@@ -0,0 +1,75 @@
+---
+tags:
+- InvertedPendulum-v2
+- deep-reinforcement-learning
+- reinforcement-learning
+- custom-implementation
+library_name: cleanrl
+model-index:
+- name: DDPG
+  results:
+  - task:
+      type: reinforcement-learning
+      name: reinforcement-learning
+    dataset:
+      name: InvertedPendulum-v2
+      type: InvertedPendulum-v2
+    metrics:
+    - type: mean_reward
+      value: 709.70 +/- 443.44
+      name: mean_reward
+      verified: false
+---
+
+# (CleanRL) **DDPG** Agent Playing **InvertedPendulum-v2**
+
+This is a trained model of a DDPG agent playing InvertedPendulum-v2.
+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/ddpg_continuous_action_jax.py).
+
+## Get Started
+
+To use this model, please install the `cleanrl` package with the following command:
+
+```
+pip install "cleanrl[ddpg_continuous_action_jax]"
+python -m cleanrl_utils.enjoy --exp-name ddpg_continuous_action_jax --env-id InvertedPendulum-v2
+```
+
+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/InvertedPendulum-v2-ddpg_continuous_action_jax-seed1/raw/main/ddpg_continuous_action_jax.py
+curl -OL https://huggingface.co/cleanrl/InvertedPendulum-v2-ddpg_continuous_action_jax-seed1/raw/main/pyproject.toml
+curl -OL https://huggingface.co/cleanrl/InvertedPendulum-v2-ddpg_continuous_action_jax-seed1/raw/main/poetry.lock
+poetry install --all-extras
+python ddpg_continuous_action_jax.py --track --capture-video --save-model --hf-entity cleanrl --upload-mode --env-id InvertedPendulum-v2 --seed 1
+```
+
+# Hyperparameters
+```python
+{'batch_size': 256,
+ 'buffer_size': 1000000,
+ 'capture_video': True,
+ 'env_id': 'InvertedPendulum-v2',
+ 'exp_name': 'ddpg_continuous_action_jax',
+ 'exploration_noise': 0.1,
+ 'gamma': 0.99,
+ 'hf_entity': 'cleanrl',
+ 'learning_rate': 0.0003,
+ 'learning_starts': 25000.0,
+ 'noise_clip': 0.5,
+ 'policy_frequency': 2,
+ 'save_model': True,
+ 'seed': 1,
+ 'tau': 0.005,
+ 'total_timesteps': 1000000,
+ 'track': True,
+ 'upload_model': True,
+ 'wandb_entity': None,
+ 'wandb_project_name': 'cleanRL'}
+```
+    

ddpg_continuous_action_jax.py

@@ -0,0 +1,335 @@
+# docs and experiment results can be found at https://docs.cleanrl.dev/rl-algorithms/ddpg/#ddpg_continuous_action_jaxpy
+import argparse
+import os
+import random
+import time
+from distutils.util import strtobool
+from typing import Sequence
+
+import flax
+import flax.linen as nn
+import gymnasium as 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="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("--buffer-size", type=int, default=int(1e6),
+        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=0.005,
+        help="target smoothing coefficient (default: 0.005)")
+    parser.add_argument("--batch-size", type=int, default=256,
+        help="the batch size of sample from the reply memory")
+    parser.add_argument("--exploration-noise", type=float, default=0.1,
+        help="the scale of exploration noise")
+    parser.add_argument("--learning-starts", type=int, default=25e3,
+        help="timestep to start learning")
+    parser.add_argument("--policy-frequency", type=int, default=2,
+        help="the frequency of training policy (delayed)")
+    parser.add_argument("--noise-clip", type=float, default=0.5,
+        help="noise clip parameter of the Target Policy Smoothing Regularization")
+    args = parser.parse_args()
+    # fmt: on
+    return args
+
+
+def make_env(env_id, seed, idx, capture_video, run_name):
+    def thunk():
+        if capture_video:
+            env = gym.make(env_id, render_mode="rgb_array")
+        else:
+            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.action_space.seed(seed)
+        env.observation_space.seed(seed)
+        return env
+
+    return thunk
+
+
+# ALGO LOGIC: initialize agent here:
+class QNetwork(nn.Module):
+    @nn.compact
+    def __call__(self, x: jnp.ndarray, a: jnp.ndarray):
+        x = jnp.concatenate([x, a], -1)
+        x = nn.Dense(256)(x)
+        x = nn.relu(x)
+        x = nn.Dense(256)(x)
+        x = nn.relu(x)
+        x = nn.Dense(1)(x)
+        return x
+
+
+class Actor(nn.Module):
+    action_dim: Sequence[int]
+    action_scale: Sequence[int]
+    action_bias: Sequence[int]
+
+    @nn.compact
+    def __call__(self, x):
+        x = nn.Dense(256)(x)
+        x = nn.relu(x)
+        x = nn.Dense(256)(x)
+        x = nn.relu(x)
+        x = nn.Dense(self.action_dim)(x)
+        x = nn.tanh(x)
+        x = x * self.action_scale + self.action_bias
+        return x
+
+
+class TrainState(TrainState):
+    target_params: flax.core.FrozenDict
+
+
+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"
+"""
+        )
+    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, actor_key, qf1_key = jax.random.split(key, 3)
+
+    # 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.Box), "only continuous action space is supported"
+
+    max_action = float(envs.single_action_space.high[0])
+    envs.single_observation_space.dtype = np.float32
+    rb = ReplayBuffer(
+        args.buffer_size,
+        envs.single_observation_space,
+        envs.single_action_space,
+        device="cpu",
+        handle_timeout_termination=False,
+    )
+
+    # TRY NOT TO MODIFY: start the game
+    obs, _ = envs.reset()
+
+    action_scale = np.array((envs.action_space.high - envs.action_space.low) / 2.0)
+    action_bias = np.array((envs.action_space.high + envs.action_space.low) / 2.0)
+    actor = Actor(
+        action_dim=np.prod(envs.single_action_space.shape),
+        action_scale=action_scale,
+        action_bias=action_bias,
+    )
+    qf1 = QNetwork()
+    actor_state = TrainState.create(
+        apply_fn=actor.apply,
+        params=actor.init(actor_key, obs),
+        target_params=actor.init(actor_key, obs),
+        tx=optax.adam(learning_rate=args.learning_rate),
+    )
+    qf1_state = TrainState.create(
+        apply_fn=qf1.apply,
+        params=qf1.init(qf1_key, obs, envs.action_space.sample()),
+        target_params=qf1.init(qf1_key, obs, envs.action_space.sample()),
+        tx=optax.adam(learning_rate=args.learning_rate),
+    )
+    actor.apply = jax.jit(actor.apply)
+    qf1.apply = jax.jit(qf1.apply)
+
+    @jax.jit
+    def update_critic(
+        actor_state: TrainState,
+        qf1_state: TrainState,
+        observations: np.ndarray,
+        actions: np.ndarray,
+        next_observations: np.ndarray,
+        rewards: np.ndarray,
+        dones: np.ndarray,
+    ):
+        next_state_actions = (actor.apply(actor_state.target_params, next_observations)).clip(-1, 1)  # TODO: proper clip
+        qf1_next_target = qf1.apply(qf1_state.target_params, next_observations, next_state_actions).reshape(-1)
+        next_q_value = (rewards + (1 - dones) * args.gamma * (qf1_next_target)).reshape(-1)
+
+        def mse_loss(params):
+            qf1_a_values = qf1.apply(params, observations, actions).squeeze()
+            return ((qf1_a_values - next_q_value) ** 2).mean(), qf1_a_values.mean()
+
+        (qf1_loss_value, qf1_a_values), grads = jax.value_and_grad(mse_loss, has_aux=True)(qf1_state.params)
+        qf1_state = qf1_state.apply_gradients(grads=grads)
+        return qf1_state, qf1_loss_value, qf1_a_values
+
+    @jax.jit
+    def update_actor(
+        actor_state: TrainState,
+        qf1_state: TrainState,
+        observations: np.ndarray,
+    ):
+        def actor_loss(params):
+            return -qf1.apply(qf1_state.params, observations, actor.apply(params, observations)).mean()
+
+        actor_loss_value, grads = jax.value_and_grad(actor_loss)(actor_state.params)
+        actor_state = actor_state.apply_gradients(grads=grads)
+        actor_state = actor_state.replace(
+            target_params=optax.incremental_update(actor_state.params, actor_state.target_params, args.tau)
+        )
+        qf1_state = qf1_state.replace(
+            target_params=optax.incremental_update(qf1_state.params, qf1_state.target_params, args.tau)
+        )
+        return actor_state, qf1_state, actor_loss_value
+
+    start_time = time.time()
+    for global_step in range(args.total_timesteps):
+        # ALGO LOGIC: put action logic here
+        if global_step < args.learning_starts:
+            actions = np.array([envs.single_action_space.sample() for _ in range(envs.num_envs)])
+        else:
+            actions = actor.apply(actor_state.params, obs)
+            actions = np.array(
+                [
+                    (jax.device_get(actions)[0] + np.random.normal(0, action_scale * args.exploration_noise)[0]).clip(
+                        envs.single_action_space.low, envs.single_action_space.high
+                    )
+                ]
+            )
+
+        # TRY NOT TO MODIFY: execute the game and log data.
+        next_obs, rewards, terminateds, truncateds, infos = envs.step(actions)
+
+        # TRY NOT TO MODIFY: record rewards for plotting purposes
+        if "final_info" in infos:
+            for info in infos["final_info"]:
+                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)
+                break
+
+        # TRY NOT TO MODIFY: save data to reply buffer; handle `terminal_observation`
+        real_next_obs = next_obs.copy()
+        for idx, d in enumerate(truncateds):
+            if d:
+                real_next_obs[idx] = infos["final_observation"][idx]
+        rb.add(obs, real_next_obs, actions, rewards, terminateds, infos)
+
+        # TRY NOT TO MODIFY: CRUCIAL step easy to overlook
+        obs = next_obs
+
+        # ALGO LOGIC: training.
+        if global_step > args.learning_starts:
+            data = rb.sample(args.batch_size)
+            qf1_state, qf1_loss_value, qf1_a_values = update_critic(
+                actor_state,
+                qf1_state,
+                data.observations.numpy(),
+                data.actions.numpy(),
+                data.next_observations.numpy(),
+                data.rewards.flatten().numpy(),
+                data.dones.flatten().numpy(),
+            )
+            if global_step % args.policy_frequency == 0:
+                actor_state, qf1_state, actor_loss_value = update_actor(
+                    actor_state,
+                    qf1_state,
+                    data.observations.numpy(),
+                )
+
+            if global_step % 100 == 0:
+                writer.add_scalar("losses/qf1_loss", qf1_loss_value.item(), global_step)
+                writer.add_scalar("losses/actor_loss", actor_loss_value.item(), global_step)
+                writer.add_scalar("losses/qf1_values", qf1_a_values.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)
+
+    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(
+                    [
+                        actor_state.params,
+                        qf1_state.params,
+                    ]
+                )
+            )
+        print(f"model saved to {model_path}")
+        from cleanrl_utils.evals.ddpg_jax_eval import evaluate
+
+        episodic_returns = evaluate(
+            model_path,
+            make_env,
+            args.env_id,
+            eval_episodes=10,
+            run_name=f"{run_name}-eval",
+            Model=(Actor, QNetwork),
+            exploration_noise=args.exploration_noise,
+        )
+        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, "DDPG", f"runs/{run_name}", f"videos/{run_name}-eval")
+
+    envs.close()
+    writer.close()

events.out.tfevents.1688055374.3090-172.1331626.0

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2024f38367e87f0c0b84dbfb9fded590ed202bb8cce73ffc5e3174e02435ee36
+size 2931277

pyproject.toml

@@ -0,0 +1,113 @@
+[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 <[email protected]>"]
+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
+]
+
+[[tool.poetry.source]]
+name = "tsinghua"
+url = "https://pypi.tuna.tsinghua.edu.cn/simple"
+default = true