pushing model

README.md

@@ -0,0 +1,78 @@
+---
+tags:
+- Hopper-v5
+- deep-reinforcement-learning
+- reinforcement-learning
+- custom-implementation
+library_name: cleanrl
+model-index:
+- name: TD3
+  results:
+  - task:
+      type: reinforcement-learning
+      name: reinforcement-learning
+    dataset:
+      name: Hopper-v5
+      type: Hopper-v5
+    metrics:
+    - type: mean_reward
+      value: 1329.44 +/- 156.77
+      name: mean_reward
+      verified: false
+---
+
+# (CleanRL) **TD3** Agent Playing **Hopper-v5**
+
+This is a trained model of a TD3 agent playing Hopper-v5.
+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/td3.py).
+
+## Get Started
+
+To use this model, please install the `cleanrl` package with the following command:
+
+```
+pip install "cleanrl[td3]"
+python -m cleanrl_utils.enjoy --exp-name td3 --env-id Hopper-v5
+```
+
+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/jacksonhack/Hopper-v5-td3-seed1/raw/main/td3.py
+curl -OL https://huggingface.co/jacksonhack/Hopper-v5-td3-seed1/raw/main/pyproject.toml
+curl -OL https://huggingface.co/jacksonhack/Hopper-v5-td3-seed1/raw/main/poetry.lock
+poetry install --all-extras
+python td3.py --save_model --upload_model --track
+```
+
+# Hyperparameters
+```python
+{'batch_size': 256,
+ 'buffer_size': 1000000,
+ 'capture_video': False,
+ 'cuda': True,
+ 'env_id': 'Hopper-v5',
+ 'exp_name': 'td3',
+ 'exploration_noise': 0.1,
+ 'gamma': 0.99,
+ 'hf_entity': 'jacksonhack',
+ 'learning_rate': 0.0003,
+ 'learning_starts': 25000.0,
+ 'noise_clip': 0.5,
+ 'policy_frequency': 2,
+ 'policy_noise': 0.2,
+ 'save_model': True,
+ 'seed': 1,
+ 'tau': 0.005,
+ 'torch_deterministic': True,
+ 'total_timesteps': 1000000,
+ 'track': True,
+ 'upload_model': True,
+ 'wandb_entity': None,
+ 'wandb_project_name': 'cleanRL'}
+```
+    

events.out.tfevents.1732713429.DESKTOP-3BC7099.129416.0

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7cd5c52ac63984c5e5f38cc2591f606271c4796402720f8b9a111df17d9b2f7f
+size 4386563

pyproject.toml

@@ -0,0 +1,33 @@
+[tool.poetry]
+name = "rl"
+version = "0.1.0"
+description = ""
+authors = ["jackson <[email protected]>"]
+readme = "README.md"
+packages = [
+    {include = "rl"},
+    {include = "rl_utils"},
+]
+
+
+[tool.poetry.dependencies]
+python = "^3.10"
+gymnasium = {extras = ["box2d"], version = "^1.0.0"}
+tensorboard = "^2.18.0"
+huggingface-hub = "^0.26.2"
+tyro = "^0.8.14"
+torch = "^2.5.1"
+stable-baselines3 = "^2.3.2"
+numpy = "^1.21.6"
+tenacity = "^9.0.0"
+mujoco = "2.3.3"
+
+
+[tool.poetry.group.dev.dependencies]
+black = "^24.10.0"
+wandb = "^0.18.7"
+moviepy = "^2.1.1"
+
+[build-system]
+requires = ["poetry-core"]
+build-backend = "poetry.core.masonry.api"

td3.py

@@ -0,0 +1,302 @@
+# docs and experiment results can be found at https://docs.cleanrl.dev/rl-algorithms/td3/#td3_continuous_actionpy
+import os
+import random
+import time
+from dataclasses import dataclass
+
+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
+import tyro
+from stable_baselines3.common.buffers import ReplayBuffer
+from torch.utils.tensorboard import SummaryWriter
+
+
+@dataclass
+class Args:
+    exp_name: str = os.path.basename(__file__)[: -len(".py")]
+    """the name of this experiment"""
+    seed: int = 1
+    """seed of the experiment"""
+    torch_deterministic: bool = True
+    """if toggled, `torch.backends.cudnn.deterministic=False`"""
+    cuda: bool = True
+    """if toggled, cuda will be enabled by default"""
+    track: bool = False
+    """if toggled, this experiment will be tracked with Weights and Biases"""
+    wandb_project_name: str = "cleanRL"
+    """the wandb's project name"""
+    wandb_entity: str = None
+    """the entity (team) of wandb's project"""
+    capture_video: bool = False
+    """whether to capture videos of the agent performances (check out `videos` folder)"""
+    save_model: bool = False
+    """whether to save model into the `runs/{run_name}` folder"""
+    upload_model: bool = False
+    """whether to upload the saved model to huggingface"""
+    hf_entity: str = "jacksonhack"
+    """the user or org name of the model repository from the Hugging Face Hub"""
+
+    # Algorithm specific arguments
+    env_id: str = "Hopper-v5"
+    """the id of the environment"""
+    total_timesteps: int = 1000000
+    """total timesteps of the experiments"""
+    learning_rate: float = 3e-4
+    """the learning rate of the optimizer"""
+    buffer_size: int = int(1e6)
+    """the replay memory buffer size"""
+    gamma: float = 0.99
+    """the discount factor gamma"""
+    tau: float = 0.005
+    """target smoothing coefficient (default: 0.005)"""
+    batch_size: int = 256
+    """the batch size of sample from the reply memory"""
+    policy_noise: float = 0.2
+    """the scale of policy noise"""
+    exploration_noise: float = 0.1
+    """the scale of exploration noise"""
+    learning_starts: int = 25e3
+    """timestep to start learning"""
+    policy_frequency: int = 2
+    """the frequency of training policy (delayed)"""
+    noise_clip: float = 0.5
+    """noise clip parameter of the Target Policy Smoothing Regularization"""
+
+
+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.action_space.seed(seed)
+        return env
+
+    return thunk
+
+
+# ALGO LOGIC: initialize agent here:
+class QNetwork(nn.Module):
+    def __init__(self, env):
+        super().__init__()
+        self.fc1 = nn.Linear(np.array(env.single_observation_space.shape).prod() + np.prod(env.single_action_space.shape), 256)
+        self.fc2 = nn.Linear(256, 256)
+        self.fc3 = nn.Linear(256, 1)
+
+    def forward(self, x, a):
+        x = torch.cat([x, a], 1)
+        x = F.relu(self.fc1(x))
+        x = F.relu(self.fc2(x))
+        x = self.fc3(x)
+        return x
+
+
+class Actor(nn.Module):
+    def __init__(self, env):
+        super().__init__()
+        self.fc1 = nn.Linear(np.array(env.single_observation_space.shape).prod(), 256)
+        self.fc2 = nn.Linear(256, 256)
+        self.fc_mu = nn.Linear(256, np.prod(env.single_action_space.shape))
+        # action rescaling
+        self.register_buffer(
+            "action_scale", torch.tensor((env.action_space.high - env.action_space.low) / 2.0, dtype=torch.float32)
+        )
+        self.register_buffer(
+            "action_bias", torch.tensor((env.action_space.high + env.action_space.low) / 2.0, dtype=torch.float32)
+        )
+
+    def forward(self, x):
+        x = F.relu(self.fc1(x))
+        x = F.relu(self.fc2(x))
+        x = torch.tanh(self.fc_mu(x))
+        return x * self.action_scale + self.action_bias
+
+
+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 = tyro.cli(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=False,
+            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, 0, args.capture_video, run_name)])
+    assert isinstance(envs.single_action_space, gym.spaces.Box), "only continuous action space is supported"
+
+    actor = Actor(envs).to(device)
+    qf1 = QNetwork(envs).to(device)
+    qf2 = QNetwork(envs).to(device)
+    qf1_target = QNetwork(envs).to(device)
+    qf2_target = QNetwork(envs).to(device)
+    target_actor = Actor(envs).to(device)
+    target_actor.load_state_dict(actor.state_dict())
+    qf1_target.load_state_dict(qf1.state_dict())
+    qf2_target.load_state_dict(qf2.state_dict())
+    q_optimizer = optim.Adam(list(qf1.parameters()) + list(qf2.parameters()), lr=args.learning_rate)
+    actor_optimizer = optim.Adam(list(actor.parameters()), lr=args.learning_rate)
+
+    envs.single_observation_space.dtype = np.float32
+    rb = ReplayBuffer(
+        args.buffer_size,
+        envs.single_observation_space,
+        envs.single_action_space,
+        device,
+        handle_timeout_termination=False,
+    )
+    start_time = time.time()
+
+    # TRY NOT TO MODIFY: start the game
+    obs, _ = envs.reset(seed=args.seed)
+    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:
+            with torch.no_grad():
+                actions = actor(torch.Tensor(obs).to(device))
+                actions += torch.normal(0, actor.action_scale * args.exploration_noise)
+                actions = actions.cpu().numpy().clip(envs.single_action_space.low, envs.single_action_space.high)
+
+        # TRY NOT TO MODIFY: execute the game and log data.
+        next_obs, rewards, terminations, truncations, 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
+        
+        if "episode" in infos:
+            print(f"global_step={global_step}, episode_return={infos['episode']['r'][infos['_episode']][0]}")
+            writer.add_scalar("charts/episodic_return", infos["episode"]["r"][infos["_episode"]][0], global_step)
+            writer.add_scalar("charts/episodic_length", infos["episode"]["l"][infos["_episode"]][0], global_step)
+
+        # TRY NOT TO MODIFY: save data to reply buffer; handle `final_observation`
+        real_next_obs = next_obs.copy()
+        # for idx, trunc in enumerate(truncations):
+        #     if trunc:
+        #         real_next_obs[idx] = infos["final_observation"][idx]
+
+        rb.add(obs, real_next_obs, actions, rewards, terminations, 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)
+            with torch.no_grad():
+                clipped_noise = (torch.randn_like(data.actions, device=device) * args.policy_noise).clamp(
+                    -args.noise_clip, args.noise_clip
+                ) * target_actor.action_scale
+
+                next_state_actions = (target_actor(data.next_observations) + clipped_noise).clamp(
+                    envs.single_action_space.low[0], envs.single_action_space.high[0]
+                )
+                qf1_next_target = qf1_target(data.next_observations, next_state_actions)
+                qf2_next_target = qf2_target(data.next_observations, next_state_actions)
+                min_qf_next_target = torch.min(qf1_next_target, qf2_next_target)
+                next_q_value = data.rewards.flatten() + (1 - data.dones.flatten()) * args.gamma * (min_qf_next_target).view(-1)
+
+            qf1_a_values = qf1(data.observations, data.actions).view(-1)
+            qf2_a_values = qf2(data.observations, data.actions).view(-1)
+            qf1_loss = F.mse_loss(qf1_a_values, next_q_value)
+            qf2_loss = F.mse_loss(qf2_a_values, next_q_value)
+            qf_loss = qf1_loss + qf2_loss
+
+            # optimize the model
+            q_optimizer.zero_grad()
+            qf_loss.backward()
+            q_optimizer.step()
+
+            if global_step % args.policy_frequency == 0:
+                actor_loss = -qf1(data.observations, actor(data.observations)).mean()
+                actor_optimizer.zero_grad()
+                actor_loss.backward()
+                actor_optimizer.step()
+
+                # update the target network
+                for param, target_param in zip(actor.parameters(), target_actor.parameters()):
+                    target_param.data.copy_(args.tau * param.data + (1 - args.tau) * target_param.data)
+                for param, target_param in zip(qf1.parameters(), qf1_target.parameters()):
+                    target_param.data.copy_(args.tau * param.data + (1 - args.tau) * target_param.data)
+                for param, target_param in zip(qf2.parameters(), qf2_target.parameters()):
+                    target_param.data.copy_(args.tau * param.data + (1 - args.tau) * target_param.data)
+
+            if global_step % 100 == 0:
+                writer.add_scalar("losses/qf1_values", qf1_a_values.mean().item(), global_step)
+                writer.add_scalar("losses/qf2_values", qf2_a_values.mean().item(), global_step)
+                writer.add_scalar("losses/qf1_loss", qf1_loss.item(), global_step)
+                writer.add_scalar("losses/qf2_loss", qf2_loss.item(), global_step)
+                writer.add_scalar("losses/qf_loss", qf_loss.item() / 2.0, global_step)
+                writer.add_scalar("losses/actor_loss", actor_loss.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"
+        torch.save((actor.state_dict(), qf1.state_dict(), qf2.state_dict()), model_path)
+        print(f"model saved to {model_path}")
+        from rl_utils.evals.td3_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),
+            device=device,
+            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 rl_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, "TD3", f"runs/{run_name}", f"videos/{run_name}-eval")
+
+    envs.close()
+    writer.close()