pushing model

.gitattributes

@@ -32,3 +32,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+DQPN_p1.cleanrl_model filter=lfs diff=lfs merge=lfs -text

DQPN_p1.cleanrl_model

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d1190bc028a2a0418cf30d35e65d57d9beb5c638a00187b9a84c1f9d72a304f2
+size 6751815

README.md

@@ -0,0 +1,83 @@
+---
+tags:
+- Pong-v4
+- deep-reinforcement-learning
+- reinforcement-learning
+- custom-implementation
+library_name: cleanrl
+model-index:
+- name: DQN
+  results:
+  - task:
+      type: reinforcement-learning
+      name: reinforcement-learning
+    dataset:
+      name: Pong-v4
+      type: Pong-v4
+    metrics:
+    - type: mean_reward
+      value: 2.70 +/- 7.47
+      name: mean_reward
+      verified: false
+---
+
+# (CleanRL) **DQN** Agent Playing **Pong-v4**
+
+This is a trained model of a DQN agent playing Pong-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/DQPN_p1.py).
+
+## Get Started
+
+To use this model, please install the `cleanrl` package with the following command:
+
+```
+pip install "cleanrl[DQPN_p1]"
+python -m cleanrl_utils.enjoy --exp-name DQPN_p1 --env-id Pong-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/pfunk/Pong-v4-DQPN_p1-seed1/raw/main/dqpn_atari.py
+curl -OL https://huggingface.co/pfunk/Pong-v4-DQPN_p1-seed1/raw/main/pyproject.toml
+curl -OL https://huggingface.co/pfunk/Pong-v4-DQPN_p1-seed1/raw/main/poetry.lock
+poetry install --all-extras
+python dqpn_atari.py --exp-name DQPN_p1 --start-policy-f 1000 --end-policy-f 1000 --evaluation-fraction 1.00 --target-tau 1.0 --policy-tau 1.00 --track --wandb-entity pfunk --wandb-project-name dqpn --save-model true --upload-model true --hf-entity pfunk --env-id Pong-v4 --seed 1 --total-timesteps 10000000
+```
+
+# Hyperparameters
+```python
+{'batch_size': 32,
+ 'buffer_size': 1000000,
+ 'capture_video': False,
+ 'cuda': True,
+ 'end_e': 0.01,
+ 'end_policy_f': 1000,
+ 'env_id': 'Pong-v4',
+ 'evaluation_fraction': 1.0,
+ 'exp_name': 'DQPN_p1',
+ 'exploration_fraction': 0.1,
+ 'gamma': 0.99,
+ 'hf_entity': 'pfunk',
+ 'learning_rate': 0.0001,
+ 'learning_starts': 80000,
+ 'policy_tau': 1.0,
+ 'save_model': True,
+ 'seed': 1,
+ 'start_e': 1,
+ 'start_policy_f': 1000,
+ 'target_network_frequency': 1000,
+ 'target_tau': 1.0,
+ 'torch_deterministic': True,
+ 'total_timesteps': 10000000,
+ 'track': True,
+ 'train_frequency': 4,
+ 'upload_model': True,
+ 'wandb_entity': 'pfunk',
+ 'wandb_project_name': 'dqpn'}
+```
+    

dqpn_atari.py

@@ -0,0 +1,295 @@
+# 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 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
+
+
+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="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("--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("--target-tau", type=float, default=1.,
+        help="the target network update rate")
+    parser.add_argument("--policy-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("--start-policy-f", type=int, default=5000,
+        help="the starting timesteps it takes to update the policy network")
+    parser.add_argument("--end-policy-f", type=int, default=5000,
+        help="the ending timesteps it takes to update the policy 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("--evaluation-fraction", type=float, default=0.10,
+        help="the fraction of `total-timesteps` it takes from start-policy-f to go end-policy-f")
+    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")
+    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 = 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)
+        env = gym.wrappers.FrameStack(env, 4)
+        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):
+    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)
+
+
+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, args.seed, 0, args.capture_video, run_name)])
+    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)
+    policy_network = QNetwork(envs).to(device)
+    target_network.load_state_dict(q_network.state_dict())
+    policy_network.load_state_dict(q_network.state_dict())
+    policy_network_frequency = args.start_policy_f
+
+    rb = ReplayBuffer(
+        args.buffer_size,
+        envs.single_observation_space,
+        envs.single_action_space,
+        device,
+        optimize_memory_usage=True,
+        handle_timeout_termination=True,
+    )
+    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 = policy_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, 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:
+            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.target_tau * q_network_param.data + (1.0 - args.target_tau) * target_network_param.data
+                    )
+
+            # update policy network
+            if global_step % policy_network_frequency == 0:
+                for policy_network_param, q_network_param in zip(policy_network.parameters(), q_network.parameters()):
+                    policy_network_param.data.copy_(
+                        args.policy_tau * q_network_param.data + (1.0 - args.policy_tau) * policy_network_param.data
+                    )
+                for target_network_param, q_network_param in zip(target_network.parameters(), q_network.parameters()):
+                    target_network_param.data.copy_(
+                        args.target_tau * q_network_param.data + (1.0 - args.target_tau) * target_network_param.data
+                    )
+                policy_network_frequency = int(linear_schedule(
+                    args.start_policy_f, args.end_policy_f,
+                    args.evaluation_fraction * args.total_timesteps, global_step))
+                # print(args.policy_network_frequency)
+
+    if args.save_model:
+        model_path = f"runs/{run_name}/{args.exp_name}.cleanrl_model"
+        torch.save(policy_network.state_dict(), model_path)
+        print(f"model saved to {model_path}")
+        from cleanrl_utils.evals.dqn_eval import evaluate
+
+        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:
+            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.1675623355.cronus.78788.0

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1de15e1fd8b15b09ae92ea372a3aad704c3ed2aa208760ebbf8949fc0237bba8
+size 17709739

pyproject.toml

@@ -0,0 +1,178 @@
+[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,<3.10"
+tensorboard = "^2.10.0"
+wandb = "^0.13.6"
+gym = "0.23.1"
+torch = ">=1.12.1"
+stable-baselines3 = "1.2.0"
+gymnasium = "^0.26.3"
+moviepy = "^1.0.3"
+pygame = "2.1.0"
+huggingface-hub = "^0.11.1"
+
+ale-py = {version = "0.7.4", optional = true}
+AutoROM = {extras = ["accept-rom-license"], version = "^0.4.2"}
+opencv-python = {version = "^4.6.0.66", optional = true}
+pybullet = {version = "3.1.8", optional = true}
+procgen = {version = "^0.10.7", optional = true}
+pytest = {version = "^7.1.3", optional = true}
+mujoco = {version = "^2.2", 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}
+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}
+rich = {version = "<12.0", 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 = "^0.1.0", optional = true}
+dm-control = {version = "^1.0.8", optional = true}
+
+[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]
+mujoco = "^2.2"
+imageio = "^2.14.1"
+
+[tool.poetry.group.mujoco_py]
+optional = true
+[tool.poetry.group.mujoco_py.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}
+
+[tool.poetry.group.dm_control]
+optional = true
+[tool.poetry.group.dm_control.dependencies]
+shimmy = "^0.1.0"
+dm-control = "^1.0.8"
+mujoco = "^2.2"
+
+[build-system]
+requires = ["poetry-core"]
+build-backend = "poetry.core.masonry.api"
+
+[tool.poetry.extras]
+atari = ["ale-py", "AutoROM", "opencv-python"]
+pybullet = ["pybullet"]
+procgen = ["procgen"]
+plot = ["pandas", "seaborn"]
+pytest = ["pytest"]
+mujoco = ["mujoco", "imageio"]
+mujoco_py = ["free-mujoco-py"]
+jax = ["jax", "jaxlib", "flax"]
+docs = ["mkdocs-material", "markdown-include"]
+envpool = ["envpool"]
+optuna = ["optuna", "optuna-dashboard", "rich"]
+pettingzoo = ["PettingZoo", "SuperSuit", "multi-agent-ale-py"]
+cloud = ["boto3", "awscli"]
+dm_control = ["shimmy", "dm-control", "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
+]