(CleanRL) PPO Agent Playing Hero-v5
This is a trained model of a PPO agent playing Hero-v5. The model was trained by using CleanRL and the most up-to-date training code can be found here.
Get Started
To use this model, please install the cleanrl package with the following command:
pip install "cleanrl[jax,envpool,atari]"
python -m cleanrl_utils.enjoy --exp-name cleanba_ppo_envpool_impala_atari_wrapper --env-id Hero-v5
Please refer to the documentation for more detail.
Command to reproduce the training
curl -OL https://huggingface.co/cleanrl/Hero-v5-cleanba_ppo_envpool_impala_atari_wrapper-seed3/raw/main/cleanba_ppo_envpool_impala_atari_wrapper.py
curl -OL https://huggingface.co/cleanrl/Hero-v5-cleanba_ppo_envpool_impala_atari_wrapper-seed3/raw/main/pyproject.toml
curl -OL https://huggingface.co/cleanrl/Hero-v5-cleanba_ppo_envpool_impala_atari_wrapper-seed3/raw/main/poetry.lock
poetry install --all-extras
python cleanba_ppo_envpool_impala_atari_wrapper.py --distributed --learner-device-ids 1 2 3 --track --wandb-project-name cleanba --save-model --upload-model --hf-entity cleanrl --env-id Hero-v5 --seed 3
Hyperparameters
{'actor_device_ids': [0],
'actor_devices': ['gpu:0'],
'anneal_lr': True,
'async_batch_size': 20,
'async_update': 3,
'batch_size': 15360,
'capture_video': False,
'clip_coef': 0.1,
'concurrency': True,
'cuda': True,
'distributed': True,
'ent_coef': 0.01,
'env_id': 'Hero-v5',
'exp_name': 'cleanba_ppo_envpool_impala_atari_wrapper',
'gae_lambda': 0.95,
'gamma': 0.99,
'global_learner_decices': ['gpu:1',
'gpu:2',
'gpu:3',
'gpu:5',
'gpu:6',
'gpu:7'],
'hf_entity': 'cleanrl',
'learner_device_ids': [1, 2, 3],
'learner_devices': ['gpu:1', 'gpu:2', 'gpu:3'],
'learning_rate': 0.00025,
'local_batch_size': 7680,
'local_minibatch_size': 1920,
'local_num_envs': 60,
'local_rank': 0,
'max_grad_norm': 0.5,
'minibatch_size': 3840,
'norm_adv': True,
'num_envs': 120,
'num_minibatches': 4,
'num_steps': 128,
'num_updates': 3255,
'profile': False,
'save_model': True,
'seed': 3,
'target_kl': None,
'test_actor_learner_throughput': False,
'torch_deterministic': True,
'total_timesteps': 50000000,
'track': True,
'update_epochs': 4,
'upload_model': True,
'vf_coef': 0.5,
'wandb_entity': None,
'wandb_project_name': 'cleanba',
'world_size': 2}