Merge pull request #293 from huangshiyu13/main

update atari

examples/atari/README.md

@@ -8,12 +8,19 @@ Then install auto-rom via:
 or:
 ```shell
 pip install autorom
-
 AutoROM --accept-license
 ```
 
+or, if you can not download the ROMs, you can download them manually from [Google Drive](https://drive.google.com/file/d/1agerLX3fP2YqUCcAkMF7v_ZtABAOhlA7/view?usp=sharing).
+Then, you can install the ROMs via:
+```shell
+pip install autorom
+AutoROM --source-file <path-to-Roms.tar.gz>
+````
+
+
 ## Usage
 
 ```shell
-python train_ppo.py --config atari_ppo.yaml
+python train_ppo.py
 ```

examples/atari/atari_ppo.yaml

@@ -2,22 +2,27 @@ seed: 0
 lr: 2.5e-4
 critic_lr: 2.5e-4
 episode_length: 128
-ppo_epoch: 4
+gamma: 0.99
+ppo_epoch: 3
 gain: 0.01
 use_linear_lr_decay: true
 use_share_model: true
 entropy_coef: 0.01
 hidden_size: 512
-num_mini_batch: 4
-clip_param: 0.1
+num_mini_batch: 8
+clip_param: 0.2
 value_loss_coef: 0.5
+max_grad_norm: 10
+
 run_dir: ./run_results/
-experiment_name: atari_ppo
+
 log_interval: 1
 use_recurrent_policy: false
 use_valuenorm: true
 use_adv_normalize: true
+
 wandb_entity: openrl-lab
+experiment_name: atari_ppo
 
 vec_info_class:
   id: "EPS_RewardInfo"

examples/atari/train_ppo.py

@@ -43,11 +43,11 @@
 
 def train():
     cfg_parser = create_config_parser()
-    cfg = cfg_parser.parse_args()
+    cfg = cfg_parser.parse_args(["--config", "atari_ppo.yaml"])
 
     # create environment, set environment parallelism to 9
     env = make(
-        "ALE/Pong-v5", env_num=9, cfg=cfg, asynchronous=True, env_wrappers=env_wrappers
+        "ALE/Pong-v5", env_num=16, cfg=cfg, asynchronous=True, env_wrappers=env_wrappers
     )
 
     # create the neural network
@@ -56,7 +56,7 @@ def train():
         env, cfg=cfg, device="cuda" if "macOS" not in get_system_info()["OS"] else "cpu"
     )
     # initialize the trainer
-    agent = Agent(net, use_wandb=True)
+    agent = Agent(net, use_wandb=True, project_name="Pong-v5")
     # start training, set total number of training steps to 20000
 
     agent.train(total_time_steps=5000000)

openrl/configs/config.py

@@ -726,8 +726,8 @@ def create_config_parser():
     parser.add_argument(
         "--max_grad_norm",
         type=float,
-        default=10.0,
-        help="max norm of gradients (default: 0.5)",
+        default=10,
+        help="max norm of gradients (default: 10)",
     )
     parser.add_argument(
         "--use_gae",

setup.py

@@ -41,6 +41,7 @@ def get_install_requires() -> list:
         "mujoco",
         "tqdm",
         "Jinja2",
+        "pettingzoo",
     ]