Views
No views yet
1from skrl.utils.huggingface import download_model_from_huggingface
2
3# assuming that there is an agent named `agent`
4path = download_model_from_huggingface("skrl/OmniIsaacGymEnvs-AllegroHand-PPO")
5agent.load(path)1# https://skrl.readthedocs.io/en/latest/modules/skrl.agents.ppo.html#configuration-and-hyperparameters
2cfg_ppo = PPO_DEFAULT_CONFIG.copy()
3cfg_ppo["rollouts"] = 16 # memory_size
4cfg_ppo["learning_epochs"] = 5
5cfg_ppo["mini_batches"] = 4 # 16 * 8192 / 32768
6cfg_ppo["discount_factor"] = 0.99
7cfg_ppo["lambda"] = 0.95
8cfg_ppo["learning_rate"] = 5e-4
9cfg_ppo["learning_rate_scheduler"] = KLAdaptiveRL
10cfg_ppo["learning_rate_scheduler_kwargs"] = {"kl_threshold": 0.02}
11cfg_ppo["random_timesteps"] = 0
12cfg_ppo["learning_starts"] = 0
13cfg_ppo["grad_norm_clip"] = 1.0
14cfg_ppo["ratio_clip"] = 0.2
15cfg_ppo["value_clip"] = 0.2
16cfg_ppo["clip_predicted_values"] = True
17cfg_ppo["entropy_loss_scale"] = 0.0
18cfg_ppo["value_loss_scale"] = 2.0
19cfg_ppo["kl_threshold"] = 0
20cfg_ppo["rewards_shaper"] = lambda rewards, timestep, timesteps: rewards * 0.01
21cfg_ppo["state_preprocessor"] = RunningStandardScaler
22cfg_ppo["state_preprocessor_kwargs"] = {"size": env.observation_space, "device": device}
23cfg_ppo["value_preprocessor"] = RunningStandardScaler
24cfg_ppo["value_preprocessor_kwargs"] = {"size": 1, "device": device}
25# logging to TensorBoard and writing checkpoints
26cfg_ppo["experiment"]["write_interval"] = 800
27cfg_ppo["experiment"]["checkpoint_interval"] = 8000