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-Crazyflie-PPO")
5agent.load(path)1# https://skrl.readthedocs.io/en/latest/modules/skrl.agents.ppo.html#configuration-and-hyperparameters
2cfg_agent = PPO_DEFAULT_CONFIG.copy()
3cfg_agent["rollouts"] = 16 # memory_size
4cfg_agent["learning_epochs"] = 8
5cfg_agent["mini_batches"] = 1 # 16 * 512 / 8192
6cfg_agent["discount_factor"] = 0.99
7cfg_agent["lambda"] = 0.95
8cfg_agent["learning_rate"] = 3e-4
9cfg_agent["learning_rate_scheduler"] = KLAdaptiveRL
10cfg_agent["learning_rate_scheduler_kwargs"] = {"kl_threshold": 0.008}
11cfg_agent["random_timesteps"] = 0
12cfg_agent["learning_starts"] = 0
13cfg_agent["grad_norm_clip"] = 1.0
14cfg_agent["ratio_clip"] = 0.2
15cfg_agent["value_clip"] = 0.2
16cfg_agent["clip_predicted_values"] = True
17cfg_agent["entropy_loss_scale"] = 0.0
18cfg_agent["value_loss_scale"] = 2.0
19cfg_agent["kl_threshold"] = 0
20cfg_agent["rewards_shaper"] = lambda rewards, timestep, timesteps: rewards * 0.1
21cfg_agent["state_preprocessor"] = RunningStandardScaler
22cfg_agent["state_preprocessor_kwargs"] = {"size": env.observation_space, "device": device}
23cfg_agent["value_preprocessor"] = RunningStandardScaler
24cfg_agent["value_preprocessor_kwargs"] = {"size": 1, "device": device}
25# logging to TensorBoard and writing checkpoints
26cfg_agent["experiment"]["write_interval"] = 16
27cfg_agent["experiment"]["checkpoint_interval"] = 80