chore: update yaml (#93)

omnisafe/configs/on-policy/PPO.yaml

@@ -286,136 +286,3 @@ SafetyRacecarGoal1-v0:
     max_len: 100
     # The number of threads used to sample data
     num_threads: 20
-
-
-SafetyHumanoidVelocity-v4:
-  # --------------------------------------Basic Configurations----------------------------------- #
-  ## -----------------------------Basic configurations for base class PG------------------------ ##
-  # The random seed
-  seed: 0
-  # The torch device
-  device: cpu
-  # The torch device id
-  device_id: 0
-  # The environment wrapper type
-  wrapper_type: CMDPWrapper
-  # Number of epochs
-  epochs: 500
-  # Number of steps per epoch
-  steps_per_epoch: 32768
-  # Number of update iteration for Actor network
-  actor_iters: 40
-  # Number of update iteration for Critic network
-  critic_iters: 40
-  # Check if all models own the same parameter values every `check_freq` epoch
-  check_freq: 25
-  # Save model to disk every `check_freq` epochs
-  save_freq: 50
-  # Entropy coefficient for PPO loss
-  entropy_coef: 0.0
-  # The max length of per epoch
-  max_ep_len: 1000
-  # The size of mini batch
-  num_mini_batches: 64
-  # The learning rate of Actor network
-  actor_lr: 0.0003
-  # The learning rate of Critic network
-  critic_lr: 0.0003
-  # The Address for saving training process data
-  data_dir: "./runs"
-  ## ---------------------------Basic configurations for derived class PPO---------------------- ##
-  # The thereshold for KL early stopping
-  target_kl: 0.02
-  # The size of batch for policy update
-  batch_size: 10000
-  # The clip range for PPO loss
-  clip: 0.2
-
-  # ---------------------------------------Optional Configuration-------------------------------- #
-  ## -----------------------------------Configuration For Cost Critic--------------------------- ##
-  # Whether to use cost critic
-  use_cost: False
-  # Cost discounted factor
-  cost_gamma: 1.0
-  # Whether to use linear decay of learning rate
-  linear_lr_decay: True
-  # Whether to use exploration noise anneal
-  exploration_noise_anneal: False
-  # The coefficient of reward penalty
-  penalty_param: 0.0
-  # Whether to use KL early stopping
-  kl_early_stopping: True
-  # Whether to use max gradient norm
-  use_max_grad_norm: True
-  # The thereshold of max gradient norm
-  max_grad_norm: 40
-  # Whether to use critic network norm
-  use_critic_norm: True
-  # The norm coefficient of critic network
-  critic_norm_coeff: 0.001
-  ## ---------------------------------------Configuration For Model----------------------------- ##
-  model_cfgs:
-    # Whether to share the weight of Actor network with Critic network
-    shared_weights: False
-    # The mode to initiate the weight of network, choosing from "kaiming_uniform", "xavier_normal", "glorot" and "orthogonal".
-    weight_initialization_mode: "kaiming_uniform"
-    # Type of Actor, choosing from "gaussian_annealing", "gaussian_std_net_actor", "gaussian_learning_actor", "categorical_actor"
-    actor_type: gaussian
-    # Configuration of Actor and Critic network
-    ac_kwargs:
-      # Configuration of Actor network
-      pi:
-        # Size of hidden layers
-        hidden_sizes: [64, 64]
-        # Activation function
-        activation: tanh
-        # Output activation function
-        output_activation: identity
-        # Whether to scale action.
-        scale_action: False
-        # Whether to clip action.
-        clip_action: False
-        # Whther to learn the standard deviation of Gaussian noise
-        std_learning: True
-        # The initial value of standard deviation of Gaussian noise
-        std_init: 1.0
-      # Configuration of Critic network
-      val:
-        # Number of critic networks
-        num_critics: 1
-        # Size of hidden layers
-        hidden_sizes: [64, 64]
-        # Activation function
-        activation: tanh
-  ## --------------------------------------Configuration For Buffer----------------------------- ##
-  buffer_cfgs:
-    # Reward discounted factor
-    gamma: 0.99
-    # Parameters used to estimate future rewards in GAE
-    lam: 0.95
-    # Parameters used to estimate future costs in GAE
-    lam_c: 0.95
-    # Method to estimate the advantage reward/cost, choosing from "gae", "plain", "vtrace"
-    adv_estimation_method: gae
-    # Whether to use reward standardized advantage estimation
-    standardized_rew_adv: True
-    # Whether to use cost standardized advantage estimation
-    standardized_cost_adv: True
-  ## --------------------------------------Configuration For Environment------------------------ ##
-  env_cfgs:
-    # The seed of environment
-    env_seed: 0
-    # The number of parallel environments
-    num_envs: 8
-    # Whether to use async environment
-    async_env: True
-    # Whether to use standardized reward
-    normalized_rew: True
-    # Whether to use standardized cost
-    normalized_cost: False
-    # Whether to use standardized obs
-    normalized_obs: False
-    # The maximum length of record queue
-    max_len: 100
-    # The number of threads used to sample data
-    num_threads: 20