b240a19ceb
- Fix missing MUON optimizer by replacing with optax.adam - Fix Hydra configuration parameter paths (env.name instead of env_name) - Fix BraxGymnaxWrapper method signatures to accept params argument - Fix training loop division by zero with proper total_time_steps - Fix incorrect algorithm name in wandb (reppo instead of sac) - Fix JAX key batching error in BraxGymnaxWrapper reset method - Add comprehensive HoReKa SLURM integration with wandb logging - Update README with detailed bug documentation and fixes
88 lines
1.9 KiB
YAML
88 lines
1.9 KiB
YAML
defaults:
|
|
- env: brax
|
|
- platform: torch
|
|
- _self_
|
|
|
|
hyperparameters:
|
|
# env and run settings (mostly don't touch)
|
|
total_time_steps: 50_000_000
|
|
normalize_env: true
|
|
max_episode_steps: 1000
|
|
eval_interval: 2
|
|
num_eval: 20
|
|
|
|
# optimization settings (seem very stable)
|
|
lr: 3e-4
|
|
anneal_lr: false
|
|
max_grad_norm: 0.5
|
|
polyak: 1.0 # maybe ablate ?
|
|
|
|
# problem discount settings (need tuning)
|
|
gamma: 0.99
|
|
lmbda: 0.95
|
|
lmbda_min: 0.50 # irrelevant if no exploration noise is added
|
|
|
|
# batch settings (need tuning for MJX humanoid)
|
|
num_steps: 128
|
|
num_mini_batches: 128
|
|
num_envs: 1024
|
|
num_epochs: 4
|
|
|
|
# exploration settings (currently not touched)
|
|
exploration_noise_max: 1.0
|
|
exploration_noise_min: 1.0
|
|
exploration_base_envs: 0
|
|
|
|
# critic architecture settings (need to be increased for MJX humanoid)
|
|
critic_hidden_dim: 512
|
|
actor_hidden_dim: 512
|
|
vmin: ${env.vmin}
|
|
vmax: ${env.vmax}
|
|
num_bins: 151
|
|
hl_gauss: true
|
|
use_critic_norm: true
|
|
num_critic_encoder_layers: 2
|
|
num_critic_head_layers: 2
|
|
num_critic_pred_layers: 2
|
|
use_simplical_embedding: False
|
|
|
|
# actor architecture settings (seem stable)
|
|
use_actor_norm: true
|
|
num_actor_layers: 3
|
|
actor_min_std: 0.0
|
|
|
|
# actor & critic loss settings (seem remarkably stable)
|
|
## kl settings
|
|
kl_start: 0.01
|
|
kl_bound: 0.1 # switched to tighter bounds for MJX
|
|
reduce_kl: true
|
|
reverse_kl: false # previous default "false"
|
|
update_kl_lagrangian: true
|
|
actor_kl_clip_mode: "clipped" # "full", "clipped", "kl_relu_clipped", "kl_bound_clipped", "value"
|
|
## entropy settings
|
|
ent_start: 0.01
|
|
ent_target_mult: 0.5
|
|
update_entropy_lagrangian: true
|
|
## auxiliary loss settings
|
|
aux_loss_mult: 1.0
|
|
|
|
|
|
measure_burnin: 3
|
|
|
|
|
|
name: "reppo"
|
|
seed: 0
|
|
num_seeds: 1
|
|
tune: false
|
|
checkpoint_dir: null
|
|
num_trials: 10
|
|
tags: ["experimental"]
|
|
wandb:
|
|
mode: "online" # set to online to activate wandb
|
|
entity: "viper_svg"
|
|
project: "online_sac"
|
|
|
|
hydra:
|
|
job:
|
|
chdir: True
|