defaults:
  - _self_
hydra:
  run:
    dir: ${logdir}
_target_: agent.finetune.train_ppo_diffusion_agent.TrainPPODiffusionAgent

name: ${env_name}_ft_diffusion_mlp_ta${horizon_steps}_td${denoising_steps}_tdf${ft_denoising_steps}
logdir: ${oc.env:DPPO_LOG_DIR}/furniture-finetune/${name}/${now:%Y-%m-%d}_${now:%H-%M-%S}_${seed}
base_policy_path: ${oc.env:DPPO_LOG_DIR}/furniture-pretrain/lamp/lamp_low_dim_pre_diffusion_mlp_ta8_td100/2024-07-23_01-28-20/checkpoint/state_8000.pt
normalization_path: ${oc.env:DPPO_DATA_DIR}/furniture/${env.specific.furniture}_${env.specific.randomness}/normalization.pth

seed: 42
device: cuda:0
env_name: ${env.specific.furniture}_${env.specific.randomness}_dim
obs_dim: 44
action_dim: 10
transition_dim: ${action_dim}
denoising_steps: 100
ft_denoising_steps: 5
cond_steps: 1
horizon_steps: 8
act_steps: 8
use_ddim: True

env:
  n_envs: 1000
  name: ${env_name}
  env_type: furniture
  max_episode_steps: 1000
  best_reward_threshold_for_success: 2
  specific:
    headless: true
    furniture: lamp
    randomness: low
    normalization_path: ${normalization_path}
    obs_steps: ${cond_steps}
    act_steps: ${act_steps}
    sparse_reward: True

wandb:
  entity: ${oc.env:DPPO_WANDB_ENTITY}
  project: furniture-${env.specific.furniture}-${env.specific.randomness}-finetune
  run: ${now:%H-%M-%S}_${name}

train:
  n_train_itr: 1000
  n_critic_warmup_itr: 1
  n_steps: ${eval:'round(${env.max_episode_steps} / ${act_steps})'}
  gamma: 0.999
  actor_lr: 1e-5
  actor_weight_decay: 0
  actor_lr_scheduler:
    first_cycle_steps: 10000
    warmup_steps: 10
    min_lr: 1e-6
  critic_lr: 1e-3
  critic_weight_decay: 0
  critic_lr_scheduler:
    first_cycle_steps: 10000
    warmup_steps: 10
    min_lr: 1e-3
  save_model_freq: 50
  val_freq: 10
  render:
    freq: 1
    num: 0
  # PPO specific
  reward_scale_running: True
  reward_scale_const: 1.0
  gae_lambda: 0.95
  batch_size: 8800
  update_epochs: 5
  vf_coef: 0.5
  target_kl: 1

model:
  _target_: model.diffusion.diffusion_ppo.PPODiffusion
  # HP to tune
  gamma_denoising: 0.9
  clip_ploss_coef: 0.001
  clip_ploss_coef_base: 0.001
  clip_ploss_coef_rate: 3
  randn_clip_value: 3
  min_sampling_denoising_std: 0.04
  #
  use_ddim: ${use_ddim}
  ddim_steps: ${ft_denoising_steps}
  learn_eta: False
  eta:
    base_eta: 1
    input_dim: ${obs_dim}
    mlp_dims: [256, 256]
    action_dim: ${action_dim}
    min_eta: 0.1
    max_eta: 1.0
    _target_: model.diffusion.eta.EtaFixed
  network_path: ${base_policy_path}
  actor:
    _target_: model.diffusion.mlp_diffusion.DiffusionMLP
    time_dim: 32
    mlp_dims: [1024, 1024, 1024, 1024, 1024, 1024, 1024]
    cond_mlp_dims: [512, 64]
    use_layernorm: True # needed for larger MLP
    residual_style: True
    cond_dim: ${eval:'${obs_dim} * ${cond_steps}'}
    horizon_steps: ${horizon_steps}
    transition_dim: ${transition_dim}
  critic:
    _target_: model.common.critic.CriticObs
    cond_dim: ${eval:'${obs_dim} * ${cond_steps}'}
    mlp_dims: [512, 512, 512]
    activation_type: Mish
    residual_style: True
  ft_denoising_steps: ${ft_denoising_steps}
  horizon_steps: ${horizon_steps}
  obs_dim: ${obs_dim}
  action_dim: ${action_dim}
  denoising_steps: ${denoising_steps}
  device: ${device}