defaults:
  - _self_
hydra:
  run:
    dir: ${logdir}
_target_: agent.eval.eval_diffusion_agent.EvalDiffusionAgent

name: ${env_name}_eval_diffusion_mlp_ta${horizon_steps}_td${denoising_steps}
logdir: ${oc.env:DPPO_LOG_DIR}/furniture-eval/${name}/${now:%Y-%m-%d}_${now:%H-%M-%S}_${seed}
base_policy_path:
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: 58
action_dim: 10
denoising_steps: 100
cond_steps: 1
horizon_steps: 8
act_steps: 8
use_ddim: True
ddim_steps: 5

n_steps: ${eval:'round(${env.max_episode_steps} / ${act_steps})'}
render_num: 0

env:
  n_envs: 1000
  name: ${env_name}
  env_type: furniture
  max_episode_steps: 700
  best_reward_threshold_for_success: 1
  specific:
    headless: true
    furniture: one_leg
    randomness: low
    normalization_path: ${normalization_path}
    obs_steps: ${cond_steps}
    act_steps: ${act_steps}
    sparse_reward: True

model:
  _target_: model.diffusion.diffusion.DiffusionModel
  predict_epsilon: True
  denoised_clip_value: 1.0
  randn_clip_value: 3
  #
  use_ddim: ${use_ddim}
  ddim_steps: ${ddim_steps}
  network_path: ${base_policy_path}
  network:
    _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}
    action_dim: ${action_dim}
  horizon_steps: ${horizon_steps}
  obs_dim: ${obs_dim}
  action_dim: ${action_dim}
  denoising_steps: ${denoising_steps}
  device: ${device}