defaults:
  - _self_
hydra:
  run:
    dir: ${logdir}
_target_: agent.pretrain.train_diffusion_agent.TrainDiffusionAgent

name: ${env}_pre_diffusion_mlp_ta${horizon_steps}_td${denoising_steps}
logdir: ${oc.env:DPPO_LOG_DIR}/furniture-pretrain/${name}/${now:%Y-%m-%d}_${now:%H-%M-%S}_${seed}
train_dataset_path: ${oc.env:DPPO_DATA_DIR}/furniture/${task}_${randomness}/train.pkl

seed: 42
device: cuda:0
task: one_leg
randomness: med
env: ${task}_${randomness}_dim
obs_dim: 58
action_dim: 10
transition_dim: ${action_dim}
denoising_steps: 100
horizon_steps: 8
cond_steps: 1

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

train:
  n_epochs: 8000
  batch_size: 256
  learning_rate: 1e-4
  weight_decay: 1e-6
  lr_scheduler:
    first_cycle_steps: 10000
    warmup_steps: 100
    min_lr: 1e-5
  epoch_start_ema: 20
  update_ema_freq: 10
  save_model_freq: 1000

model:
  _target_: model.diffusion.diffusion.DiffusionModel
  predict_epsilon: True
  denoised_clip_value: 1.0
  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}
    transition_dim: ${transition_dim}
  horizon_steps: ${horizon_steps}
  obs_dim: ${obs_dim}
  action_dim: ${action_dim}
  transition_dim: ${transition_dim}
  denoising_steps: ${denoising_steps}
  cond_steps: ${cond_steps}
  device: ${device}

ema:
  decay: 0.995

train_dataset:
  _target_: agent.dataset.sequence.StitchedActionSequenceDataset
  dataset_path: ${train_dataset_path}
  horizon_steps: ${horizon_steps}
  cond_steps: ${cond_steps}
  device: ${device}