defaults:
  - _self_
hydra:
  run:
    dir: ${logdir}
_target_: agent.eval.eval_diffusion_img_agent.EvalImgDiffusionAgent

name: ${env_name}_eval_diffusion_unet_img_ta${horizon_steps}_td${denoising_steps}
logdir: ${oc.env:DPPO_LOG_DIR}/robomimic-eval/${name}/${now:%Y-%m-%d}_${now:%H-%M-%S}_${seed}
base_policy_path:
robomimic_env_cfg_path: cfg/robomimic/env_meta/${env_name}-img.json
normalization_path: ${oc.env:DPPO_DATA_DIR}/robomimic/${env_name}-img/normalization.npz

seed: 42
device: cuda:0
env_name: transport
obs_dim: 18
action_dim: 14
denoising_steps: 100
cond_steps: 1
img_cond_steps: 1
horizon_steps: 16
act_steps: 8
use_ddim: True
ddim_steps: 5

n_steps: 400  # each episode takes max_episode_steps / act_steps steps
render_num: 0

env:
  n_envs: 30  # reduce gpu usage
  name: ${env_name}
  best_reward_threshold_for_success: 1
  max_episode_steps: 800
  save_video: False
  use_image_obs: True
  wrappers:
    robomimic_image:
      normalization_path: ${normalization_path}
      low_dim_keys: ['robot0_eef_pos',
                     'robot0_eef_quat',
                     'robot0_gripper_qpos',
                     "robot1_eef_pos",
                     "robot1_eef_quat",
                     "robot1_gripper_qpos"]
      image_keys: ['shouldercamera0_image', 
                   'shouldercamera1_image']
      shape_meta: ${shape_meta}
    multi_step:
      n_obs_steps: ${cond_steps}
      n_action_steps: ${act_steps}
      max_episode_steps: ${env.max_episode_steps}
      reset_within_step: True

shape_meta:
  obs:
    rgb:
      shape: [6, 96, 96]
    state:
      shape: [18]
  action: 
    shape: [14]

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.unet.VisionUnet1D
    backbone:
      _target_: model.common.vit.VitEncoder
      obs_shape: ${shape_meta.obs.rgb.shape}
      num_channel: ${eval:'3 * ${img_cond_steps}'} # each image patch is history concatenated
      img_h: ${shape_meta.obs.rgb.shape[1]}
      img_w: ${shape_meta.obs.rgb.shape[2]}
      cfg:
        patch_size: 8
        depth: 1
        embed_dim: 128
        num_heads: 4
        embed_style: embed2
        embed_norm: 0
    img_cond_steps: ${img_cond_steps}
    augment: False
    num_img: 2
    spatial_emb: 128
    diffusion_step_embed_dim: 32
    dim: 64
    dim_mults:
    - 1
    - 2
    kernel_size: 5
    n_groups: 8
    smaller_encoder: false
    cond_predict_scale: true
    action_dim: ${action_dim}
    cond_dim: ${eval:'${obs_dim} * ${cond_steps}'}
  horizon_steps: ${horizon_steps}
  obs_dim: ${obs_dim}
  action_dim: ${action_dim}
  denoising_steps: ${denoising_steps}
  device: ${device}