train_data:
  dir: sample/motion/processed
  mask: [src, tgt]

representation:
  skel: [lo, go] # S
  pose: [q, p, r, pv, qv, pprev, c] # D_t
  out: [r, q, c] # \hat{D_t}

model:
  z_dim: 32
  Encoder:
      type: GATEnc
      hid_lyrs: [16, 16, 16]
      heads_num: 16
  Decoder:
      type: GATDec
      hid_lyrs: [16, 16, 16]
      heads_num: 16
      tgt_all_lyr: True
  # load: 
  #   - dir: blahblah
  #     epoch: 
  #     prefix: encoder
  #     freeze: True

train:
  consq_n: 8
  batch_size: 256 # 512 if GPU memory allows
  grad_max_norm: 0.5
  learning_rate: 0.01
  
  lr_schedule: 
    type: exponential
    gamma: 0.99
    min: 0.01

  epoch_num: 300
  save_per: 30

  # load trainer to continue training
  # load:
  #   dir: prev_trainer
  #   epoch: None

  # misc
  copy_orig_contact: False

  loss:
    q: 5
    p: 0.01
    r: 10
    pv: 1
    c: 1
    cv: 6
    z: 1
    pen: 0.1
    slide: 6
    jerk: 0.2

  metric: [qR, ra_xz, pa, slide, jerk, pen]