mixed_precision: bf16

stage1:
  _target_: lofa.hyper_modulator_mask.TransformerV2
  context_dim: 4096
  d_model: 768
  n_heads: 12
  n_blocks: 8
  patch_size: 96
  num_type: 8
  lora_rank: 32
  num_depth: 30
  factorized: false
  qk_norm: true
  use_cls: true
  decode_cls: false
  tau: 0.02
  temperature: 0.1
  out_act: softplus

stage2:
  _target_: lofa.hyper_modulator_transformer.TransformerV2_stage2
  context_dim: 4096
  d_model: 768
  n_heads: 12
  n_blocks: 8
  patch_size: 96
  num_type: 8
  lora_rank: 32
  num_depth: 30
  factorized: false
  qk_norm: true
  use_cls: true
  decode_cls: false