hparams:
  hidden_dim: 512
  nhead: 8
  low_rank_dim: 64
  query_residual_weight: 0.5
  anchor_scale: 0.8
  num_steps: 4
  num_layers: 6
  embedding_dim: 768
  use_fro_norm: true
feature_extractor: openai/clip-vit-large-patch14