pretrain_protbert.yaml

output_dir: ~/scratch/protst_output/

dataset:
  class: UniProtSeqText
  path: ~/scratch/protein-datasets/
  seq_lazy: True
  atom_feature: null
  bond_feature: null
  transform:
    class: Compose
    transforms:
      - class: TruncateProtein
        max_length: 450
        random: True
      - class: ProteinView
        view: "residue"

fix_text_model: True

lr_ratio: 0.1  # add this line when fine-tuning the protein encoder

task:
  class: ProtSTMMP
  protein_model:
    class: PretrainProtBert
    path: ~/scratch/protbert-model-weights/
    mask_modeling: True
    output_dim: 512
    readout: pooler
  text_model:
    class: PubMedBERT
    model: {{ text_model }} # PubMedBERT-abs / PubMedBERT-full
    output_dim: 512 # NOTE: same as the output_dim of protein_model
    readout: mean
  fusion_model:
    class: CrossAttention
    hidden_dim: 512
    num_layers: 1
    num_heads: 8
    batch_norm: True
  protein2text: True
  text2protein: True
  mlm_weight: 1.0
  mmp_weight: 1.0
  mask_rate: 0.15
  num_mlp_layer: 2
  global_contrast: True

optimizer:
  class: Adam
  lr: 1.0e-5

engine:
  gpus: [0, 1, 2, 3]
  batch_size: 3

train: 
  num_epoch: 20