# Tutorial 2: Training a Property Regressor
# Usage: MolCraftDiff train docs/cfg_examples/train_regressor
#
# Trains an EGCL regressor to predict molecular properties.
# The trained model can also be used as a gradient-guidance model
# (see Tutorial 7).

# @package _global_
defaults:
  - data: mol_dataset
  - tasks: regression
  - engine: original
  - logger: default
  - trainer: regression
  - hydra: default
  - _self_

name: "EGCL_reg_my_property"
tags: ["regression"]
seed: 42

logger:
  project_wandb: "MolCraftDiffusion"

tasks:
  task_learn: ["my_property"]   # <-- property column name(s) in your dataset
  num_sublayers: 5
  dropout: 0.1
  prediction_mlp_type: padded
  target_normalization: True
  mlp_batch_norm: batchnorm

data:
  data_type: "pyg"               # CRITICAL: must be "pyg" for regression
  batch_size: 48
  root: data/
  filename: data/my_dataset/metadata.csv   # <-- path to your CSV
  xyz_dir: data/my_dataset/xyz/            # <-- path to your XYZ directory
  dataset_name: my_dataset
  max_atom: 50

trainer:
  output_path: "training_outputs/${name}"
  lr: 8e-4
  validation_interval: 2
  scheduler: reducelronplateau
  scheduler_kwargs:
    mode: "min"
    factor: 0.5
    patience: 10