ml4g_project_1/train_exp_config.yaml

data:
  bigwig:
    raw:
      root_dir: '/home/juanbermeo/Documents/ml4g_project_1/data/raw_bigwig'
    h5:
      root_dir: '/home/juanbermeo/Documents/ml4g_project_1/data/preprocessed/bigwig_50k_window'

  bed:
    h5:
      root_dir: '/home/juanbermeo/Documents/ml4g_project_1/data/preprocessed/resampled_bed'
      features: [
        'DNase',
        'H3K27ac',
        'H3K27me3',
        'H3K36me3',
        'H3K4me1',
        'H3K4me3',
      ]

preprocessing:
  bed:
    features: [
      'DNase',
      'H3K27ac',
      'H3K27me3',
      'H3K36me3',
      'H3K4me1',
      'H3K4me3',
    ]
    num_resampled_bins: 1000

  raw_bigwig:
    features: [
      'DNase',
      'H3K27ac',
      'H3K27me3',
      'H3K36me3',
      'H3K4me1',
      'H3K4me3',
      # 'H3K9me3',
      'DNAhg38'
    ]
    n_features: 10                  # Using DNAhg38 adds 4 features. The markers adds 1 each

model:
  bigwig:
    C: 16                           # Number of output channels of conv layers
    N: 4                            # Number of Conv blocks
    M: 4                            # Number of dilated Conv blocks to use

  bed:
    C: 16
    N: 3
    M: 3

  mlp:
    H: 16


train:
  dataset_type: 'bigwig_bed_h5'    # ['bigwig_raw', 'bigwig_h5', 'bed_h5', 'bigwig_bed_h5'] are the options
  mini_batch_size: 8
  num_workers: 10
  device: 'cuda:0'
  run_dir_path: './train_runs'
  cell_lines: ['X1']              # ['X1', 'X2'] for reporting on test
  lr: 0.0001
  gamma: 0.8
  random_shift: 10
  momentum: 0.9
  weight_decay: 0.1
  n_warmup_epochs: 1
  max_grad_norm: 10000


validation:
  cell_lines: ['X2']