rpe1/config.yaml

data:
  name: PerturbationDataModule
  kwargs:
    toml_config_path: /large_storage/ctc/ML/state_sets/replogle_proper/rpe1.toml
    embed_key: X_hvg
    output_space: gene
    pert_rep: onehot
    basal_rep: sample
    num_workers: 12
    pin_memory: true
    n_basal_samples: 1
    basal_mapping_strategy: random
    should_yield_control_cells: true
    batch_col: ct_gem_group
    pert_col: gene
    cell_type_key: cell_type
    control_pert: non-targeting
    map_controls: true
    perturbation_features_file: null
    store_raw_basal: false
    int_counts: false
  output_dir: null
  debug: true
model:
  name: PertSets
  checkpoint: null
  device: cuda
  kwargs:
    cell_set_len: 32
    extra_tokens: 1
    blur: 0.05
    hidden_dim: 128
    loss: energy
    confidence_token: false
    n_encoder_layers: 4
    n_decoder_layers: 4
    predict_residual: true
    freeze_pert_backbone: false
    finetune_vci_decoder: false
    residual_decoder: false
    batch_encoder: true
    nb_decoder: false
    decoder_loss_weight: 1.0
    mask_attn: false
    distributional_loss: energy
    regularization: 0.0
    init_from: null
    transformer_backbone_key: GPT2
    transformer_backbone_kwargs:
      max_position_embeddings: 32
      n_positions: 32
      hidden_size: 128
      n_embd: 128
      n_layer: 8
      n_head: 8
      resid_pdrop: 0.0
      embd_pdrop: 0.0
      attn_pdrop: 0.0
      use_cache: false
training:
  wandb_track: false
  weight_decay: 0.0005
  batch_size: 128
  lr: 0.0001
  max_steps: 200000
  train_seed: 42
  val_freq: 2000
  ckpt_every_n_steps: 4000
  gradient_clip_val: 10
  loss_fn: mse
wandb:
  entity: arcinstitute
  project: vci1
  local_wandb_dir: /large_storage/ctc/userspace/aadduri/wandb_dir/vci1
  tags:
  - replogle_proper
name: rpe1
output_dir: /large_storage/ctc/userspace/aadduri/preprint/replogle_hvg_proper_cs32_batch
use_wandb: true
overwrite: false
return_adatas: false
pred_adata_path: null
true_adata_path: null