Skip to content

Latest commit

 

History

History
76 lines (47 loc) · 4.7 KB

input.md

File metadata and controls

76 lines (47 loc) · 4.7 KB

Input Files and Parameters

This document describes the input files and parameters required by scGenAI for different modes of operation: Train, Predict, and Finetune.

Input Mode

  • mode: Specifies the mode of operation. Choices: Train, Predict, Finetune. Default: Train.

Input Files and Dir

Input Files

These files are required in most configurations. Depending on the mode (Train, Predict, Finetune), some may be required:

  • train_file: Required when mode is Train or Finetune. The primary training data file in .h5ad format.
  • val_file: Optional. The validation data file used during training.
  • train_ADTfile: Required when multiomics is set to Yes and mode is Train or Finetune. The training file for ADT (antibody-derived tags) data.
  • val_ADTfile: Optional. The validation file for ADT data.
  • cytofile: Required when context_method is set to genomic and mode is Train. The cytoband data file used in genomic context. It should be a tab separated files contains the two columns, gene_symbol and cytobandID.
  • gmtfile: Required when context_method is set to biofounction and mode is Train. The gmt format file for biofunction context.
  • glstfile: Required when context_method is set to genelist and mode is Train or Finetune. A gene list file used in genelist context. It should be a single column no header with gene names.
  • predict_file: Required when mode is Predict. The input file for prediction tasks.
  • predict_ADTfile: Required when multiomics is set to Yes and mode is Predict. The ADT file for prediction tasks (if using multi-omics data).
  • outputfile: Required when mode is Predict. The output CSV file where the prediction results will be saved.
  • cache_dir: Required. Directory to cache models template during training and prediction, when the first time of using model template, the cache files will be saved in this folder, and will be loaded directly for reuse.

Input Dir

  • log_dir: Required when savelog is set to Yes. Directory to store logs. Default: examples/logs.
  • model_dir: Required when mode is Train or Finetune. Directory where models are saved during training, which is also the input folder in the Finetune mode.
  • finetune_dir: Required when mode is Finetune. Directory used for saving fine-tuned models. Default: examples/finetune.

Input Parameters

Preprocessing Settings

  • min_cells: Optional. Minimum number of cells required for filtering. Default: 50.
  • target_feature: Required. The feature to predict (e.g., celltype). Default: celltype.
  • multiomics: Optional. Whether multi-omics data (e.g., RNA and ADT) is used. Choices: Yes, No. Default: No.
  • savelog: Optional. Whether to save logs to the log_dir. Choices: Yes, No. Default: No.
  • num_bins: Number of bins used for gene expression normalization. Default: 10.

Model Settings

  • model_backbone_name: Required when mode is Train. The backbone model template to use. Choices: gpt, bigbird, llama, scGenT. Default: llama.
  • model_backbone_size: Required when mode is Train. The size of the model. Choices: small, normal, large. Default: normal. Suggested size: small for llama and bigbird; normal for gpt and scGenT.
  • context_method: Method for generating context. Choices: random, genomic, biofounction, genelist. Default: random.

Training Settings

  • batch_size: Required. The batch size for training. Default: 8. We suggest the user begin with 1 to avoid the error of out of GPU memory.
  • learning_rate: Optional. The learning rate for the model trainning. Default: 1e-5.
  • num_epochs: Optional. Number of training epochs. Default: 30. We suggest use 30 for training, 20~30 for finetune mode.
  • weight_decay: Optional. Weight decay rate for the optimizer. Default: 0.01.
  • depth: Optional. Depth for the genomic/biofounction/genelist context method. Default: 2.
  • seed: Optional. Random seed for reproducibility. Default: 1314521.
  • max_length: Required when mode is Train. Maximum sequence length for tokenization. Default: 1024. Suggested size: 5120 for llama; 1024 for gpt; 4096 for bigbird; 2048 for scGenT.

Next Step: Build the YAML configuration file

All the required parameters must be defined in the YAML configuration file. Please read the doc of Configuration for details