Inference with custom model

[avantikalal]

Changes made to allow easier inference and evaluation of custom Decima models, including models trained on non-hg38 genomes.

src/decima/hub/init.py
Changed load_decima_model to:

• take as input ckpt files as well as safetensors. This is important because custom models are saved as .ckpt files, and the user should not need to convert them into another format before being able to use them.
• load multiple custom models and ensemble them.

src/decima/cli/predict_genes.py
Allow cli_predict_genes to take as input a list of model paths and ensemble them.

src/decima/core/result.py
Added a genome argument to the function prepare_one_hot so that one-hot encoded sequences can be produced for non-hg38 genomes.

src/decima/data/dataset.py
Added a genome argument to GeneDataset so that gene expression prediction can be performed for non-hg38 genomes.

src/decima/tools/inference.py
Applied the genome argument during inference.
The function predict_gene_expression ran both prediction and evaluation. I split the evaluation part into a separate function evaluate_gene_expression_predictions to be more modular.

docs/tutorials/3-finetune.ipynb
Added a section to the tutorial showing how to apply and evaluate a custom model.

[MuhammedHasan]

Looks good but I will add support for passing list of models to other endpoints as well such as attributions and vep.

[MuhammedHasan]

• New endpoints to download model files and metadata.
• Support for vep and attribution api with model paths are added.
• name parameter in the LigthiningModel is compulsory.