cell-classification

Instructions

Download the dataset from HuggingFace

Download the files here and put it in the working directory in the cell_type_train_data.dataset folder.

The workflow will upload these files and cache them in S3.

Setup

pip install outerbounds
outerbounds configure <>

If you want to make a private version of the Docker image for the fine-tuning step of the workflow, see the Dockerfile in the root of this repository.

docker build -t geneformer .

Iterate

You can find hyperparameters to tune in config.py.

Run the workflow

Run the workflow locally

Download Geneformer and install dependencies in the virtual environment where you are running Metaflow

git clone https://huggingface.co/ctheodoris/Geneformer
cd Geneformer
pip install .

Then, run the flow, passing a parameter to the workflow that is used to load the model with HuggingFace:

python flow.py run --path <PATH/TO/Geneformer>                    # default model
python flow.py run --path <PATH/TO/Geneformer/geneformer-12L-30M> # bigger model

Run the workflow remotely

python flow.py run # by default, try to use geneformer-12L-30M. use --path as before if you change stuff in the docker image
python flow.py run --path /Geneformer # to use the default model.