keiserlab/exceiver

exceiver

A pretrained single cell gene expression language model.

install

1. clone this repository: git clone git@github.com:keiserlab/exceiver.git
2. install lightweight packaging tool: conda install flit
3. install this repo in a new environment: flit install -s

usage

see notebooks/example.ipynb for loading pretrained models:

from exceiver.models import Exceiver

model = Exceiver.load_from_checkpoint("../pretrained_models/exceiver/pretrained_TS_exceiver.ckpt")

preprocessing

1. downlaod the Tabula Sapiens dataset from figshare, specifically TabulaSapiens.h5ad.zip (careful this link will likely begin download: https://figshare.com/ndownloader/files/34702114)
2. run scripts/preprocess.py:

python preprocess.py --ts_path /path/to/download/TabulaSapiens.h5ad
                     --out_path /path/to/prepocessed/TabulaSapiens

training

pytorch_lightning makes distributed training easy and CLI access to a host of hyperparameters by running scripts/train.py:

python train.py --name MODELNAME 
                --data_path /path/to/prepocessed/TabulaSapiens 
                --logs path/to/model/logs
                --frac 0.15 
                --num_layers 1 
                --nhead 4 
                --query_len 128 
                --batch_size 64 
                --min_epochs 5 
                --max_epochs 10 
                --strategy ddp 
                --gpus 0,1,2,3