Installation#

ReconEval supports two install paths.

Metrics only (lightweight pip install)#

If you just want to score your own (true, reconstructed) AnnData pair with the metrics API and run tutorials/metrics.ipynb, a vanilla pip install is enough:

git clone https://github.com/theislab/ReconEval.git
cd ReconEval
python -m venv .venv && source .venv/bin/activate
pip install -e .

Verify:

python -c "from sc_reconstruction.metrics import compute_all_metrics, funky_heatmap; print('OK')"

Full benchmark (conda env)#

To run the other three tutorials (end_to_end.ipynb, fm.ipynb, latent_shift.ipynb) or any of the training and eval drivers under experiments/, use the conda environment file. It pins compatible versions of torch, scvi-tools, lightning, scvi, dask, and the STATE foundation-model package.

conda env create -f envs/cstm_scvi_env.yaml
conda activate reconeval

Optional environment variables#

The training and eval drivers read a small set of env vars:

Variable	What it controls
`RECONEVAL_OUT`	Output root for weights, checkpoints, Hydra outputs (default `~/reconeval_outputs/`).
`CELLFLOW_SRC`	Path to a cellflow source clone for `03_latent_shift`. Skipped if unset.
`STATE_SRC`	Path to a STATE source clone for SE embedding. Skipped if unset.