Pepsickle for Debian
===================

The upstream gradient boosting models were saved with scikit-learn 1.2.0.
Newer scikit-learn versions cannot load that format. During the Debian build,
debian/convert-models.py extracts the existing tree nodes, leaf values and
class priors into JSON. The installed package evaluates those trees with
NumPy, preserving the trained models without depending on scikit-learn's
private serialization format. The original joblib files remain in the source
package. The neural network weights are unchanged and use Debian's PyTorch.

Upstream describes the training methods and provides training scripts in
https://github.com/pdxgx/pepsickle-paper (see also the upstream README).

Example FASTA input is installed in
/usr/share/doc/python3-pepsickle/examples/P43357.fasta.