I'd love to have a minimal version for deployment that uses more stable exports of the models. I think ONNX looks really promising. As far as a preliminary test showed, it should be relatively easy to convert tensorflow/keras, pytorch, and scikit-learn models to the onnx format.
Ideally this would remove the large tensorflow dependency and the pinned scikit-learn dependency in the production code. What's blocking this is the very involved nature of the cipherTypeDetection module with its dependency on those libraries.
The pinned dependencies also pose a challenge as the base python version (3.11) cannot be updated.
I'd love to have a minimal version for deployment that uses more stable exports of the models. I think ONNX looks really promising. As far as a preliminary test showed, it should be relatively easy to convert tensorflow/keras, pytorch, and scikit-learn models to the onnx format.
Ideally this would remove the large tensorflow dependency and the pinned scikit-learn dependency in the production code. What's blocking this is the very involved nature of the
cipherTypeDetectionmodule with its dependency on those libraries.The pinned dependencies also pose a challenge as the base python version (3.11) cannot be updated.