adaptive-machine-learning / CapyMOA / 31173798693

80%

push

github

fix(drift): make ABCD usable without PyTorch, and default to PCA

ABCD reconstructs instances with one of three encoder-decoder models. Only
the autoencoder needs PyTorch; PCAModel and KernelPCAModel use scikit-learn,
which is a core dependency. All three lived in feature_extraction.py, whose
module-level `import torch.nn` forced the whole detector behind the torch
gate, so `pip install capymoa` lost the only multivariate drift detector even
for configurations that never touch PyTorch.

Move AutoEncoder to its own module. It subclasses nn.Module, so the import
cannot be deferred into a method; a separate file is what lets abcd.py load it
lazily, and only when model_id="ae" is actually requested. Asking for that
model without the extra now raises OptionalDependencyError naming the model
rather than the detector, since the other two remain available.

Switch the default from "ae" to "pca" so a plain ABCD() works in a default
install. This changes the model an existing caller gets from the default, and
belongs in the release notes; on the univariate stream in the drift detection
notebook both models behave identically, so its stored output is unchanged.

Also give the unknown-model_id branch a message instead of a bare ValueError.

61 of 65 new or added lines in 4 files covered. (93.85%)

8783 of 10928 relevant lines covered (80.37%)

0.8 hits per line