Open-Source Contributions

I contribute performance, correctness and documentation fixes to the scientific Python ecosystem. Statuses below are as of October 2026.

scikit-learn

• FastICA performance (#33269, fixes #32914). In the deflation algorithm, FastICA calls the _logcosh non-linearity on 1-D arrays. The existing code then computed the derivative one element at a time in a Python loop. The PR adds a vectorised path for 1-D input and keeps the memory-saving loop for 2-D input. For FastICA(algorithm="deflation", fun="logcosh") on a 2000 × 50 dataset, fitting time fell from 424.0 s to 17.7 s (≈ 24×). A regression test was added.
• Spurious feature-name warning (#33268, fixes #31947). PowerTransformer and QuantileTransformer wrongly warned “X has feature names, but … was fitted without feature names” in inverse_transform. The fix validates input with check_array and adds regression tests. It builds on earlier work by another contributor in a closed PR.

Core change in #33269 (sklearn/decomposition/_fastica.py):

 def _logcosh(x, fun_args=None):
     alpha = fun_args.get("alpha", 1.0)

     x *= alpha
     gx = np.tanh(x, x)  # apply the tanh inplace
+
+    if x.ndim == 1:
+        return gx, alpha * (1 - gx**2)
+
+    # When the input is 2D, compute in a loop to avoid extra allocation
+    # of array of shape x.shape
     g_x = np.empty(x.shape[0], dtype=x.dtype)
     for i, gx_i in enumerate(gx):
         g_x[i] = (alpha * (1 - gx_i**2)).mean()
     return gx, g_x

scikit-bio

• Import latency (#2427). Moved the optional requests and h5py imports into the functions that use them, so import skbio no longer loads them. Mean import time fell from 112.7 ms to 70.3 ms (−37%) over 5 fresh-process runs, and all 368 I/O tests pass. This is a step towards the project’s broader import-time issue (#2170).

Keras Hub

• LiteRT export (#2555, fixes #2493). The LiteRT (TFLite) export test helper cast boolean attention masks to int32, which broke export of the Moonshine speech model. Keeping the masks boolean lets the previously skipped Moonshine export test run again.

Own packages

• flwr-trees: federated learning for tree ensembles with a scikit-learn API (PyPI).