Q&A

[ruiqwy]

This post summarizes the issues raised and provides answers. Further questions can also be commented here.

[Barry57]

1. Regarding Bimax:

• We have fixed the convergence issue.
• A synthetic dataset is now included.
• For visualization, Bimax reuses the bcheatmap function already available for the SSVD method, which draws boxes around the detected biclusters in the heatmap.

[Barry57]

2. Regarding fem_bifunc:

• We have removed all np.matrix calls and replaced them with standard ndarrays to ensure full compatibility with scikit-learn.

• The update has been pushed to BiFuncLib; we are now running comprehensive tests to confirm that nothing else breaks.

[Barry57]

3. Regarding the ImportError: cannot import name '__all__' from 'numpy.linalg':

• Root cause: a binary-interface mismatch between an old NumPy (≤ 1.20.x) and a newer SciPy (≥ 1.9) that was compiled against NumPy ≥ 1.22.

• Fix: create a clean environment and install the package versions listed in BiFuncLib’s requirements; this guarantees ABI compatibility and eliminates the import error.

[Barry57]

4. Regarding the ValueError: could not broadcast input array from shape (40,5) into shape (5,) in lbm_bifunc:

• Root cause: with the chosen init='kmeans' and basis_name='spline', the spline basis collapses one class to a single observation, yielding an incompatible array shape.

• Immediate fix: switch to basis_name='fourier' (or any other supported basis); Fourier series avoids the collapse and runs without broadcasting errors.