Question about eigendecomposition

[mynameisissac]

I am a big fan of TSV, thank you for the interesting work. I have a small question about the code implementation.
For the eigen-decompose version, why do we need to sort the singular vectors in ascending before eigen-decomposition?

task_singular_vectors/src/utils/TSVM_utils.py

Lines 443 to 445 in 9e3a822

	sum_s, indices = torch.sort(sum_s, stable=True)
	sum_u = torch.index_select(sum_u, 1, indices)

[AntoAndGar]

Hi, thank you for the kind words.
The eigen-decomposition version needs a sorting step, mostly because the output of torch.linalg.eigh returns the eigenvalues in ascending order instead torch.linalg.svd returns them in descending order, so in a few words, you want to match the correct singular/eigen-vector with the correct singular/eigen-value when you recompute the merged matrix.
Please let me know if you have any doubts or if I have explained it sufficiently. Maybe I can explain this step with more details if needed 😄.