[Warning Report] "is" with a literal about sslrec HOT 5 CLOSED

A cleaner approach is to add a small epsilon to the degree to avoid zeros, before taking the reciprocal square root.

Here is one way I provide to fix it:

def _normalize_adj(self, mat):
    # Add epsilon to avoid divide by zero
    degree = np.array(mat.sum(axis=-1)) + 1e-10
    d_inv_sqrt = np.reshape(1.0 / np.sqrt(degree), [-1])
    d_inv_sqrt[np.isinf(d_inv_sqrt)] = 0.0
    d_inv_sqrt_mat = sp.diags(d_inv_sqrt)
    return mat.dot(d_inv_sqrt_mat).transpose().dot(d_inv_sqrt_mat).tocoo()

By adding a small epsilon value, it guarantees there are no zeros in degree, avoiding the divide by zero.

The tiny epsilon doesn't meaningfully change the normalization, but avoids the need to handle inf/NaN values separately.

I hope it will help. By the way, I wanted to mention that I originally intended to submit code to a dev branch, but noticed that the project only has a master branch. In the future, would you consider adding a dev branch or something similar to facilitate the contributions? Thank you.

from sslrec.

Hi!

Thanks for your interest in SSLRec, and we appreciate your contribution!
Perhaps you could consider submitting a pull request so that we can merge your changes into the project ;)

Best regards,
Xubin

from sslrec.

Of course, I am more than happy to make a contribution. Currently, I have already submitted a pull request #7. If there are any issues with the code, please let me know. Thanks.

from sslrec.

Hi! Thank you for your contribution. It has been merged into the main branch :)

from sslrec.

Thank you for your recognition, good night~

from sslrec.