Open
Description
🐛 Describe the bug
Hi this is an old bug described in pull request #8562 .
As mentioned, the implementation of select function in sagpool reuses the select class of topkpool, which introduces an extra learnable weight. As a consequence, this select function learns to take the opposite of the scores if the weight is negative, which is not consistent with the paper, neither the codes of the authors. In my observation, this change influences the effectiveness of sagpool significantly in some situations.
Considering the future work may use the sagpool layer in Pyg for evaluation, would you mind fixing this bug so that it doesn't influence their results and conclusion?
Versions
I cant get the collect_env tool. The connection refused.