Federico Baldassarre

Hi, that sure looks like it should be `edge_index[1]`. I don't remember exactly what those two lines do, but it wouldn't make much sense to compare the same thing twice....

Example to reproduce ```python # main.py import sys import torch import torch.utils.data import torch.distributed import torchmetrics import tqdm class Dataset(torch.utils.data.Dataset): def __init__(self, K, H, W): self.K = K self.H =...

I ran another test on a different machine. System: - NVIDIA Driver Version: 470.86 - CUDA Version: 11.4 - 2x NVIDIA GeForce GTX 1080 Ti Setup: ```bash conda create -y...

Same bug with version 21.1.2 running in linux. Curiously, if the generated multi-page PDF is included in Latex as a figure, only the first page is shown, while opening the...

Hello @SkafteNicki, thanks for looking into the issue and for the proposed solution. However, I wonder if there is some underlying bug with memory management that we are overlooking here....

All right, it's probably related to pytorch then. I don't have enough knowledge of pytorch's internals to open an issue there, but let's hope it gets fixed in future releases....

I've noticed the speed issue, I think it could be related to the left/right side, which at least in this version gives correct results. Anyway, I'll look into speeding up...

Hi, just checked. Here are a few comments about the code: In the pytest test [here](https://github.com/aliutkus/torchsearchsorted/blob/ef21b0c33db45e3de318f272d7e6e07d819b5e78/test/test_searchsorted.py#L27-L36), why do we have separate `Ba, Bv` and then skip the test if `Ba...

Well, I argue that we do care about what happens in case the value to insert is already present in the array. The whole point of the `side` parameter is...

Ah about the broadcasting behavior for the batch dimension. I understand now what that test and the `sel` function were about. I think it could be written more clearly by...