Ferdinand John Briones

I think I sort of get it now. I think the zero array labels is meant to indicate the positive label for one pass. Since at Line 51, they put...

``` random_labels = torch.randint(low=0, high=logits.shape[1], size=(logits.shape[0],1)).to(device) index = torch.arange(logits.shape[0]).to(device).unsqueeze(1) labels_access = torch.cat([index, random_labels], 1) labels_access = torch.transpose(labels_access, 0, 1) temp = logits[tuple(labels_access)] logits[:,0] = temp logits[tuple(labels_access)] = positives.squeeze()` logits =...

I believe their notebook implementation is already suited for that task.

@moured You can try this [notebook](https://github.com/fjbriones/DocSegTr/blob/experiments/experiment.ipynb) I created. You may have to edit some codes in detectron2 to make it work (covert to cpu I think).