How to advancing like this?

1

For example, I use numpy:

2017-12-05-170222_544x733_scrot
2017-12-05-170222_544x733_scrot544×733 41.1 KB

Then I would like to do it in PyTorch, but it failed, I do not know how to solve it.

2017-12-05-170351_1042x765_scrot
2017-12-05-170351_1042x765_scrot1042×765 47 KB

I know this is duplicated with How to get a matrix top-n in its second axis? However, topk still does not meet my real need that I wand to use b to index another variablec too. For example:

2017-12-05-185038_596x158_scrot

@tom Hello, could you answer this question too? thank you!

2

My advand indexing in numpy is wrong too, forget it.

3

That is a weird thing to do. Did you intend to extract smallest 2 element for each row? If so, use this one liner

a.topk(2, dim=1, largest=False)

if you don’t need result to be sorted, you can add sorted=False (defaulted to True).

4

What I want is below:

if CONFIG[config_key].RPN_PRE_NMS_TOP_N > 0:
    # TODO: omit 0 dim bug for PyTorch
    _, order = probs.sort(1, descending=True)
    keep = order[:, :CONFIG[config_key].RPN_PRE_NMS_TOP_N]
    index = torch.arange(0, boxes.size()[0]).unsqueeze(1).long().cuda()
    boxes = boxes[index, keep, :]
    probs = probs[index, keep]

So it is solved, but thank you anyway! @SimonW