1top and 5top errors in Pytorch

1

Hi all, How can I calculate the k-top error in Pytorch? any help would be appreciated.

2

EDIT: This is likely not what you’re looking for. This is a top-k function in PyTorch to get the highest value of a Tensor (including a loss Tensor), but likely you’re looking for the meaning of “top1” and “top5” accuracy like in ImageNet. See my reply below.

Hello, have a look at @InnovArul solution to my similar question here.

You would have to slightly adapt it, but it’s the same basic approach outlined under “Ignoring top-k losses”:

import torch
loss = torch.rand(6, requires_grad=True)
k = 3
sorted_loss, indices = torch.sort(loss)
top_k_losses = sorted_loss[-k:]  # this is how you identify the top losses

Depending on your specific use case, you can also look at torch.topk.

1 Like
5

Hi, thank you for your response. So, is the top1 error, the largest error value?

6

Hi, thank you for your response. So, is the top1 error, the largest error value?

Yes, should be.

1 Like
8

What about top5? For it there are 5 values but in the technical report or research paper one value as top5 error is reported. Is it the avarage of the five values?

9

Hmm, I think your original question was motivated by something different than the highest loss.

Likely what you’re looking for with top1 / top5 is this instead:

  • top 1 means whether the highest probability model prediction matches the label
  • top 5 means that one of the 5 highest probability predictions match the label

See more here.

1 Like