Why do you evaluate on cpu(not on gpu)?

[Jumabek]

Hi @kuangliu, in the evaluation code you are decoding boxes on the cpu . https://github.com/kuangliu/torchcv/blob/master/examples/fpnssd/eval.py#L57 My question is why don't use gpu? UPDATE 1: below measurement is actually incorrrect, because in the beginning of the evaluation most of the time is spent on waiting to load the images.

I loaded anchor boxes to GPU and did box decoding on GPU for the bach_size=1. Here is the result:

CPU : 17sec GPU: 7 sec

This is for the batch size of 1, I am sure using larger batch_sizes will give more gain in time for GPU