Liu Songtao

@najingligong1111 your lr is too large, did you change something in the code?

Great, thanks for your verification！

@ys0823 It seems a bug exists on the voc evaluator. I will check this at the weekend.

I have met the NaN loss problem when I use large lr. And this problem occurs in the first epoch only when I use FP16 training with mobilenet. So I...

Yes, it's worth to try first.

What's your batchsize. And you disable the sycn bn, mixup, smooth label, and other tricks? The loss becomes NaN in the 4th epoch, so I guess the LR is still...

You only have one gpu and the total batchsize is 5? With sycn bn, the equivalent total batch size of my model is 64 or 60 (with 4 v100 or...

You may take a look at the "warm-up" strategy.

Oh...yes, you are right, it's my fault. Thank you for pointing out the error, I will fix this error in the paper, and I may also run their code to...

I can't reproduce your error. I notice your test size is 240, in yolov3, we only use the test size that can be divisible by 32.