slryou41 / anchorloss Goto Github PK

Hi, thanks for your awesome work!
I have one question when reading paper. In human pose estimation training, it said

Note that we put sigmoid activation layer on top of the standard architecture to perform classifacation.

The question is, how you process learning label? do you use gaussian heatmap or circle plate as used in grmi?

Much thanks!

Dear Authors,
with all respect to your work, based on my experiments I feel that eighter the presented text in the paper does not reflect the reality or the code that you have published contain bugs. For the purpose of empirical comparison with the other loss functions for my research I have tried to train Resnet50 on Cifar10, with the warm-up, without, with commenting log.sum(1), with clamping log(1 - pt)... in all cases, the loss is going to nan... Moreover, I can see in the issues that its not only for me, its common, no one can use the proposed loss and Im not even talking about improvements...
It is very frustrating to see that the research publish in such a high venue as ICCV is not only not reproducible but just simply not working (at least implementation provided by the authors).

I read your paper on anchor loss with curiosity.
Right now I'm trying to implement the code using tensorflow, and I came up with some questions.

1. For logpt_neg, you calculated F.logsigmoid(1-input), but as far as I understand, if the input parameter is non-sigmoid logits, then it should be torch.log(1-sig(input)), isn't it?

2. Did you experience any 'NaN' loss problem during training? if you did, what was your solution for the problem?

thanks for your sharing.
Could you tell me how to apply the anchor loss in pose estimation?

My input is heatmap, it's shape : N, C, H, W. and my target also is N, C, H, W.

And, how to apply the anchor loss in pose estimation?

looking forward your reply!

thx for your sharing
could you please tell me how to apply anchor loss in pose estimation?
i try to use MSELoss multiply by anchor_weight which is :
anchor_weight = torch.where(
torch.eq(targets, 0),
torch.pow(1 + predictions - pt_pos, gamma),
torch.full_like(targets,1)
)
the result is unideal..

thank you very much

Hi, thanks for your work.
I used this anchor loss in my project to do classification job, I have tried set warmup = False and warmup = True, after several epochs, both loss become Nan. I have changed nothing in the source code anchor_loss.py. So, could you please give me some advice how can I fix it? The loss when setting warmup = True like this:

Thanks a lot!

Hi I troubled with loss is NAN(first epoch loss is 5021 ,it is too large).Can you give me some advices?
Thanks very much.