Some doubts about your programs about racnn-pytorch

[jeong-tae]

the 'i' in https://github.com/jeong-tae/RACNN-pytorch/blob/master/models/Loss.py#L7 is for the each scales. I currently use 3 different scales. so 'i' is range in (0, 3)
labels are same for each scales.

[Ostnie]

You know, I still have some doubt about backpropagation for APN, I think it is incorrect to use a model's result as another model's loss, all the parameters should be calculated by the chain rule in backwardpropagation and Lrank_loss does not match this.

But in formula 11

I think there is no way to calculate the left side of the formula, but the right side of the formula can be calculated. You said that your apn does not converging well, and I didn't find a conversion method like formula 11 in your program. I don't know if this is the cause of the problem, this is just a guess .

[jeong-tae]

I also can't sure to calculate formula 11. i just refer the original caffe code for the attention crop backward

[jeong-tae]

i will close this issue. if you still have any question, re-open or make a new issue to contribute.
i will wait a good news to improve current repository. thx