if I set alpha the program crash

[stephane-archer]

~/FullyConnected/focalloss.py in forward(self, input, target)
     32 
     33         loss = -1 * (1-pt)**self.gamma * logpt
---> 34         if self.size_average: return loss.mean()
     35         else: return loss.sum()

RuntimeError: cuda runtime error (59) : device-side assert triggered at /opt/conda/conda-bld/pytorch_1524586445097/work/aten/src/THC/generated/../THCReduceAll.cuh:339

[rxqy]

Met exactly the same problem.
I found that the problem is coming from

focal_loss_pytorch/focalloss.py

Line 30 in e11e75b

at = self.alpha.gather(0,target.data.view(-1))

Any ideas? @stephane-archer

[rxqy]

Hi, for anyone who met the same problem, I might have a solution.
You can add one line before the following line
select = (target!=0).type(torch.LongTensor).cuda()

focal_loss_pytorch/focalloss.py

Line 30 in e11e75b

at = self.alpha.gather(0,target.data.view(-1))

and set this line to
at = self.alpha.gather(0,select.data.view(-1))

i.e alpha controls the balance between positive samples and negative samples.
so in the test code, you can set gamma=0, alpha=0.5, and multiple the loss by 2.
This should give a similar result. ( It's also the case in the original implementation of detectron)

The problem is that, when you are doing index select, alpha only has two dimensions
however, if you use target to select the index, you might have more than two labels in your target.
BTW, I'm using pytorch 0.3.1, don't know if it's a bug @clcarwin @stephane-archer