Should the loss, do a mean over all the classes of an anchor. For example if data comes in like this:
N * (Anum_classes) * H * W, where A is the number of anchors per cell. H, = height of feature map and W = width of feature map. Our targets are NAHW.

Now, we need to do a mean over each the num_classes of each A, of our loss computed over our N * (Anum_classes) * H * W input, to get it to NAHW.

This NAH*W should be summed over and then normalised by the number of anchors assigned to a ground truth box only, as the paper states the rest of the value is effectively zero.

When I run this program , an error occurs:

"NameError: name 'long' is not defined"

it points to line 11 of file focalloss.py:

if isinstance(alpha,(float,int,long)): self.alpha = torch.Tensor([alpha,1-alpha])

Thank you very much @clcarwin

gamma, alpha and size_average.
What are your recommendation when you init the loss function?
How gamma and alpha affect the result?

During the adjustment process of gamma, when the value is between 0-1, the loss function has a gradient explosion. If it exceeds 1, it will not. May I ask what caused it

• your code:
  

  focal_loss_pytorch/focalloss.py

  Lines 17 to 19 in e11e75b

  	input = input.view(input.size(0),input.size(1),-1) # N,C,H,W => N,C,H*W
  	input = input.transpose(1,2) # N,C,H*W => N,H*W,C
  	input = input.contiguous().view(-1,input.size(2)) # N,H*W,C => N*H*W,C

hi, thank you for your code. and i have a question, why you can`t reshape the tensor from N,C,H,W => N* H* W,C directly

Hi 👋

I've accidentally stumbled upon your code. Looks nice, but I have a question though.

Why do you use Softmax instead of Sigmoid function?

In the paper authors clearly stated the usage of sigmoid function, which makes sense. On the contrary, softmax doesn't (at least, it seems to me)

As a reference, implementation in fvcore has sigmoid function

I would be glad if someone could explain to me, whether it's a bug or there's some intuition behind this.

Thanks in advance!

Hello! Thank for your focal loss implementation. But I have question. I guessed we also have to consider negative cases of confidence in loss, I mean cases when pt = 1 - p. I see there are only pt = p(positive) cases, but no pt = 1 - p cases. Could you comment it?

I notice that it seems focal_loss.py only implements focal loss on the binary classification setting. Do there exist any implementations about focal loss for a multi-class classification setting?

In my experiments, the the loss of FocalLoss with gamma=0 is much lower than the loss of CrossEntropyLoss. What makes it?

The paper mentions that the loss layer is combined with the sigmoid computation and not softmax. More speciafically this line

Finally,
we note that the implementation of the loss layer combines
the sigmoid operation for computing p with the loss computation, resulting in greater numerical stability.

So isn't the author saying that we should use sigmoid activation over the last layer. The softmax usage maybe could lead to a lower accuracy.

It seems the loss=-alpha*(1-y')^{gamma}log(y') when y=1 in the paper.
But in the code:
loss=-(1-alpha)*(1-y')^{gamma}log(y') when y=1
So the alpha should be set in an oppisite way?...
In the code:
{
(alpha,(float,int)): self.alpha = torch.Tensor([alpha,1-alpha]),
at = self.alpha.gather(0,target.data.view(-1))
logpt = logpt * Variable(at)
}
Is this right?

I find the calculation of pt using the data, which means your implementation not allowing pt Backpropagatio. So I argue should this allow bp?