Hi,
Thanks for your excellent work! One simple question here. When I read the code, I found that the focal loss code:
def sigmoid_focal_loss_cpu(logits, targets, gamma, alpha): num_classes = logits.shape[1] gamma = gamma alpha = alpha dtype = targets.dtype device = targets.device class_range = torch.arange(1, num_classes+1, dtype=dtype, device=device).unsqueeze(0) t = targets.unsqueeze(1) p = torch.sigmoid(logits) term1 = (1 - p) ** gamma * torch.log(p) term2 = p ** gamma * torch.log(1 - p) return -(t == class_range).float() * term1 * alpha - ((t != class_range) * (t >= 0)).float() * term2 * (1 - alpha)
You name it as sigmoid_focal_loss_cpu.
I am a little bit confused since this code can be run on GPU (if the input tensor device is GPU). Am I right?
Many Thanks!
Hi, Thanks for your excellent work! One simple question here. When I read the code, I found that the focal loss code:
def sigmoid_focal_loss_cpu(logits, targets, gamma, alpha): num_classes = logits.shape[1] gamma = gamma alpha = alpha dtype = targets.dtype device = targets.device class_range = torch.arange(1, num_classes+1, dtype=dtype, device=device).unsqueeze(0) t = targets.unsqueeze(1) p = torch.sigmoid(logits) term1 = (1 - p) ** gamma * torch.log(p) term2 = p ** gamma * torch.log(1 - p) return -(t == class_range).float() * term1 * alpha - ((t != class_range) * (t >= 0)).float() * term2 * (1 - alpha)You name it assigmoid_focal_loss_cpu. I am a little bit confused since this code can be run on GPU (if the input tensor device is GPU). Am I right? Many Thanks!