A small question about the code for training on SOP

[Dyfine]

Hi, thanks very much for sharing the code. When I use it to train models on SOP dataset, I get unexpected low results. I check the code and find in ProxyGML/loss/ProxyGML.py that one line (line36) has been commented in the function below.

ProxyGML/loss/ProxyGML.py

Lines 35 to 39 in f26101e

	def scale_mask_softmax(self,tensor,mask,softmax_dim,scale=1.0):
	#scale = 1.0 if self.opt.dataset != "online_products" else 20.0
	scale_mask_exp_tensor = torch.exp(tensor* scale) * mask.detach()
	scale_mask_softmax_tensor = scale_mask_exp_tensor / (1e-8 + torch.sum(scale_mask_exp_tensor, dim=softmax_dim)).unsqueeze(softmax_dim)
	return scale_mask_softmax_tensor

I uncomment this line and get the expected results, i.e., 78.0 R@1. So I think you may comment this line by mistake?

[YuehuaZhu]

Thank you for your attention and you are right.