When I ran the code, some errors occurred. about bert-kpe HOT 3 CLOSED

Hi Leo, thanks for using our code.

I don't think it's a dimension error. Take the error in the figure as an example, input1 (size=[1,1]), input2 (size[1,431]), and flag (size=[1]), actually, margin_ranking_loss will scale them to the same dimensions, i.e. input1 (size=[1,431]), input2 (size=[1,431]) and flag (size=[1,431]).

Here, I try to reproduce the error, but the command ran properly without bugs. (PyTorch=1.8)
Could you provide your PyTorch version or more implementation details?

from bert-kpe.

@SunSiShining Thank you for your reply. My PyTorch version is 1.11.0. And I took a look at source code of PyTorch 1.8 and PyTorch 1.11. Find out that the MarginRankingLoss class in PyTorch 1.11 will check the input dimension before processing. I solve that problem by using 'torch.margin_ranking_loss' function.

input_tensor = torch.randn(3).unsqueeze(-1)
input_tensor1 = torch.randn(3).unsqueeze(0)
targe_tensor = torch.FloatTensor([1])
print(input_tensor.size())
print(input_tensor1.size())
print(torch.margin_ranking_loss(input_tensor, input_tensor1, targe_tensor, 1, 1))    # 1 stands for reduction = 'mean'

from bert-kpe.

Wow, that's great! thanks for letting me know.

from bert-kpe.