Closed k920049 closed 8 months ago
According to the implementation, you used torch.nn.functional.cross_entropy for computation. This ignores off-diagonal components in similarity matrix. I implemented a NCE version of this and want to contribute to this repo.
torch.nn.functional.cross_entropy
According to the implementation, you used
torch.nn.functional.cross_entropy
for computation. This ignores off-diagonal components in similarity matrix. I implemented a NCE version of this and want to contribute to this repo.