Did you try use alignment and uniformity as loss instead?

[jiqiujia]

In my dataset, I found in-batch sampled softmax loss donesn't work, and using alignment and uniformity as loss works well. I doubt if it related to the number of items.

[yxni98]

Hello @jiqiujia . I also tried the latter one before. Alignment and uniformity provide good guidance for learning sometimes, that's why they are kept in our code. Though in-batch helps accelerate the convergence, the learning performance highly relies on item distribution within datasets.