About The InfoNCE Loss temperature parameter Setting of τ

[wlfxy]

Thank you very much for your paper and code, and I would like to ask you about The InfoNCE Loss Does temperature parameter τ have an initial value? According to the thesis, it is a learnable parameter. How does it learn by itself

[Skyy93]

Thank you for your interest in our work. We set the temperature parameter in the model.py to

self.logit_scale = torch.nn.Parameter(torch.ones([]) * np.log(1 / 0.07))

Since we use it in the loss calculation we can learn it.

[wlfxy]

Thank you for your answer