BUG in the code

[panmianzhi]

Thanks for your solid work! I think there is a potential bug in the implementation of ConR. For one anchor, if there is no positive samples in the batch, then denorm (as follows) will be zero, which consequently causes division by zero and the loss will be nan.

  # Loss = sum over all samples in the batch (sum over (positive dot product/(negative dot product+positive dot product)))
  denom = pos_i.sum(1)

[mahsakeramati]

@panmianzhi Hi, Thank you very much for your interest in our work. In this implementation, there always would be at least one positive pair. The reason is that for each training sample, we create 2 augmented samples, one is considered as a potential anchor and the other one is considered as a positive pair.