Questions about NCEAverage.py

[LourisXu]

Hi, Thanks a lot for sharing this great code.
In NCEAverage.py, the code is shown as the following:

        if idx is None:
             idx = self.multinomial.draw(batchSize * (self.K + 1)).view(batchSize, -1)
             idx.select(1, 0).copy_(y.data)
        # sample
        weight_l = torch.index_select(self.memory_l, 0, idx.view(-1)).detach()
        weight_l = weight_l.view(batchSize, K + 1, inputSize)
        out_ab = torch.bmm(weight_l, ab.view(batchSize, inputSize, 1))

Select batchsize * (K + 1) features from memory bank and then calculate the similarity matrix. How to determine a positive pair and K negative pairs for each sample?

In MoCo, the aug views of the same image, v1 and v2, are a positive pairs, besides，v1 and other k views from the memory bank of the encoder_k are the k negative pairs. In CMC, this seems to be K + 1 negative pairs, because the memory bank is updated late and it seems that it is not guaranteed to obtain the positive pair view of each sample in the same batch.