Question about conditional sampling and moco-k during training

[andsild]

Hi,

Thanks for the code and paper :)

I'm trying to understand this line, where the model checks for conditional sampling:

conditional_ssl_hist/condssl/builder.py

Lines 144 to 148 in 32c0551

	if self.condition:
	# conditional ssl
	logits = torch.mm(q, k.T) / self.T
	labels = torch.arange(logits.shape[0], dtype=torch.long).cuda()
	return logits, labels

It seems to me that this skips updating the self.queue which I believe corresponds to the "dictionary" in the MoCo paper (section 3.1)

Am I right? If so, does this mean you effectively train with args.moco-k at the same as the batch size (m = 128)? If so, when you compared against an model without conditional sampling in your paper, did you also set moco-k to 128 or was it the default 65536?