NaNs for CTCF and other multibatch cases when number of rounds is one

[ilibarra]

The behavior, when the number of rounds was only one, was some weights becoming NaNs during training, for log_etas weights, particularly due to this step in the code.

It was something associated with any of these lines. I think it was the normalization being always one.

mubind/mubind/models/models.py

Lines 1635 to 1649 in 270dc2a

	out = None
	if self.enr_series:
	out = torch.cumprod(binding_scores, dim=1) # cum product between rounds 0 and N
	else:
	out = binding_scores
	# multiplication in one step
	etas = torch.exp(self.log_etas)
	out = out * etas[batch, :]
	# fluorescent data e.g. PBM, does not require scaling, to keep numbers beyond range [0 - 1]
	if not kwargs.get('scale_countsum', True):
	return out
	results = out.T / torch.sum(out, dim=1)

If prevented for the CTCF and discarding the column [1], one could then load multiple samples with non-similar k-mers, and only one round.
https://github.com/theislab/mubind/blob/fix-scatac/notebooks/batch/01_CTCF_two_batches.ipynb

@johschnee do you remember/can you point out, in case you remember what the crucial step was, and if a likely fix is possible? Thank you.