IndexError when predicting monomers with more than 3 recycles about alphapulldown HOT 3 CLOSED

Unfortunately, this problem roots in AlphaFold's module.py, as shown here: https://github.com/google-deepmind/alphafold/blob/632ef575c64eff9eb5ed96c8c7b055bf675421ac/alphafold/model/modules.py#L331

In the case of monomer modelling, Deepmind never exposed num_recycles to the users thus num_recycles is always hard-coded as 3 in the configuration. The batch size is 4 in the above snipped, hard-coded as well by Deepmind. Thus, when num_recycles is smaller or equals to 3, this line of num_ensemble = batch_size // (self.config.num_recycle + 1) assigns num_ensemble to 1. However, if num_recyle is greater than 3, then num_ensemble = batch_size // (self.config.num_recycle + 1) will make num_ensemble 0. num_ensemble=0 will mess up the downstream data processing like here:
https://github.com/google-deepmind/alphafold/blob/632ef575c64eff9eb5ed96c8c7b055bf675421ac/alphafold/model/modules.py#L336
and here https://github.com/google-deepmind/alphafold/blob/632ef575c64eff9eb5ed96c8c7b055bf675421ac/alphafold/model/modules.py#L166

Multimer modelling works for whichever number of recyles cuz in multimer_module.py, this step is completely got rid off.
I think the solution now is that we should add an extra check in AlphaPulldown so that if it models a monomer structure, we set the upper limit of num_recyles.

from alphapulldown.

Crazy! I think ColabFold might have fixed that by editing this line:
https://github.com/sokrypton/alphafold/blob/7c09316f71893e532bc15528cce4c1d3cb705f8f/alphafold/model/modules.py#L421
Do I understand that correct? Maybe we could do the same? Allowing for more recycles for monomers is a useful feature, esp. if they are big.

from alphapulldown.

Crazy! I think ColabFold might have fixed that by editing this line: https://github.com/sokrypton/alphafold/blob/7c09316f71893e532bc15528cce4c1d3cb705f8f/alphafold/model/modules.py#L421 Do I understand that correct? Maybe we could do the same? Allowing for more recycles for monomers is a useful feature, esp. if they are big.

Thanks Jan you're right ColabFold guys got rid of this limit in these lines. We can do the same but it has to be done on our AlphaFold"s fork instead of AlphaPulldown.

from alphapulldown.