How `num_recycles` work in UnimolConfGModel?

[saltball]

I'm trying to use UnimolConfGModel for molecular conformation generation because my molecules are hard to generate initial conformations with rdkit, so I'm trying to hack into the model forward and use coords2unimol to generate inputs.

I found something strange, the parameter num_recycles of the model doesn't seem to have anything to do with convergence, the atoms of the output seem to push together when num_recycles value is large.

I want to make sure that I should only use the default value of num_recycles (4) and that this value has to be the same during training and with testing if the model is trained?

[guolinke]

yeah, it is better to use the same num_recycles for training and inference.

[saltball]

Thanks for the reply, I guess I can assume that the model is learning to get results at specific num_recycles? Is this considered fixed as a model structural parameter?