With some low (but non-trivial) frequency, processing through the convolutional layers diverge and node attributes become a mixture of nan and inf (which seem to all turn to nan). This later throws an exception during the Kabsch transform, which ruins results for the whole complex. Setting these to 0 basically skips an iteration, at worst it ruins one of the sampled complexes, but leaves the others. Note this is only applied to the main model, *not* the confidence model.
* Ensure we calculate rotatable bonds on the version of the ligand with no hydrogens. Also fix spelling of rotable -> rotatable. Closes GH-220 (@Nobody-Zhang)
* Vectorize SO3 calculations. Closes PR GH-218 (@tornikeo)
* Pin pytorch-lightning version. Closes GH-193 (@mikael-h-christensen)
* Guard against divide by zero in torus.py. Closes GH-161 (@amorehead)
* Update e3nn version to 0.5.1. Closes GH-155 (@amorehead)
* Add a little more info on docker container to README.md
