Zeming Lin

Forward takes Variant class

16

One of the bigger refactors, we introduce a variant class to wrap inputs and outputs in, so you at least get run time type checking of your arguments, much like...

They don't own the objects, which means you can't really create and pass in a TensorList, unless it's backed by a owning list with the tensors. Just use a std::vector....

I think there is no way to realistically do this, since the parameters are added outside of the container. Best to just error out, or at least warn. Specifically, you...

This let's people do tricky things if they don't want to have a model to contain the parameters.

Implement the rest of the optimizers

2

All the optimizers listed [here](http://pytorch.org/docs/master/optim.html#algorithms): - [x] SGD - [x] Adam - [x] RMSProp - [x] Adagrad - [ ] LBFGS - [ ] Adadelta - [ ] SparseAdam -...

It not only displays an error, but sometimes will hang. We should get rid of the error message and only set seeds for visible devices.

Apparently a lot of sparse operators don't work very well, and none of the optimizers are coded with respect to sparse support.

Implement LookupTable

2

Sparse gradients are the hardest part.

pdbx get_structure with use_author_fields gives different results

1

On 7yxg, if we load the CIF file with `pdbx.get_structure(file, use_author_fields=False)` vs `pdbx.get_structure(file, use_author_fields=True)`, we get a different number of atoms. I believe this is because the author residue index...