Felix Musil

I implemented a [TorchScript-able neighbor list](https://github.com/felixmusil/torch_nl) recently that is as general as ASE's and it is somewhat faster. It might be useful while waiting for a general NNPOps implementation of...

Thanks for the quick reply. > Have you run explicit checks that torch.nn.functional.x and torch.nn.X give identical answers? Running this snippet shows that the results are the same within machine...

Hi @Linux-cpp-lisp, Sorry for disappearing on you. I reverted the unnecessary commits and updated the PR branch to the current `develop`. So it's finally ready to be merged unless you...

> Re multi-GPU support, take a look at this recent issue: https://github.com/mir-group/nequip/issues/210. Basically, unless you are really married to PyTorch Lightning, I would recommend looking into our horovod branch, which...

Hello, could you try to set `"expansion_by_species_method":"structure wise"` in your input for `SphericalInvariants` ? In principle, it should give similar timings as using directly numpy. If it does not then...

Looks ok to me, thank for your tests. @Luthaf, @ceriottm and @agoscinski what do you think ? We could also use a numpy implementation rather than the rascal one in...

A simpler (in terms of code modifications) alternative would be to use a variadic templated external function that takes the calculator and mangers as argument, forwards the template parameters for...

Right, I got these runtime errors compiling/running test_calculator. Also I use the GCC 9 available to Ubuntu 18.04. I did not try other compilers as I was investigating a weird...

In release mode only, debug mode did not have the same bug so I did not test this either. I am not sure about the version of the GCC 9...

I ran the same comparison with the latest version of QE/GIPAW 6.1 and there are no problems there. My collaborators and I ran many calculations with QE 6.1 svn rev....