Brian DeCost

Results 5 issues of Brian DeCost

this PR addresses autograd forces in ALIGNNAtomWise, in light of discussion at #113 - moves autograd toggle on relative displacement vectors to the start of ALIGNNAtomWise.forward - re-computes bond angle...

Benchmark ReLU vs Swish performance and workload -- I ran some informal comparisons between ReLU networks and swish networks this summer, but didn't record computational workload or run the tests...

[current version logging](https://github.com/usnistgov/alignn/blob/4e5e1ea47ae379204480e0e5a10244c158993d06/alignn/config.py#L21) assumes that code is running from a working directory under the alignn repository, and logs the current commit hash to the full configuration datafile. to support use...

we've ended up with a lot of complex logic in the main training script -- I think we should consider moving ignite event handler definitions out of the main `train_dgl`...

e.g. the MEGNet dataset is rather large; so are OQMD and AFLOW proposal: store graph data in hdf5 using dataset index as main dataset keys Alternate key: use the structure...