Add grace to train - #663
Conversation
There was a problem hiding this comment.
This file would not be required if we upgraded to tensorpotential 0.5.5 since that add extxyz support.
It is small though, and it requires various modifications to the xyz's we have as well
There was a problem hiding this comment.
We definitely should upgrade at some point, but it introduces conflicts with basically everything else via PyTorch/CUDA conflicts (see ICAMS/grace-tensorpotential#23), if I remember correctly
This may be more tractable once we cut out some of the unsupported MLIPs from our extras
6602b00 to
ab4fba4
Compare
Now mace and nequip both output to ./janus_results by default.
Co-authored-by: Jacob Wilkins <46597752+oerc0122@users.noreply.github.com>
Co-authored-by: Elliott Kasoar <45317199+ElliottKasoar@users.noreply.github.com>
ab4fba4 to
6a7e525
Compare
Might as well do this Co-authored-by: Jacob Wilkins <46597752+oerc0122@users.noreply.github.com>
Now mace and nequip both output to ./janus_results by default.
Co-authored-by: Jacob Wilkins <46597752+oerc0122@users.noreply.github.com>
6a7e525 to
1726b85
Compare
|
@harveydevereux I'm not expecting you to finish this off at this point, but if you have a second, can you remind me what still needs doing? Apart from now needing to rebase, was it mostly just the testing/docs (and trying to upgrade GRACE)? |
2 participants
Grace is quite slow it seems (60s for 1 epoch fine-tuning) and memory hungry (13Gb)