NAN values in NNP training

[zhenming-xu]

Hello! Recently, I encountered a problem of some NAN values in the first column of the weights.XXX.data file when NNP training.

The following is my NNP training paramaters,

"nnp.train(train_structures=train_structures, train_energies=train_energies, train_forces=train_forces, cutoff_type=1, r_etas=[0.5,2.0], a_etas=[0.5,2.0], r_shift=[0.0], zetas=[1.0,4.0], r_cut=4.2, hidden_layers=[4,4], epochs=5)"

Very strangely, running the example(https://github.com/materialsvirtuallab/maml/blob/master/notebooks/pes/nnp/example.ipynb) in my cumputer is OK!

[JiQi535]

Hello! Recently, I encountered a problem of some NAN values in the first column of the weights.XXX.data file when NNP training.

The following is my NNP training paramaters,

"nnp.train(train_structures=train_structures, train_energies=train_energies, train_forces=train_forces, cutoff_type=1, r_etas=[0.5,2.0], a_etas=[0.5,2.0], r_shift=[0.0], zetas=[1.0,4.0], r_cut=4.2, hidden_layers=[4,4], epochs=5)"

Very strangely, running the example(https://github.com/materialsvirtuallab/maml/blob/master/notebooks/pes/nnp/example.ipynb) in my cumputer is OK!

Hi Zhenming, thanks for reporting this issue. Since you mentioned that the notebook can be correctly executed, I suspect that the input you gave to the nnp.train() method could be problematic. I'm not sure how you chose the parameters like r_shift, but please make sure they are reasonable values. For example, your setting of r_shift=[0.0] is not found in the recommended settings given by n2p2 (https://github.com/CompPhysVienna/n2p2/blob/master/examples/input.nn.recommended).

You may try the default recommended settings for the symmetry function and see if the error exists. If so, please report this issue to the n2p2 repo (https://github.com/CompPhysVienna/n2p2), as the weight.XXX.data file is generated in the n2p2 training process by the nnp executable.

Please kindly let us know how you figured out the issue. Thanks.

[JiQi535]

Hi @zhenming-xu , can I ask if this issue is solved?

[zhenming-xu]

Hi,

I have submitted this issue to n2p2, https://github.com/CompPhysVienna/n2p2/issues/164 https://github.com/CompPhysVienna/n2p2/issues/164 Now, this is no reply.

Regards!

Zhenming Xu

@.***

2022年4月27日 02:26，JiQi535 @.***> 写道：

Hi @zhenming-xu https://github.com/zhenming-xu , can I ask if this issue is solved?

— Reply to this email directly, view it on GitHub https://github.com/materialsvirtuallab/maml/issues/425#issuecomment-1110117721, or unsubscribe https://github.com/notifications/unsubscribe-auth/AHGEUOO7CXLWORIS2M3GYNDVHAYO5ANCNFSM5SI7IUWQ. You are receiving this because you were mentioned.