The size of tensor Runtime Error

[umichwbaek]

Hi

I got an Runtime error while running the nequip-train of my work.

File "/home/software/anaconda3/envs/NequiP/lib/python3.10/site-packages/nequip/train/metrics.py", line 183, in call error_N = error / N
RuntimeError: The size of tensor a (350) must match the size of tensor b (5) at non-singleton dimension 0

I found that the error comes from the metrics_components. When I mute metrics_components, it runs without the errors, but I want to see the output with components I put.

Are there have any additional parameters that I need to set?

[Linux-cpp-lisp]

Hi @umichwbaek ,

Can you please provide more details about your config and dataset?

[umichwbaek]

I used almost the same input parameters of example.yaml in allegro except for the following:

r_max: 7.0

dataset: ase
ase_args:
format: extxyz
key_mapping:
z: atomic_numbers
E: total_energy
F: forces
R: pos

validation_dataset: ase

n_train: 160
n_val: 1044

batch_size: 2
max_epochs: 1000

metrics_components:

• • total_energy
  • mae
  • PerAtom: True
• • total_energy
  • rmse
  • PerAtom: True
• • forces
  • mae
  • PerAtom: True
• • forces
  • rmse
  • PerAtom: True

The type of train and validation dataset is xyz. I have separate validation dataset file, so I did not use 'train_val_split'. It did not show any dataset error, so the format would not be the problem

[Linux-cpp-lisp]

Ah I see it--- you should not set PerAtom: True for metrics on per-atom quantities like forces. Perhaps that is a little unclear, but it specifically is only for getting the version of a global quantity normalized by the number of atoms. I'll try to add a clearer error.

[umichwbaek]

Oh I see. Then how can I get per-atom values of mse and rmse? Is it the mse or rmse divided by the total number of atoms of the sum of whole data, or average?

[Linux-cpp-lisp]

You mean you want to report the force RMSE on a specific atom?

[umichwbaek]

Both. For overall and specific if possible.

[Linux-cpp-lisp]

What do you mean by "For overall and specific if possible."?

[umichwbaek]

I want to know 'rmse/the number of all atom' and if possible 'rmse/individual atoms'.

[Linux-cpp-lisp]

Is that supposed to be division in the first case and not in the second?

If you want to do some post processing on per-atom level errors, see https://github.com/mir-group/nequip/blob/develop/examples/parity_plot.py for an example. If you want to track the error of one specific atom during training, that will take a little bit of coding (though not much). If you want per-species errors, see https://github.com/mir-group/nequip/blob/main/configs/full.yaml#LL259C31-L259C31