How do I find the model attributes for the elements in the dataset?

[vsa1920]

I am working on training on Cu-dataset as provided here.

They also have a pretrained model available here.

I am sorry, I am a newbie at both running LAMMPS and using DeePMD. Whenever I try to run a LAMMPS simulation using these models, I get this error:

INVALID_ARGUMENT: Tensor spin_attr/ntypes_spin:0, specified in either feed_devices or fetch_devices was not found in the Graph
  >>> Info of model(s):
  using   1 model(s): Cu-dpgen-updated.pb 
  rcut in model:      8
  ntypes in model:    1
ERROR: pair_coeff: element Cu not found in the model (/project/source/lmp/pair_deepmd.cpp:1100)
Last command: pair_coeff     * * Cu

I have been trying to debug this issue for a week now. Any help would be greatly appreciated. Thanks!

[njzjz]

The elements must be defined during the training time to map the elements, otherwise, only the default order can be used. Please read the documentation for details.

Probably, this model wasn't given the type map. CC @AnguseZhang

[vsa1920]

Earlier, I did. Then I retrained a model and tried to use it for a LAMMPS simulation, but it didn't work out. I am still trying out some stuff.

Thank you for the Princeton workshop reference, I will check it out.

[njzjz]

You can also remove atomic types from pair_coeff, i.e. pair_coeff * *, if the type information is not given by the model. The types will be assumed to be the same as training data.

[njzjz]

Could you provide input.json and the version you use to help locate the problem?

[vsa1920]

Thank you so much for your help!
Here's a link to input.json: https://drive.google.com/open?id=1cIdK1-dFkwQ7n8FCqWc_5Q-jqbY1wk-4&usp=drive_copy

[njzjz]

I don't have access to this file.

[vsa1920]

You can also remove atomic types from pair_coeff, i.e. pair_coeff * *, if the type information is not given by the model. The types will be assumed to be the same as training data.

I got this error when I tried:

Generated 0 of 0 mixed pair_coeff terms from geometric mixing rule
ERROR: Not all per-type masses are set. Type 1 is missing. (src/src/atom.cpp:726)
Last command: run 5000

[njzjz]

It's just a LAMMPS error. It seems that you don't set the mass for atomic types.

[vsa1920]

It's just a LAMMPS error. It seems that you don't set the mass for atomic types.

Hi! Thanks again for your help. It's invaluable for someone who is new to both LAMMPS and DeePMD. I am trying to run this on a local GPU cluster. Earlier, I was able to run this on Colab GPU but that was still incredibly slow ~ 500 LAMMPS steps in a small system size in 8 minutes. I am getting this error:

2023-07-31 15:59:35.884004: E tensorflow/core/common_runtime/session.cc:91] Failed to create session: INTERNAL: cudaGetDevice() failed. Status: cudaGetErrorString symbol not found.
INTERNAL: cudaGetDevice() failed. Status: cudaGetErrorString symbol not found.
ERROR on proc 0: DeePMD-kit C API Error: DeePMD-kit Error: TensorFlow Error: INTERNAL: cudaGetDevice() failed. Status: cudaGetErrorString symbol not found. (/project/source/lmp/pair_deepmd.cpp:838)
Last command: pair_style deepmd Cu-dpgen-model-updated.pb

Any thoughts on this?

[njzjz]

It should be CUDA version mismatch. What's your CUDA version and how did you install deepmd-kit?

[vsa1920]

I used easy install using pip to install deepmd; using - pip install deepmd-kit[gpu,cu11,lmp]
The cuda version on the machine is 12.0, same as on colab. I checked through nvidia-smi.

[njzjz]

The laster stable TF version only supports CUDA 11.x. You can check LD_LIBRARY_PATH and see whether CUDA 12.0 is loaded.

[vsa1920]

I was checking the CUDA version from LD_LIBRARY_PATH and its 11.0. I am wondering if I did it correctly: