Error running ase interface

[marian-code]

Hi I am getting this error when trying to compute predictions for Ge crystal frames using dpdata.

Traceback (most recent call last):
File "/opt/miniconda3/lib/python3.8/site-packages/tensorflow/python/framework/importer.py", line 496, in _import_graph_def_internal
results = c_api.TF_GraphImportGraphDefWithResults(
tensorflow.python.framework.errors_impl.InvalidArgumentError: NodeDef mentions attr 'T' not in Op<name=DescrptSeA; signature=coord:double, type:int32, natoms:int32, box:double, mesh:int32, davg:double, dstd:double -> descrpt:double, descrpt_deriv:double, rij:double, nlist:int32; attr=rcut_a:float; attr=rcut_r:float; attr=rcut_r_smth:float; attr=sel_a:list(int); attr=sel_r:list(int)>; NodeDef: {{node load/DescrptSeA}}. (Check whether your GraphDef-interpreting binary is up to date with your GraphDef-generating binary.).

I am using deepmd kit version 1.3.3 with GPU acceleration and cuda10.1. This is a fresh installation.

The commands are just:

dp train control_file.yaml
dp freeze -d . -o model.pb

I have an similar instalation on another PC with deepmdkit v1.3.2 and repeating this procedure there works OK.

[njzjz]

Do you use model v1.3 in v1.2 program?

[njzjz]

Could you post all your installation scripts? I'm going to see what happened here.

[marian-code]

Yes of course. I have attached all the important files in zip attachement (files.zip), except the training files as those are too big.

I have a mininonda instalation. See base.yml for exported base environment, although I don't know if this is really important.
The environment I am using for training and inference is exported in file dpmd_gpu.yml. I have also included one trained potential file in zip which is throwing an error for me.

Here are all the steps needed to reproduce the error:

conda env create -f dpmd_gpu.yml
conda activate dpmd_gpu
dp train ge_se_a.yaml  # the training file also in zip attachement
dp freeze -d . -o ge_all.pb

The number of batches in the training file is unrealistically small but this is just a minimal example, the problem persists if the number of batches is raised to reasonable numbers e.g. 10^6

Then to get the error you just have to run this snippet in python:

from deepmd.calculator import DP
calc = DP(model="ge_all_aa.pb") # in attachement

And you will get the error

Thank you and please let me know if you need any other details.

[njzjz]

You didn't attach ge_se_a.yaml, so I use water_se_a.yaml for test. I cannot reproduce the error.

I also use your ge_all_aa.pb and it works well. Your model is good.

I notice your base.yml uses v1.2.2 but dpmd_gpu.yml uses v1.3. Please check it.

[njzjz]

> grep deepmd *.yml
base.yml:  - deepmd-kit=1.2.2=py38_0_cuda10.1_gpu
base.yml:  - libdeepmd=1.2.2=0_cuda10.1_gpu
dpmd_gpu.yml:  - deepmd-kit=1.3.3=py38_1_cuda10.1_gpu
dpmd_gpu.yml:  - libdeepmd=1.3.3=0_cuda10.1_gpu

[marian-code]

Sorry about the yaml file I have somehow forgotten to pack it. Anyway the problem was solved by updating deepmd in base environment as you suggested. Thanks for noticing.

This surprised me. Something must have gone wrong with my conda installation. Apparently when training in dpmd_gpu env the model was trained correctly with deepmd v1.3.3 but when trying to calculate with the ase interface in the same environment it seems that deepmd v.1.2.2 was somehow loaded. I do not understand why.