loc_frame support in v2.0.3

[njzjz]

Discussed in #1244

^{Originally posted by zrzrv5 October 26, 2021}
Hi,

I just updated deepmd-kit from v1.3.3 to v2.03 on my local offline workstation, and I can't train a model using loc_frame descriptor. When I train the model using loc_frame, the code in entrypoints/train.py checks jdata['sel'], but loc_frame only has sel_a and sel_r.
So I made the following changes on my machine:

1. skip update_sel(jdata) in entrypoints/train.py line 90
2. add @Descriptor.register('loc_frame') in descriptor/loc_frame.py line8

Now I can start the training, but lammps crashes when using this model. I also tried converting a model from 1.3.3, also doesn't work in lammps.

Is loc_frame not available in this version? Or do I need to change other codes?

Also, is there any way to check the lammps and deeppot error log?

Thanks,

[njzjz]

The second error message (Segmentation fault) is 0x00007fffb8161ac5 in DescrptOp<Eigen::ThreadPoolDevice, double>::_Compute(tensorflow::OpKernelContext*) ()

0x00007fffb80fc510 in format_nlist_i_fill_a(std::vector<int, std::allocator<int> >&, std::vector<int, std::allocator<int> >&, std::vector<double, std::allocator<double> > const&, int const&, std::vector<int, std::allocator<int> > const&, SimulationRegion<double> const&, bool const&, int const&, std::vector<int, std::allocator<int> > const&, std::vector<int, std::allocator<int> > const&, double const&, std::vector<int, std::allocator<int> > const&, std::vector<int, std::allocator<int> > const&) () from /home/jz748/anaconda3/envs/dpdev113/bin/../lib/./././libdeepmd.so

located at

deepmd-kit/source/op/descrpt.cc

Line 272 in 1a25414

if ((ret = format_nlist_i_fill_a (fmt_nlist_a, fmt_nlist_r, d_coord3, ntypes, d_type, region, b_pbc, ii, d_nlist_a[ii], d_nlist_r[ii], rcut_r, sec_a, sec_r)) != -1){