Polarizability issue

[jakubks]

I'm trying to train a model to predict the polarizability of the system.
This seems to work with the example water scripts and data but crashes when I try to do the same with a smaller system. I get the error below when I try to use it on a system with 31 atoms or fewer. With 32 or more atoms, it works fine.

This seems to me to be a bug unless I'm missing something? (That 32 number seems to be independent of input parameters.)

Traceback (most recent call last):
File "/deepmdkit/bin/dp", line 10, in
sys.exit(main())
File "/deepmdkit/lib/python3.10/site-packages/deepmd/entrypoints/main.py", line 562, in main
train_dp(**dict_args)
File "/deepmdkit/lib/python3.10/site-packages/deepmd/entrypoints/train.py", line 107, in train
_do_work(jdata, run_opt, is_compress)
File "/projects/pr7/js190/deepmdkit/lib/python3.10/site-packages/deepmd/entrypoints/train.py", line 163, in _do_work
model.build(train_data, stop_batch)
File "/deepmdkit/lib/python3.10/site-packages/deepmd/train/trainer.py", line 325, in build
self.model.data_stat(data)
File "/deepmdkit/lib/python3.10/site-packages/deepmd/model/tensor.py", line 76, in data_stat
self._compute_input_stat (m_all_stat, protection = self.data_stat_protect)
File "/deepmdkit/lib/python3.10/site-packages/deepmd/model/tensor.py", line 87, in _compute_input_stat
self.fitting.compute_input_stats(all_stat, protection = protection)
File "//deepmdkit/lib/python3.10/site-packages/deepmd/fit/polar.py", line 269, in compute_input_stats
polar_bias.append(all_stat['polarizability'][ss].reshape((1,9)))
ValueError: cannot reshape array of size 27 into shape (1,9)

[njzjz]

What is the shape of polarizability.npy and coord.npy?

[jakubks]

I have a set of 1333 geometries and a 12 atom system.
np.shape(polarizability) is (1333,9)
np.shape(coord) is (1333,36)

If I add 20 dummy atoms to the coord.npy file, I don't get the error above.

[wanghan-iapcm]

It is a bug fixed by #3426

[wanghan-iapcm]

As a fast fix you may revise the line raising the value error as

[jakubks]

I'm afraid that this (and reinstalling deepmd) gives a different error:

File "/deepmd/lib/python3.10/site-packages/deepmd/fit/polar.py", line 217, in compute_input_stats
np.concatenate(sys_matrix, axis=0),
ValueError: need at least one array to concatenate

Also, a separate question:
I tried running the example (which works since it has 192 atoms). But I'm confused why the resulting model predicts a polarizability of the shape (64,9) for a single geometry of 192 atoms taken from the training set. Shouldn't it predict all atomic tensors [and so for a single geometry output something with a shape (1,192,9) ] ?

[njzjz]

I tried running the example (which works since it has 192 atoms). But I'm confused why the resulting model predicts a polarizability of the shape (64,9) for a single geometry of 192 atoms taken from the training set. Shouldn't it predict all atomic tensors [and so for a single geometry output something with a shape (1,192,9) ] ?

Only the atoms whose type is in sel_types are output.