Compress a model: zero-size array to reduction operation minimum which has no identity

[superstar54]

Dear all,

I am using Deepmd-2.0.0-beta-v3. I got the following error when compressing a model:

Traceback (most recent call last):
File "/storage/homefs/xw20n572/softwares/Anaconda3/deepmd-kit-2.0.0-gpu/bin/dp", line 10, in
sys.exit(main())
File "/storage/homefs/xw20n572/softwares/Anaconda3/deepmd-kit-2.0.0-gpu/lib/python3.9/site-packages/deepmd/entrypoints/main.py", line 429, in main
compress(**dict_args)
File "/storage/homefs/xw20n572/softwares/Anaconda3/deepmd-kit-2.0.0-gpu/lib/python3.9/site-packages/deepmd/entrypoints/compress.py", line 97, in compress
train(
File "/storage/homefs/xw20n572/softwares/Anaconda3/deepmd-kit-2.0.0-gpu/lib/python3.9/site-packages/deepmd/entrypoints/train.py", line 212, in train
_do_work(jdata, run_opt)
File "/storage/homefs/xw20n572/softwares/Anaconda3/deepmd-kit-2.0.0-gpu/lib/python3.9/site-packages/deepmd/entrypoints/train.py", line 262, in _do_work
model.build(train_data, stop_batch)
File "/storage/homefs/xw20n572/softwares/Anaconda3/deepmd-kit-2.0.0-gpu/lib/python3.9/site-packages/deepmd/train/trainer.py", line 306, in build
= self.neighbor_stat.get_stat(data)
File "/storage/homefs/xw20n572/softwares/Anaconda3/deepmd-kit-2.0.0-gpu/lib/python3.9/site-packages/deepmd/utils/neighbor_stat.py", line 85, in get_stat
dt = np.min(dt)
File "<array_function internals>", line 5, in amin
File "/storage/homefs/xw20n572/softwares/Anaconda3/deepmd-kit-2.0.0-gpu/lib/python3.9/site-packages/numpy/core/fromnumeric.py", line 2858, in amin
return _wrapreduction(a, np.minimum, 'min', axis, None, out,
File "/storage/homefs/xw20n572/softwares/Anaconda3/deepmd-kit-2.0.0-gpu/lib/python3.9/site-packages/numpy/core/fromnumeric.py", line 87, in _wrapreduction
return ufunc.reduce(obj, axis, dtype, out, **passkwargs)
ValueError: zero-size array to reduction operation minimum which has no identity

Could anyone help me with how to avoid this?

Best,
Xing

[njzjz]

Duplicate of #795. This should be a bug and waits to be fixed.

[nicklin96]

can you share your input data? I am trying to reproduce the error.

[superstar54]

Here are the input.json and graph.pb files. The command I used is:

dp compress input.json -i graph.pb -o graph-compress.pb

Thanks very much for your help!

compress-test.zip

[nicklin96]

It seems that the error is related to the training data. An empty array of minimum distance between neighbors was return.
Do you mind sharing your training data?

[superstar54]

Yes, you are right. In my training data, I have single atom data sets for each element. This is an extreme case, which has no neighboring atom. This data set is very helpful for the training, because most of the energy (> 99%) in a chemical system is the sum of the energy of the single atom, while the energy of chemical bonds is only attributed to a small portion.

I deleted the single atom data sets in the input.json file. And it running normally. I checked the simulation results. The compressed model shows almost the same results as the normal model. Do you think is it OK that the data sets used for compressing the model are different from the previous training process?

Here is a example for single atom data set:
Ag.zip

[nicklin96]

I think a slight difference is fine.
Besides, model compression only affects the time-consuming embedding procedure, while the knowledge in descriptor net remains.
but I think a fix to deal with such a special case should be made.
Thanks for your valuable feedback!

[superstar54]

Thanks very much for your help.

[njzjz]

Fixed in #882.