The performance of the DPA2 model is an order of magnitude worse than that of DPA1

[duduRUHappy]

When I used DPA1 for testing on a new dataset, the results were as follows: rmse_energy_valid = 0.0361 eV and rmse_force_valid = 0.29 eV/Å. These results are in line with expectations. However, when I used the single task model version of DPA2 for training and testing, the results were rmse_energy_valid = 0.1361 eV and rmse_force_valid = 1.8297 eV/Å. As the results indicate, the performance of the DPA2 model is an order of magnitude worse than that of DPA1. Could you help analyze how to improve the DPA2 results?

I adopted an input.json similar to the one published for the DPA2 model. The details of input.json of DPA1 and DPA2 are included in the attachment, please take a look.
dpa1_ener_input.json
dpa2_st_ener_input.json

[wanghan-iapcm]

from the input script it seems that you are using a very old version of the code. please report the version, or you may want to try the latest devel branch or the 2024Q1 stable branch.

[duduRUHappy]

Yes, I used the 'devel' branch from git@github.com:dptech-corp/deepmd-pytorch.git.

[duduRUHappy]

[wanghan-iapcm]

the deepmd-pytorch code is no longer supported

[duduRUHappy]

Ok, I got it. Then if I wanna test DPA2, should I use v3.0.0a0 or the latest v2.2.10?

[njzjz]

It's not in v2. Perhaps a deprecation notice should be added to deepmd-pytorch. cc @iProzd