finetune dpa3 #4830
Unanswered
chtchelkatchev
asked this question in
Q&A
finetune dpa3
#4830
Replies: 1 comment 2 replies
-
|
@chtchelkatchev Hi, there might be two primary factors causing the confusion:
First run: Second run following exactly the same process: |
Beta Was this translation helpful? Give feedback.
2 replies
-
|
Thank you for the quick response. Given the observed differences when running finetuning on a small database, I suspect that finetuning affects not only the outer layers of the fitting net but also the graph descriptor. This is supported by the out.json file, where the descriptor is marked as "trainable": true. Could you briefly clarify which parts of the DPA-3 potential are actually trained during finetuning and which remain frozen? Is there any way to manually control what gets frozen? On a related note, the developers of NEQUIP claim that even 100 configurations can sometimes be sufficient to train NEQUIP from scratch, as demonstrated in their examples. However, it appears that with the same number of configurations, deepmd-dpa3 cannot even be finetuned effectively. |
Beta Was this translation helpful? Give feedback.
-
|
By default, it's a full finetuning. However, you can freeze the descriptor if you want, see doc. |
Beta Was this translation helpful? Give feedback.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
-
I started using the finetune DPA3, as recommended in https://www.bohrium.com/notebooks/57111746135, for my tasks using the latest deepmd-3.1.0. The fine-tuning appeared to go well externally, and lcurve.out showed high convergence accuracy for all parameters. However, the dp test check revealed that, for example, the energy per atom had an accuracy much worse (by more than an order of magnitude) than what was shown in lcurve.out.
To ensure that this was not my mistake but rather a bug in DeepMD, I tried reproducing the finetune water example suggested in https://www.bohrium.com/notebooks/57111746135. Here, the situation was similar. The lcurve.out matched the one in https://www.bohrium.com/notebooks/57111746135, but dptest showed different results. The dptest values from https://www.bohrium.com/notebooks/57111746135 could not be reproduced:
lcurve.out
step rmse_val
If there is no available reference data, rmse_*_{val,trn} will print nan
my dp test
[2025-07-01 10:22:08,347] DEEPMD INFO Energy MAE/Natoms : 1.124378e-03 eV
[2025-07-01 10:22:08,347] DEEPMD INFO Energy RMSE/Natoms : 1.138879e-03 eV
https://www.bohrium.com/notebooks/57111746135 results
[2025-06-10 15:32:49,768] DEEPMD INFO Energy MAE/Natoms : 7.430266e-04 eV
[2025-06-10 15:32:49,768] DEEPMD INFO Energy RMSE/Natoms : 7.646335e-04 eV
input_finetune.json
dp --pt train input_finetune.json --skip-neighbor-stat --finetune DPA-3.1-3M.pt --model-branch MP_traj_v024_alldata_mixu < /dev/null && break
Beta Was this translation helpful? Give feedback.
All reactions