Deepmd training suggestion #6021

Activity

1. 1997sankha

   added

   wontfix

   on Sep 13, 2026
2. njzjz-bot commented on Oct 2, 2026

   

   njzjz-bot

   on Oct 2, 2026

   Contributor

   More actions

   This appears to be the same training/evaluation follow-up as #5950; the explanation there about separate energy plots also applies here. Different atom counts/compositions can give different total-energy ranges. Separate clusters alone do not establish a training failure, and combining Cartesian force components into one plot is fine.

   Before changing the descriptor or network, please evaluate each held-out system separately:

   dp --tf test -m graph.pb -s /path/to/held_out_system -n 0 -d heldout

   Replace the model and system paths with yours. In your reported v3.1.1, -n 0 removes the frame-count cap on the selected test data. Use a directory containing the intended validation data so this does not accidentally evaluate a training split.

   For each system, report the energy-per-atom and force RMSE, and plot the paired reference/prediction columns against y = x. Energy residuals (E_pred - E_ref)/N are more informative than whether different configurations share the same absolute energy range. Also verify identical atom ordering, type mapping, units, and frame alignment between coordinates and labels.

   If validation_data.numb_btch is still 1 as in the earlier configuration, increasing it (e.g. to 100) can reduce noise in training-time monitoring; it does not change model capacity or replace the held-out evaluation. If both training and validation errors remain large, a useful next diagnostic is fitting roughly 100 unique frames from one system and evaluating those same frames. Keep the validation split independent, and perform any bootstrapping only after splitting.

   I could not retrieve the attached JSON/learning-curve files from this review environment, so I cannot justify a specific learning-rate, cutoff, or loss-weight change from those attachments. Please put the per-system RMSE values and the relevant input sections in text, preferably continuing #5950 so the diagnosis stays in one thread.

   ---

   Coding agent: Codex
   Codex version: codex-cli 0.160.0
   Model: gpt-6-astra
   Reasoning effort: xhigh
3. 1997sankha commented on Oct 3, 2026

   

   1997sankha

   on Oct 3, 2026

   Last edited by 1997sankha

   Author

   More actions

   input2.json
   is this input file ok? or i have to check with other descriptors also?