Default use_gpumd_qscaler to False

README.md

@@ -142,7 +142,7 @@ function (`train_nep` / `train_nep_sharded`):
 | `recompute_q_scaler` | `False` | only with `finetune_from`: recompute the descriptor scaler on the new data instead of keeping the source model's |
 | `slim_types` | `False` | drop element types absent from the dataset |
 | `energy_key` | `"energy"` | comment-line tag read as reference energy (e.g. `"atomization_energy"`) |
-| `use_gpumd_qscaler` | `True` | reproduce GPUMD's init (SNES `mu`): re-init every parameter — descriptor coeffs **and** NN weights — uniform(−1,1), and compute `q_scaler` with coeffs `c=1`. Fresh training only |
+| `use_gpumd_qscaler` | `False` | `False`: torch's default init + self-consistent `q_scaler` — converges to better minima (600-epoch 4-seed benchmark: ~12% lower E/V RMSE, ~3% lower F). `True`: reproduce GPUMD's init (SNES `mu`, all parameters uniform(−1,1)) with the `c=1` `q_scaler` — for GPUMD-comparison runs. The saved `nep.txt` is GPUMD-compatible either way. Fresh training only |
 | `run_seed` | `None` | master RNG seed. `None` = random each run; an int makes the run reproducible (weight init + batch shuffle). Saved in `checkpoint.pt`, restored on resume |
 | `valid_file` | `None` | validation `.xyz`, `nep_best` and the plateau LR schedule follow the validation loss; writes GPUMD-style `*_test.out` |
 | `valid_ratio` | `None` | hold out this fraction (e.g. `0.1`) of `data_file` as the validation set; the split is drawn from `run_seed` and preserved on resume. Mutually exclusive with `valid_file` |

releaseNotes.md

@@ -2,6 +2,12 @@
 
 ## Unreleased
 
+- **`use_gpumd_qscaler` now defaults to `False`**: torch's default init
+  with the self-consistent q_scaler converges to clearly better minima
+  than the GPUMD-style start (600-epoch 4-seed PdCuNiP benchmark: ~12%
+  lower E/V RMSE, ~3% lower F, on train and validation alike). `True`
+  (the old default) remains available for GPUMD-comparison runs; the
+  saved nep.txt is GPUMD-compatible either way.
 - **`export_valid_split`**: write the exact `valid_ratio` split
   `train_nep` uses as verbatim GPUMD-ready `train.xyz` / `test.xyz`
   files, so the same data partition can be trained in GPUMD and the loss

torchnep/train.py

@@ -1126,7 +1126,7 @@ def train_nep(
     recompute_q_scaler: bool = False,
     slim_types: bool = False,
     energy_key: str = "energy",
-    use_gpumd_qscaler: bool = True,
+    use_gpumd_qscaler: bool = False,
     run_seed: int = None,
     valid_file: str = None,
     valid_ratio: float = None,
@@ -1180,14 +1180,18 @@ def train_nep(
     energy_key : name of the comment-line tag read as the reference energy
         (default ``"energy"``). Set to ``"atomization_energy"`` to train
         against atomization energies instead of totals.
-    use_gpumd_qscaler : Default True — reproduce GPUMD's initialization: every
-        parameter is re-initialised uniform(-1, 1) — the descriptor
-        coefficients AND the NN weights (w0/b0/w1), matching SNES's mu init —
-        and the q_scaler is computed with all coefficients = 1.0 (GPUMD's
-        generation-0 ``initial_para``). False leaves torch's default NN init in
-        place and uses the self-consistent q_scaler (computed from the model's
-        actual init coefficients). Only applies to fresh training (ignored
-        under finetune_from).
+    use_gpumd_qscaler : Default False — torch's default init with the
+        self-consistent q_scaler (computed from the model's actual init
+        coefficients). On a 600-epoch 4-seed PdCuNiP benchmark this
+        converges to clearly better minima than the GPUMD-style start
+        (~12% lower E/V RMSE, ~3% lower F, train and validation alike).
+        True reproduces GPUMD's initialization instead: every parameter
+        re-initialised uniform(-1, 1) (SNES mu init) and the q_scaler
+        computed with all coefficients = 1.0 (GPUMD's generation-0
+        ``initial_para``) — useful for GPUMD-comparison runs. Either way
+        the saved nep.txt is fully GPUMD-compatible (the scaler is stored
+        in the file). Only applies to fresh training (ignored under
+        finetune_from).
     run_seed : master RNG seed for this run. None (default) -> a fresh random
         seed each run, so repeated runs differ (independent weight init AND
         per-epoch batch shuffle) — the stochastic-testing behaviour. Pass an

torchnep/train_sharded.py

@@ -215,7 +215,7 @@ def train_nep_sharded(
     recompute_q_scaler: bool = False,
     slim_types: bool = False,
     energy_key: str = "energy",
-    use_gpumd_qscaler: bool = True,
+    use_gpumd_qscaler: bool = False,
     run_seed: int = None,
     valid_file: str = None,
     valid_ratio: float = None,