bfflearn 0.4.0__tar.gz → 0.4.2__tar.gz

bfflearn-0.4.2/CHANGELOG.md

@@ -0,0 +1,32 @@
+## `0.4.2` - 2026-10-02
+
+### Changed
+
+- Bounded the posterior sample counts used for default corner, standard
+  marginal, and QoI marginal plots. The standard marginal cap accepts `-1` or
+  `null` to use every sample. QoI likelihood attribution now runs in adaptive
+  batches to avoid CUDA memory exhaustion after learning.
+- Centered default LGP hyperparameter priors on the observed input and residual
+  target scales, avoiding noise-dominated fits for scalar QoIs whose natural
+  parameter or output scales differ substantially from one.
+- Reported the mode of each posterior marginal instead of its mean, limited
+  charge modes to three decimal places, and formatted other parameter modes
+  with three significant digits.
+- Balanced marginal plots across rows of at most five parameters, placed each
+  arbitrary `define` parameter on an independent axis, and added configurable
+  axis labels through `plots.plot_metadata`. Marginal figures now reserve
+  measured space for legends so they cannot overlap single- or multi-panel
+  plots, align y-axis labels within subplot columns, and leave sufficient
+  gutters between neighboring panel sections.
+- Made local validation progress visible immediately, including campaigns with
+  a single sample, and advanced the displayed count only after each MD job has
+  completed.
+
+## `0.4.1` - 2026-08-25
+
+### Changed
+
+- Simplified RDF calculation so selection handling and atom-type expansion
+  occur in the QoI adapter while the numerical kernel operates directly on
+  MDAnalysis AtomGroups. Dynamic selections and numerical behavior are
+  preserved.

{bfflearn-0.4.0 → bfflearn-0.4.2}/CITATION.cff

@@ -1,8 +1,8 @@
 cff-version: 1.2.0
 message: "If you use BFF, please cite the repository and the associated paper."
 title: "Bayesic Force Fields"
-version: "0.4.0"
-date-released: "2026-08-24"
+version: "0.4.2"
+date-released: "2026-10-02"
 authors:
   - family-names: Kostal
     given-names: Vojtech

{bfflearn-0.4.0/bfflearn.egg-info → bfflearn-0.4.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: bfflearn
-Version: 0.4.0
+Version: 0.4.2
 Summary: Bayesian workflows for fixed-charge molecular force fields
 Author-email: Vojtech Kostal <vojtakostal8@gmail.com>
 License-Expression: GPL-3.0-only

{bfflearn-0.4.0 → bfflearn-0.4.2}/bff/__init__.py

@@ -6,7 +6,7 @@ from importlib import import_module
 from pathlib import Path
 from typing import Any
 
-__version__ = "0.4.0"
+__version__ = "0.4.2"
 
 _INTERNAL_MODULES = {
     "bayes",

{bfflearn-0.4.0 → bfflearn-0.4.2}/bff/bayes/learning.py

@@ -369,6 +369,42 @@ def _default_checkpoint_path(fn_posterior: Path) -> Path:
     return fn_posterior.with_name(f"{stem}.ckpt{suffix}")
 
 
+def _default_lgp_hyperpriors(
+    X: torch.Tensor,
+    residuals: torch.Tensor,
+) -> Priors:
+    """Build scale-aware priors for log GP hyperparameters."""
+    input_scales = X.std(dim=0, unbiased=False)
+    input_scales = torch.where(
+        torch.isfinite(input_scales) & (input_scales > 0),
+        input_scales,
+        torch.ones_like(input_scales),
+    )
+
+    target_scale = residuals.std(unbiased=False)
+    if not torch.isfinite(target_scale) or target_scale <= 0:
+        target_scale = torch.sqrt(torch.mean(residuals.square()))
+    if not torch.isfinite(target_scale) or target_scale <= 0:
+        target_scale = torch.ones((), dtype=residuals.dtype)
+
+    # LocalGaussianProcess adds ``sigma`` directly to the covariance diagonal,
+    # so its natural scale is a fraction of the target variance.
+    noise_scale = 0.1 * target_scale.square()
+    tiny = torch.finfo(noise_scale.dtype).tiny
+    noise_scale = torch.clamp(noise_scale, min=tiny)
+
+    return Priors(
+        [
+            Prior("normal", float(torch.log(scale)), 2.0, name=f"length_{i}")
+            for i, scale in enumerate(input_scales)
+        ]
+        + [
+            Prior("normal", float(torch.log(target_scale)), 2.0, name="width"),
+            Prior("normal", float(torch.log(noise_scale)), 3.0, name="noise"),
+        ]
+    )
+
+
 def fit_lgp_committee(
     X: torch.Tensor,
     y: torch.Tensor,
@@ -400,16 +436,7 @@ def fit_lgp_committee(
 
     n_params = X.shape[1]
     if hyperpriors is None:
-        priors = Priors(
-            [
-                Prior("normal", -2.0, 2.0, name=f"length_{i}")
-                for i in range(n_params)
-            ]
-            + [
-                Prior("normal", -2.0, 2.0, name="width"),
-                Prior("normal", -2.0, 3.0, name="noise"),
-            ]
-        )
+        priors = _default_lgp_hyperpriors(X_hyper, y_hyper - y_hyper_mean)
     else:
         priors = Priors.from_any(hyperpriors)
     if len(priors) != n_params + 2:

{bfflearn-0.4.0 → bfflearn-0.4.2}/bff/bayes/utils.py

@@ -215,6 +215,11 @@ def find_max_stable_lr(
     else:
         learning_rates = list(learning_rates)
     lower, upper = param_bounds
+    # Log-scale hyperparameters can legitimately start outside the generic
+    # stability window when the underlying data are very small or very large.
+    # Preserve the explosion guard while always admitting the supplied start.
+    lower = min(lower, float(p0.min()) - 1.0)
+    upper = max(upper, float(p0.max()) + 1.0)
     for lr in learning_rates:
         x = p0.clone().detach().requires_grad_(True)
         opt = torch.optim.SGD([x], lr=lr)

{bfflearn-0.4.0 → bfflearn-0.4.2}/bff/io/logs.py

@@ -216,6 +216,7 @@ class Logger:
         level: int = 1,
         overwrite: bool = False,
         style: str | tuple[str, ...] | None = None,
+        write_file: bool = True,
     ) -> None:
         """Write a pytest-style line with a status block right aligned."""
         prefix = self._prefix(level)
@@ -223,7 +224,8 @@ class Logger:
         plain_len = len(prefix) + len(message) + len(status)
         spacing = " " * max(width - plain_len, 1)
         line = f"{prefix}{message}{spacing}{status}"
-        self._write_file(line)
+        if write_file:
+            self._write_file(line)
         self._write_console(line, overwrite=overwrite, style=style)
 
     def progress_status(
@@ -234,6 +236,7 @@ class Logger:
         *,
         level: int = 1,
         overwrite: bool = False,
+        write_file: bool = True,
     ) -> None:
         """Write a pytest-style progress line with ``[ 42%]`` on the right."""
         percent = 100 if total <= 0 else round(100 * current / total)
@@ -244,6 +247,7 @@ class Logger:
             level=level,
             overwrite=overwrite,
             style="magenta",
+            write_file=write_file,
         )
 
     def result_summary(