PyFemtet 1.8.1__tar.gz → 1.8.2__tar.gz

{pyfemtet-1.8.1 → pyfemtet-1.8.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: pyfemtet
-Version: 1.8.1
+Version: 1.8.2
 Summary: Design parameter optimization using Femtet.
 License-Expression: MIT
 License-File: LICENSE

{pyfemtet-1.8.1 → pyfemtet-1.8.2}/pyfemtet/opt/prediction/_botorch_utils.py

@@ -5,6 +5,7 @@ from packaging import version
 
 from tqdm import tqdm
 import torch
+import gpytorch
 
 from gpytorch.mlls import ExactMarginalLogLikelihood
 from gpytorch.kernels import MaternKernel, ScaleKernel  # , RBFKernel
@@ -37,8 +38,6 @@ __all__ = [
 
 
 def get_standardizer_and_no_noise_train_yvar(Y: torch.Tensor):
-    import gpytorch
-
     standardizer = Standardize(m=Y.shape[-1])
     min_noise = gpytorch.settings.min_fixed_noise.value(Y.dtype)
     standardizer.forward(Y)  # require to un-transform
@@ -50,19 +49,28 @@ def get_standardizer_and_no_noise_train_yvar(Y: torch.Tensor):
 def setup_yvar_and_standardizer(
         Y_: torch.Tensor,
         observation_noise_: str | float | None,
-) -> tuple[torch.Tensor | None, Standardize]:
-
+        use_outcome_transform: bool = True,
+) -> tuple[torch.Tensor | None, Standardize | None]:
     standardizer_ = None
     train_yvar_ = None
+
     if isinstance(observation_noise_, str):
         if observation_noise_.lower() == 'no':
-            train_yvar_, standardizer_ = get_standardizer_and_no_noise_train_yvar(Y_)
+            if use_outcome_transform:
+                # Standardize 経由で untransform して YVar を得る
+                train_yvar_, standardizer_ = get_standardizer_and_no_noise_train_yvar(Y_)
+            else:
+                # 外部で標準化済みの場合、min_noise をそのまま使用
+                min_noise = gpytorch.settings.min_fixed_noise.value(Y_.dtype)
+                train_yvar_ = torch.full_like(Y_, min_noise)
         else:
             raise NotImplementedError
     elif isinstance(observation_noise_, float):
+        # 正規化済みスケールの observation_noise を期待
         train_yvar_ = torch.full_like(Y_, observation_noise_)
 
-    standardizer_ = standardizer_ or Standardize(m=Y_.shape[-1])
+    if use_outcome_transform:
+        standardizer_ = standardizer_ or Standardize(m=Y_.shape[-1])
 
     return train_yvar_, standardizer_
 
@@ -115,12 +123,12 @@ def get_matern_kernel_with_gamma_prior_as_covar_module(
     )
 
 
-def setup_gp(X, Y, bounds, observation_noise, lh_class=None, covar_module=None):
+def setup_gp(X, Y, bounds, observation_noise, lh_class=None, covar_module=None, use_outcome_transform=True):
 
     lh_class = lh_class or ExactMarginalLogLikelihood
 
     train_yvar_, standardizer_ = setup_yvar_and_standardizer(
-        Y, observation_noise
+        Y, observation_noise, use_outcome_transform
     )
 
     model_ = SingleTaskGP(

{pyfemtet-1.8.1 → pyfemtet-1.8.2}/pyfemtet/opt/prediction/_model.py

@@ -28,6 +28,8 @@ class SingleTaskGPModel(AbstractModel):
 
     KWARGS = dict(dtype=torch.float64, device='cpu')
     gp: SingleTaskGP
+    y_mean: torch.Tensor  # Y の平均を保存（復元用）
+    y_std: torch.Tensor   # Y の標準偏差を保存（復元用）
 
     def fit(
             self,
@@ -45,6 +47,12 @@ class SingleTaskGPModel(AbstractModel):
         Y = torch.tensor(y, **self.KWARGS)
         B = torch.tensor(bounds, **self.KWARGS).transpose(1, 0) if bounds is not None else None
 
+        # Y を外部で標準化（BoTorch 内部の Standardize を使わない）
+        self.y_mean = Y.mean(dim=0, keepdim=True)
+        self.y_std = Y.std(dim=0, keepdim=True)
+        self.y_std[self.y_std == 0] = 1.0  # ゼロ除算防止
+        Y_normalized = (Y - self.y_mean) / self.y_std
+
         if covar_module_settings is not None:
             if covar_module_settings['name'] == 'matern_kernel_with_gamma_prior':
                 covar_module_settings.pop('name')
@@ -66,7 +74,8 @@ class SingleTaskGPModel(AbstractModel):
             else:
                 raise NotImplementedError(f'{covar_module_settings["name"]=}')
 
-        self.gp = setup_gp(X, Y, B, observation_noise, likelihood_class, covar_module)
+        # 標準化済み Y を渡し、outcome_transform を使用しない
+        self.gp = setup_gp(X, Y_normalized, B, observation_noise, likelihood_class, covar_module, use_outcome_transform=False)
 
     def predict(self, x: np.ndarray):
         assert hasattr(self, 'gp')
@@ -75,7 +84,14 @@ class SingleTaskGPModel(AbstractModel):
         with torch.no_grad():
             mean = post.mean.cpu().numpy()
             std = post.variance.sqrt().cpu().numpy()
-        return mean, std
+
+        # 元のスケールに復元
+        y_mean_np = self.y_mean.cpu().numpy()
+        y_std_np = self.y_std.cpu().numpy()
+        mean_restored = mean * y_std_np + y_mean_np
+        std_restored = std * y_std_np
+
+        return mean_restored, std_restored
 
 
 class PyFemtetModel:

{pyfemtet-1.8.1 → pyfemtet-1.8.2}/pyproject.toml

@@ -34,7 +34,7 @@ dependencies = [
 
 #[project.optional-dependencies]
 #brep = ["brepmatching>=0.1.8,<0.2"]
-version = "1.8.1"
+version = "1.8.2"
 
 [project.urls]
 Repository = "https://github.com/pyfemtet/pyfemtet"