pyfemtet 1.0.3__tar.gz → 1.0.5__tar.gz

{pyfemtet-1.0.3 → pyfemtet-1.0.5}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.3
 Name: pyfemtet
-Version: 1.0.3
+Version: 1.0.5
 Summary: Design parameter optimization using Femtet.
 License: MIT
 Author: kazuma.naito

{pyfemtet-1.0.3 → pyfemtet-1.0.5}/pyfemtet/_i18n/locales/ja/LC_MESSAGES/messages.po

@@ -976,11 +976,11 @@ msgstr "解析モデル互換性チェックで不明なエラーが発生しま
 
 #: pyfemtet/opt/visualization/plotter/pm_graph_creator.py:146
 msgid "Upper of pred. std-dev"
-msgstr "予測標準偏差の上限"
+msgstr "予測の標準偏差(+)"
 
 #: pyfemtet/opt/visualization/plotter/pm_graph_creator.py:150
 msgid "Lower of pred. std-dev"
-msgstr "予測標準偏差の小さい方"
+msgstr "予測の標準偏差(-)"
 
 #: pyfemtet/opt/visualization/plotter/pm_graph_creator.py:183
 msgid "Std-dev of pred."

{pyfemtet-1.0.3 → pyfemtet-1.0.5}/pyfemtet/opt/history/_history.py

@@ -261,6 +261,14 @@ class DataFrameWrapper:
 class CorrespondingColumnNameRuler:
     """:meta private:"""
 
+    @staticmethod
+    def cns_lower_bound_name(cns_name):
+        return cns_name + '_lower_bound'
+
+    @staticmethod
+    def cns_upper_bound_name(cns_name):
+        return cns_name + '_upper_bound'
+
     @staticmethod
     def direction_name(obj_name):
         return obj_name + '_direction'
@@ -453,7 +461,7 @@ class ColumnManager:
 
                     # later
                     target_cds.update({f(name): object})  # str | float
-                    target_mcs.append(f('obj'))
+                    target_mcs.append('obj.direction')
 
                 for name in extra_y_names:
                     # later
@@ -466,16 +474,35 @@ class ColumnManager:
 
             elif key == 'c':
 
+                f_lb = CorrespondingColumnNameRuler.cns_lower_bound_name
+                f_ub = CorrespondingColumnNameRuler.cns_upper_bound_name
+
                 for name in self.c_names:
                     # important
                     column_dtypes.update({name: float})
                     meta_columns.append('cns')
 
+                    # later
+                    target_cds.update({f_lb(name): float})
+                    target_mcs.append('cns.lower_bound')
+
+                    # later
+                    target_cds.update({f_ub(name): float})
+                    target_mcs.append('cns.upper_bound')
+
                 for name in extra_c_names:
                     # later
                     target_cds.update({name: float})
                     target_mcs.append('')
 
+                    # later
+                    target_cds.update({f_lb(name): float})
+                    target_mcs.append('')
+
+                    # later
+                    target_cds.update({f_ub(name): float})
+                    target_mcs.append('')
+
             # additional_data を入れる
             elif key == self._get_additional_data_column():
                 # important
@@ -682,10 +709,19 @@ class Record:
         # messages to str
         messages_str = _RECORD_MESSAGE_DELIMITER.join(d['messages'])
         d.update({'messages': messages_str})
-        
+
+        # obj
         d.update(**{k: v.value for k, v in y.items()})
-        d.update(**{f'{k}_direction': v.direction for k, v in y.items()})
+        d.update(**{f'{CorrespondingColumnNameRuler.direction_name(k)}': v.direction for k, v in y.items()})
+
+        # cns
         d.update(**{k: v.value for k, v in c.items()})
+        f_lb = CorrespondingColumnNameRuler.cns_lower_bound_name
+        d.update(**{f'{f_lb(k)}': v.lower_bound
+                    for k, v in c.items()})
+        f_ub = CorrespondingColumnNameRuler.cns_upper_bound_name
+        d.update(**{f'{f_ub(k)}': v.upper_bound
+                    for k, v in c.items()})
 
         df = pd.DataFrame(
             {k: [v] for k, v in d.items()},
@@ -1320,7 +1356,7 @@ class History:
                 if row is not None:
                     self_.postprocess_after_recording(row)
 
-                # save history
+                # save history if no FEMOpt
                 client = get_client()
                 if client is None:
                     self.save()
@@ -1332,6 +1368,16 @@ class History:
 
         The destination path is :class:`History.path`.
         """
+
+        # flask server 情報のように、最適化の途中で
+        # 書き換えられるケースがあるので
+        # additional data を再度ここで meta_columns に反映する
+        cm = self._records.column_manager
+        for i, column in enumerate(cm.column_dtypes.keys()):
+            # additional_data を入れる
+            if column == cm._get_additional_data_column():
+                cm.meta_columns[i] = json.dumps(self.additional_data or dict())
+
         self._records.save(self.path)
 
     def _create_optuna_study_for_visualization(self):

{pyfemtet-1.0.3 → pyfemtet-1.0.5}/pyfemtet/opt/interface/__init__.py

@@ -9,6 +9,16 @@ if platform.system() == 'Windows':
     from ._excel_interface import ExcelInterface
     from ._with_excel_settings import *
     from ._with_excel_settings import __all__ as _with_excel_settings__all__
+else:
+    class DummyInterface:
+        def __init__(self, args, kwargs):
+            raise RuntimeError(f'{type(self).__name__} is only for Windows OS.')
+
+    class FemtetInterface(DummyInterface, AbstractFEMInterface): pass
+    class FemtetWithNXInterface(DummyInterface, AbstractFEMInterface): pass
+    class FemtetWithSolidworksInterface(DummyInterface, AbstractFEMInterface): pass
+    class ExcelInterface(DummyInterface, AbstractFEMInterface): pass
+
 
 from ._surrogate_model_interface import AbstractSurrogateModelInterfaceBase
 from ._surrogate_model_interface import BoTorchInterface
@@ -27,4 +37,5 @@ __all__ = [
     'PoFBoTorchInterface',
 ]
 
-__all__.extend(_with_excel_settings__all__)
+if platform.system() == 'Windows':
+    __all__.extend(_with_excel_settings__all__)

{pyfemtet-1.0.3 → pyfemtet-1.0.5}/pyfemtet/opt/optimizer/optuna_optimizer/_pof_botorch/enable_nonlinear_constraint.py

@@ -113,7 +113,7 @@ class _ConvertedConstraint:
 
         """
 
-        x = X.detach().numpy()
+        x = X.detach().cpu().numpy()
         cns_value = _evaluate_pyfemtet_cns(
             self.cns,
             self.opt,
@@ -124,6 +124,8 @@ class _ConvertedConstraint:
             return Tensor([cns_value - self.cns.lower_bound - self.ce])
         elif self.ub_or_lb == 'ub':
             return Tensor([self.cns.upper_bound - cns_value - self.ce])
+        else:
+            assert False
 
 
 # list[pyfemtet.opt.Constraint] について、正規化された入力に対し、 feasible or not を返す関数
@@ -184,7 +186,7 @@ class NonlinearInequalityConstraints:
         for each_num_restarts in ic_batch:
             feasible_q_list = []
             for each_q in each_num_restarts:
-                x: np.ndarray = each_q.detach().numpy()  # normalized parameters
+                x: np.ndarray = each_q.detach().cpu().numpy()  # normalized parameters
                 if _is_feasible(self.constraints, self.opt, self.trans, x, self.ce, self.cs):
                     feasible_q_list.append(each_q)  # Keep only feasible rows
 

{pyfemtet-1.0.3 → pyfemtet-1.0.5}/pyfemtet/opt/optimizer/optuna_optimizer/_pof_botorch/pof_botorch_sampler.py

@@ -841,6 +841,7 @@ class PoFConfig:
         )
     feasibility_cdf_threshold: float | str = 0.5  # or 'sample_mean'
     feasibility_noise: float | str | None = None  # 'no' to fixed minimum noise
+    remove_hard_constraints_from_gp: bool = False  # if consider_explicit_hard_constraint is False, no effect.
 
 
 # TODO:
@@ -1079,6 +1080,8 @@ class PoFBoTorchSampler(BoTorchSampler):
         if n_trials < self._n_startup_trials:
             return {}
 
+        # ===== ここまで変更なし =====
+
         trans = _SearchSpaceTransform(search_space)
         n_objectives = len(study.directions)
         values: numpy.ndarray | torch.Tensor = numpy.empty(
@@ -1099,23 +1102,48 @@ class PoFBoTorchSampler(BoTorchSampler):
                     ):  # BoTorch always assumes maximization.
                         value *= -1
                     values[trial_idx, obj_idx] = value
+
                 if self._constraints_func is not None:
+
+                    # get constraints
                     constraints = study._storage.get_trial_system_attrs(trial._trial_id).get(
                         _CONSTRAINTS_KEY
                     )
+
+                    # Explicit hard constraints を optimize_acqf で扱うならば
+                    # GP にはそれを渡さない (option)
+                    if (
+                            constraints is not None
+                            and self.pof_config.consider_explicit_hard_constraint
+                            and self.pof_config.remove_hard_constraints_from_gp
+                    ):
+                        constraints = filter_soft_constraints_only(
+                            constraints, self.pyfemtet_optimizer
+                        )
+
                     if constraints is not None:
+
                         n_constraints = len(constraints)
 
-                        if con is None:
-                            con = numpy.full(
-                                (n_completed_trials, n_constraints), numpy.nan, dtype=numpy.float64
-                            )
-                        elif n_constraints != con.shape[1]:
-                            raise RuntimeError(
-                                f"Expected {con.shape[1]} constraints "
-                                f"but received {n_constraints}."
-                            )
-                        con[trial_idx] = constraints
+                        # remove_hard_constraints_from_gp で
+                        # constraints がなくなった場合、
+                        # そもそも今後ここに来る必要がない
+                        if n_constraints == 0:
+                            self._constraints_func = None
+
+                        else:
+
+                            if con is None:
+                                con = numpy.full(
+                                    (n_completed_trials, n_constraints), numpy.nan, dtype=numpy.float64
+                                )
+                            elif n_constraints != con.shape[1]:
+                                raise RuntimeError(
+                                    f"Expected {con.shape[1]} constraints "
+                                    f"but received {n_constraints}."
+                                )
+                            con[trial_idx] = constraints
+
             elif trial.state == TrialState.RUNNING:
                 if all(p in trial.params for p in search_space):
                     params[trial_idx] = trans.transform(trial.params)
@@ -1162,7 +1190,6 @@ class PoFBoTorchSampler(BoTorchSampler):
         else:
             running_params = None
 
-        # ===== ここまで変更なし =====
 
         # TODO: ミーゼスなどの場合にこれらのシード固定法も試す
         # if self._seed is not None:
@@ -1247,3 +1274,28 @@ class PoFBoTorchSampler(BoTorchSampler):
         # ===== ここまで変更なし =====
 
         return trans.untransform(candidates.cpu().numpy())
+
+
+def filter_soft_constraints_only(
+        constraints: tuple[float],
+        opt: AbstractOptimizer,
+) -> list[float]:
+    # constraints を取得
+    # lb, ub の存在に応じて hard, soft を展開
+    is_soft_list = []
+    for cns in opt.constraints.values():
+        if cns.lower_bound is not None:
+            is_soft_list.append(not cns.hard)
+        if cns.upper_bound is not None:
+            is_soft_list.append(not cns.hard)
+
+    # constraints と比べる
+    assert len(constraints) == len(is_soft_list)
+
+    # soft に該当する要素のみで組立て返す
+    ret = []
+    for is_soft, value in zip(is_soft_list, constraints):
+        if is_soft:
+            ret.append(value)
+
+    return ret

{pyfemtet-1.0.3 → pyfemtet-1.0.5}/pyfemtet/opt/prediction/_botorch_utils.py

@@ -3,15 +3,17 @@ from __future__ import annotations
 
 from packaging import version
 
+from tqdm import tqdm
 import torch
 
 from gpytorch.mlls import ExactMarginalLogLikelihood
 from gpytorch.kernels import MaternKernel, ScaleKernel  # , RBFKernel
 from gpytorch.priors.torch_priors import GammaPrior  # , LogNormalPrior
-# from gpytorch.constraints.constraints import GreaterThan
+from gpytorch.constraints.constraints import GreaterThan
 
 from botorch.models import SingleTaskGP
 from botorch.models.transforms import Standardize, Normalize
+from botorch.exceptions import ModelFittingError
 
 # import fit_gpytorch_mll
 import botorch.version
@@ -22,6 +24,9 @@ if version.parse(botorch.version.version) < version.parse("0.8.0"):
 else:
     from botorch.fit import fit_gpytorch_mll
 
+from pyfemtet.logger import get_module_logger
+
+logger = get_module_logger('opt.botorch_util')
 
 __all__ = [
     'get_standardizer_and_no_noise_train_yvar',
@@ -128,6 +133,58 @@ def setup_gp(X, Y, bounds, observation_noise, lh_class=None, covar_module=None):
     )
 
     mll_ = lh_class(model_.likelihood, model_)
-    fit_gpytorch_mll(mll_)
+
+    try:
+        fit_gpytorch_mll(mll_)
+
+    except ModelFittingError as e:
+        logger.warning(f'{type(e).__name__} is raised '
+                       f'during fit_gpytorch_mll()! '
+                       f'The original message is: `{",".join(e.args)}`.')
+
+        # retry with noise level setting
+        try:
+            logger.info('Attempt to retrying...')
+
+            # noinspection PyBroadException
+            try:
+                model_.likelihood.noise_covar.register_constraint(
+                    "raw_noise", GreaterThan(1e-5))
+                logger.warning('Set the raw_noise constraint to 1e-5.')
+            except Exception:
+                logger.info('Failed to set the raw_noise constraint. '
+                            'Retrying simply.')
+
+            fit_gpytorch_mll(mll_)
+
+        except ModelFittingError:
+            logger.warning(f'{type(e).__name__} is raised '
+                           f'during *second* fit_gpytorch_mll()! '
+                           f'The original message is: `{",".join(e.args)}`.')
+
+            # retry with another way
+            logger.warning('Attempt to retrying.')
+            logger.warning('Try to use SGD algorithm...')
+
+            from torch.optim import SGD
+
+            NUM_EPOCHS = 150
+
+            optimizer = SGD([{"params": model_.parameters()}], lr=0.025)
+
+            model_.train()
+
+            for epoch in tqdm(range(NUM_EPOCHS), desc='fit with SGD'):
+                # clear gradients
+                optimizer.zero_grad()
+                # forward pass through the model to obtain the output MultivariateNormal
+                output = model_(X)
+                # Compute negative marginal log likelihood
+                loss = -mll_(output, model_.train_targets)
+                # back prop gradients
+                loss.backward()
+                optimizer.step()
+
+            model_.eval()
 
     return model_

{pyfemtet-1.0.3 → pyfemtet-1.0.5}/pyfemtet/opt/prediction/_model.py

@@ -73,8 +73,8 @@ class SingleTaskGPModel(AbstractModel):
         X = torch.tensor(x, **self.KWARGS)
         post = self.gp.posterior(X)
         with torch.no_grad():
-            mean = post.mean.numpy()
-            std = post.variance.sqrt().numpy()
+            mean = post.mean.cpu().numpy()
+            std = post.variance.sqrt().cpu().numpy()
         return mean, std
 
 

{pyfemtet-1.0.3 → pyfemtet-1.0.5}/pyproject.toml

@@ -26,14 +26,14 @@ dependencies = [
     "optuna>=3.4.0, <5.0.0",
     "optuna-integration>=3.6.0, <5.0.0",
     "botorch>=0.12.0, <0.15.0",
-    "torch",  # botorch specifies torch's version
-    "fire>=0.7",  # It probably has a strong backward compatibility.
+    "torch", # botorch specifies torch's version
+    "fire>=0.7", # It probably has a strong backward compatibility.
     "pyyaml>=6.0.2,<7",
 ]
 
 #[project.optional-dependencies]
 #brep = ["brepmatching>=0.1.8,<0.2"]
-version = "1.0.3"
+version = "1.0.5"
 
 [project.urls]
 Repository = "https://github.com/pyfemtet/pyfemtet"
@@ -58,15 +58,6 @@ dev = [
     "pytest-reporter-html-dots>=0.11.0",
 ]
 
-[[tool.uv.index]]
-name = "pytorch-cpu"
-url = "https://download.pytorch.org/whl/cpu"
-explicit = true
-
-[tool.uv.sources]
-# Defaults to installing CUDA without index-url specification only on linux
-torch = [{ index = "pytorch-cpu", marker = "sys_platform=='linux'" }]
-
 [tool.uv-dynamic-versioning]
 fallback-version = "0.0.0"