def cv(self, dataset: tk.data.Dataset,
folds: tk.validation.FoldsType) -> Model:
"""CVして保存。
Args:
dataset: 入力データ
folds: CVのindex
Returns:
self
"""
dataset = dataset.copy()
if self.preprocessors is not None:
dataset.data = self.preprocessors.fit_transform(
dataset.data, dataset.labels)
if self.postprocessors is not None:
dataset.labels = np.squeeze(
self.postprocessors.fit_transform(
np.expand_dims(dataset.labels, axis=-1)),
axis=-1,
)
self._cv(dataset, folds)
if self.save_on_cv:
self.save()
return self
def predict(self, dataset: tk.data.Dataset) -> typing.List[np.ndarray]:
"""予測結果をリストで返す。
Args:
dataset (tk.data.Dataset): 入力データ
Returns:
len(self.folds)個の予測結果
"""
dataset = dataset.copy()
if self.preprocessors is not None:
dataset.data = self.preprocessors.transform(dataset.data)
pred_list = self._predict(dataset)
if self.postprocessors is not None:
for i in range(len(pred_list)):
if pred_list[i].ndim <= 1:
pred_list[i] = np.squeeze(
self.postprocessors.inverse_transform(
np.expand_dims(pred_list[i], axis=-1)),
axis=-1,
)
else:
pred_list[i] = self.postprocessors.inverse_transform(
pred_list[i])
return pred_list
def predict(self, dataset: tk.data.Dataset, fold: int) -> np.ndarray:
"""推論結果を返す。
Args:
dataset: 入力データ
fold: 使用するモデル
Returns:
推論結果
"""
dataset = dataset.copy()
if self.preprocessors is not None:
dataset.data = self.preprocessors.transform(dataset.data)
pred = self._predict(dataset, fold)
if self.postprocessors is not None:
if isinstance(pred, np.ndarray) and pred.ndim <= 1:
pred = np.squeeze(
self.postprocessors.inverse_transform(
np.expand_dims(pred, axis=-1)),
axis=-1,
)
else:
pred = self.postprocessors.inverse_transform(pred)
return pred
def cv(
self,
dataset: tk.data.Dataset,
folds: tk.validation.FoldsType,
models_dir: pathlib.Path,
) -> dict:
"""CVして保存。
Args:
dataset: 入力データ
folds: CVのindex
models_dir: 保存先ディレクトリ (Noneなら保存しない)
Returns:
metrics名と値
"""
if models_dir is not None:
models_dir = pathlib.Path(models_dir)
models_dir.mkdir(parents=True, exist_ok=True)
dataset = dataset.copy()
if self.preprocessors is not None:
dataset.data = self.preprocessors.fit_transform(
dataset.data, dataset.labels)
if self.postprocessors is not None:
dataset.labels = np.squeeze(
self.postprocessors.fit_transform(
np.expand_dims(dataset.labels, axis=-1)),
axis=-1,
)
scores = self._cv(dataset, folds)
if models_dir is not None:
if self.preprocessors is not None:
tk.utils.dump(self.preprocessors,
models_dir / "preprocessors.pkl")
if self.postprocessors is not None:
tk.utils.dump(self.postprocessors,
models_dir / "postprocessors.pkl")
self._save(models_dir)
return scores