feat(torch): add more errors and fix imports (#16)

Author: 34j

README.md

@@ -42,6 +42,8 @@ pip install sklearn-utilities
 
 ## API
 
+See [Docs](https://sklearn-utilities.readthedocs.io/en/latest/sklearn_utilities.html) for more information.
+
 - `EstimatorWrapperBase`: base class for wrappers. Redirects all attributes which are not in the wrapper to the wrapped estimator.
 - `DataFrameWrapper`: tries to convert every estimator output to a pandas DataFrame or Series.
 - `FeatureUnionPandas`: a `FeatureUnion` that works with pandas DataFrames.
@@ -56,9 +58,10 @@ pip install sklearn-utilities
 - `RecursiveFitSubtractRegressor`: a regressor that recursively fits a regressor and subtracts the prediction from the target.
 - `SmartMultioutputEstimator`: a `MultiOutputEstimator` that supports tuple of arrays in `predict()` and supports pandas `Series` and `DataFrame`.
 - `until_event()`, `since_event()`: calculates the time since or until events (`Series[bool]`)
-- `ComposeVarEstimator`: compose mean and std/var estimators.
+- `ComposeVarEstimator`: composes mean and std/var estimators.
 - `DummyRegressorVar`: `DummyRegressor` that returns 1.0 for std/var.
 - `TransformedTargetRegressorVar`: `TransformedTargetRegressor` with std/var support.
+- `StandardScalerVar`: `StandardScaler` with std/var support.
 
 ### `sklearn_utilities.dataset`
 
@@ -70,7 +73,8 @@ pip install sklearn-utilities
 
 #### `sklearn_utilities.torch.skorch`
 
-- `SkorchReshaper`, `SkorchCNNReshaper`: reshape X and y for `nn.Linear` and `nn.Conv1d/2d` respectively. (For `nn.Conv2d`, uses `np.sliding_window_view()`.)
+- `SkorchReshaper`, `SkorchCNNReshaper`: reshapes X and y for `nn.Linear` and `nn.Conv1d/2d` respectively. (For `nn.Conv2d`, uses `np.sliding_window_view()`.)
+- `AllowNaN`: wraps a loss module and assign 0 to y and y_hat for indices where y contains NaN in `forward()`..
 
 ## See also
 

src/sklearn_utilities/__init__.py

@@ -19,6 +19,7 @@
     ComposeVarEstimator,
     DummyRegressorVar,
     PipelineVar,
+    StandardScalerVar,
     TransformedTargetEstimatorVar,
 )
 from .recursive_fit_subtract_regressor import RecursiveFitSubtractRegressor
@@ -45,4 +46,5 @@
     "DummyRegressorVar",
     "TransformedTargetEstimatorVar",
     "PipelineVar",
+    "StandardScalerVar",
 ]

src/sklearn_utilities/proba/__init__.py

@@ -1,11 +1,13 @@
 from .compose_var import ComposeVarEstimator
 from .dummy_regressor import DummyRegressorVar
 from .pipeline_var import PipelineVar
+from .standard_scaler_var import StandardScalerVar
 from .transformed_target_estimator import TransformedTargetEstimatorVar
 
 __all__ = [
     "ComposeVarEstimator",
     "DummyRegressorVar",
     "PipelineVar",
     "TransformedTargetEstimatorVar",
+    "StandardScalerVar",
 ]

src/sklearn_utilities/torch/skorch/__init__.py

@@ -1,9 +1,14 @@
 from .proba import (
+    AlgebraicErrors,
+    AllowNan,
     AsymmetricLoss,
     AsymmetricLosses,
     LNErrors,
+    LogCoshErrors,
     SkorchCNNReshaperProba,
     SkorchReshaperProba,
+    XSigmoidErrors,
+    XTanhErrors,
 )
 from .reshaper import SkorchCNNReshaper, SkorchReshaper
 
@@ -13,6 +18,11 @@
     "SkorchReshaperProba",
     "SkorchCNNReshaperProba",
     "LNErrors",
+    "LogCoshErrors",
+    "XSigmoidErrors",
+    "XTanhErrors",
+    "AlgebraicErrors",
+    "AllowNan",
     "AsymmetricLoss",
     "AsymmetricLosses",
     "SkorchReshaperProba",

src/sklearn_utilities/torch/skorch/proba.py

@@ -1,10 +1,12 @@
 from __future__ import annotations
 
+import math
 from typing import Any, Generic, Literal, Sequence
 
 import numpy as np
 import torch
 import torch.nn as nn
+import torch.nn.functional as F
 
 from ...types import TX, TY, TEstimator
 from .reshaper import SkorchCNNReshaper, SkorchReshaper
@@ -64,6 +66,78 @@ def forward(self, y_pred: torch.Tensor, y_true: torch.Tensor) -> torch.Tensor:
         return torch.abs(y_true - y_pred).pow(self.n)
 
 
+class LogCoshErrors(nn.Module):
+    """Log cosh errors.
+    Loss = log(cosh(errors + eps))
+
+    See also
+    --------
+    https://datascience.stackexchange.com/questions/96271/logcoshloss-on-pytorch
+    """
+
+    def __init__(self, *, softplus: bool = True, eps: float | None = None) -> None:
+        """Returns log(cosh(errors)).
+
+        Parameters
+        ----------
+        softplus : bool, optional
+            If True, uses softplus to get stable results, by default True"""
+        super().__init__()
+        self.softplus = softplus
+        self.eps = eps
+
+    def forward(self, y_pred: torch.Tensor, y_true: torch.Tensor) -> torch.Tensor:
+        x = y_pred - y_true
+        if self.softplus:
+            return x + F.softplus(-2.0 * x) - math.log(2.0)
+        else:
+            eps = self.eps or torch.finfo(x.dtype).eps
+            return torch.log(torch.cosh(x + eps))
+
+
+class XTanhErrors(nn.Module):
+    """XTanh errors.
+    Loss = x * tanh(x)
+
+    See also
+    --------
+    https://github.com/tuantle/regression-losses-pytorch
+    """
+
+    def forward(self, y_pred: torch.Tensor, y_true: torch.Tensor) -> torch.Tensor:
+        x = y_pred - y_true
+        return x * torch.tanh(x)
+
+
+class XSigmoidErrors(nn.Module):
+    """XSigmoid errors.
+    Loss = x * (2 * sigmoid(x) - 1)
+
+    See also
+    --------
+    https://github.com/tuantle/regression-losses-pytorch
+    """
+
+    def forward(self, y_pred: torch.Tensor, y_true: torch.Tensor) -> torch.Tensor:
+        x = y_pred - y_true
+        return x * (2 * torch.sigmoid(x) - 1)
+
+
+class AlgebraicErrors(nn.Module):
+    """Algebraic errors.
+    Loss = x^2 / sqrt(1 + x^2)
+
+    See also
+    --------
+    https://github.com/tuantle/regression-losses-pytorch
+    """
+
+    def forward(self, y_pred: torch.Tensor, y_true: torch.Tensor) -> torch.Tensor:
+        x = y_pred - y_true
+        x2 = torch.pow(x, 2)
+        return x2 / torch.sqrt(1 + x2)
+
+
 class AsymmetricLoss(nn.Module):
     def __init__(
         self,