From 5da5f4ac9a14aa0d3893b3b5f25525c323c06185 Mon Sep 17 00:00:00 2001
From: Jakob <jakob@winter-wind.de>
Date: Thu, 15 Feb 2024 13:37:31 +0100
Subject: [PATCH] add base transform class

---
 src/continuity/transforms/__init__.py  |  5 +++
 src/continuity/transforms/transform.py | 47 ++++++++++++++++++++++++++
 2 files changed, 52 insertions(+)
 create mode 100644 src/continuity/transforms/__init__.py
 create mode 100644 src/continuity/transforms/transform.py

diff --git a/src/continuity/transforms/__init__.py b/src/continuity/transforms/__init__.py
new file mode 100644
index 00000000..fbc75b71
--- /dev/null
+++ b/src/continuity/transforms/__init__.py
@@ -0,0 +1,5 @@
+from transform import Transform
+
+__all__ = [
+    "Transform",
+]
diff --git a/src/continuity/transforms/transform.py b/src/continuity/transforms/transform.py
new file mode 100644
index 00000000..a48bbccf
--- /dev/null
+++ b/src/continuity/transforms/transform.py
@@ -0,0 +1,47 @@
+import torch
+import torch.nn as nn
+import warnings
+
+from abc import ABC, abstractmethod
+
+
+class Transform(nn.Module, ABC):
+    """Abstract base class for transformations of tensors.
+
+    Transformations are applied to tensors to improve model performance, enhance generalization, handle varied input
+    sizes, facilitate specific features, reduce overfitting, improve computational efficiency or many other reasons.
+    This class takes some tensor and transforms it into some other tensor.
+    """
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+    @abstractmethod
+    def forward(self, tensor: torch.Tensor) -> torch.Tensor:
+        """Applies the transformation.
+
+        Args:
+            tensor: Tensor that should be transformed.
+
+        Returns:
+            Transformed tensor.
+        """
+
+    def backward(self, tensor: torch.Tensor) -> torch.Tensor:
+        """Applies the inverse transformation (given the transformation is bijective).
+
+        When the transformation is not bijective (one-to-one correspondence of data) the inverse/backward transformation
+        is not applied. Instead, a warning is raised.
+
+        Args:
+            tensor: Transformed tensor.
+
+        Returns:
+            Tensor with the transformation undone (given it is possible).
+        """
+        warnings.warn(
+            f"Backward pass for transformation {self.__class__.__name__} not implement! "
+            f"Backward pass is performed as identity!",
+            stacklevel=2,
+        )
+        return tensor