[MNT] Isolate cpflow package, towards fixing readthedocs build (#1775)

The readthedocs build is failing due to attempted imports of the
`cpflow` package.

This PR fixes this by isolating the `cpflow` soft dependency imports.

For this, the `_safe_import` utility from `sktime` is copied over -
mid-term, we may want to move this to `scikit-base`.

Author: fkiraly

pytorch_forecasting/metrics/_mqf2_utils.py

@@ -2,7 +2,6 @@
 
 from typing import List, Optional, Tuple
 
-from cpflows.flows import DeepConvexFlow, SequentialFlow
 import torch
 from torch.distributions import (
     AffineTransform,
@@ -12,6 +11,11 @@
 )
 import torch.nn.functional as F
 
+from pytorch_forecasting.utils._dependencies import _safe_import
+
+DeepConvexFlow = _safe_import("cpflows.flows.DeepConvexFlow")
+SequentialFlow = _safe_import("cpflows.flows.SequentialFlow")
+
 
 class DeepConvexNet(DeepConvexFlow):
     r"""

pytorch_forecasting/utils/_dependencies/__init__.py

@@ -0,0 +1,13 @@
+"""Utilities for managing dependencies."""
+
+from pytorch_forecasting.utils._dependencies._dependencies import (
+    _check_matplotlib,
+    _get_installed_packages,
+)
+from pytorch_forecasting.utils._dependencies._safe_import import _safe_import
+
+__all__ = [
+    "_get_installed_packages",
+    "_check_matplotlib",
+    "_safe_import",
+]

…torch_forecasting/utils/_dependencies.py …ing/utils/_dependencies/_dependencies.pypytorch_forecasting/utils/_dependencies.py renamed to pytorch_forecasting/utils/_dependencies/_dependencies.py pytorch_forecasting/utils/_dependencies.py renamed to pytorch_forecasting/utils/_dependencies/_dependencies.py

@@ -0,0 +1,88 @@
+"""Import a module/class, return a Mock object if import fails.
+
+Copied from sktime/skbase.
+"""
+
+import importlib
+from unittest.mock import MagicMock
+
+from pytorch_forecasting.utils._dependencies import _get_installed_packages
+
+
+def _safe_import(import_path, pkg_name=None):
+    """Import a module/class, return a Mock object if import fails.
+
+    The function supports importing both top-level modules and nested attributes:
+
+    - Top-level module: "torch" -> imports torch
+    - Nested module: "torch.nn" -> imports torch.nn
+    - Class/function: "torch.nn.Linear" -> imports Linear class from torch.nn
+
+    Parameters
+    ----------
+    import_path : str
+        The path to the module/class to import. Can be:
+
+        - Single module: "torch"
+        - Nested module: "torch.nn"
+        - Class/attribute: "torch.nn.ReLU"
+
+        Note: The dots in the path determine the import behavior:
+
+        - No dots: Imports as a single module
+        - One dot: Imports as a submodule
+        - Multiple dots: Last part is treated as an attribute to import
+
+    pkg_name : str, default=None
+        The name of the package to check for installation. This is useful when
+        the import name differs from the package name, for example:
+
+        - import: "sklearn" -> pkg_name="scikit-learn"
+        - import: "cv2" -> pkg_name="opencv-python"
+
+        If None, uses the first part of import_path before the dot.
+
+    Returns
+    -------
+    object
+        One of the following:
+
+        - The imported module if import_path has no dots
+        - The imported submodule if import_path has one dot
+        - The imported class/function if import_path has multiple dots
+        - A MagicMock object that returns an installation message if the
+          package is not found
+
+    Examples
+    --------
+    >>> # Import a top-level module
+    >>> torch = safe_import("torch")
+
+    >>> # Import a submodule
+    >>> nn = safe_import("torch.nn")
+
+    >>> # Import a specific class
+    >>> Linear = safe_import("torch.nn.Linear")
+
+    >>> # Import with different package name
+    >>> cv2 = safe_import("cv2", pkg_name="opencv-python")
+    """
+    if pkg_name is None:
+        path_list = import_path.split(".")
+        pkg_name = path_list[0]
+
+    if pkg_name in _get_installed_packages():
+        try:
+            if len(path_list) == 1:
+                return importlib.import_module(pkg_name)
+            module_name, attr_name = import_path.rsplit(".", 1)
+            module = importlib.import_module(module_name)
+            return getattr(module, attr_name)
+        except (ImportError, AttributeError):
+            return importlib.import_module(import_path)
+    else:
+        mock_obj = MagicMock()
+        mock_obj.__call__ = MagicMock(
+            return_value=f"Please install {pkg_name} to use this functionality."
+        )
+        return mock_obj