Add support for Python 3.14 (#1962)

### Summary

This PR enables compatibility with the latest Python 3.14, while
dropping support for the old Python 3.10. Removing 3.10 was necessary
because there's no numpy version which supports 3.10 and 3.14 at the
same time. This PR also upgrades some dependency packages to a version
compatible with 3.14; for breaking changes (e.g. in `shapely`) I have
updated the source code.
CI was also updated to test both the minimum (3.11) and maximum (3.14)
supported versions.

Resolves #1960 

### Checklist
- [x] I have added tests to cover my changes or documented any manual
tests.
- [x] I have updated the
[documentation](https://github.com/open-edge-platform/datumaro/tree/develop/docs)
accordingly

---------

Signed-off-by: Leonardo Lai <[email protected]>

Author: leoll2

.github/workflows/linter.yml

@@ -28,7 +28,7 @@ jobs:
           persist-credentials: false
       - uses: actions/setup-python@e797f83bcb11b83ae66e0230d6156d7c80228e7c # v6.0.0
         with:
-          python-version: "3.10"
+          python-version: "3.11"
 
       # Install pre-commit
       - name: Install pre-commit

.github/workflows/pr_check.yml

@@ -31,7 +31,7 @@ jobs:
       fail-fast: false
       matrix:
         os: ["ubuntu-24.04", "windows-2022", "macos-15"]
-        python-version: ["3.10", "3.13"]
+        python-version: ["3.11", "3.14"]
     name: pr test (${{ matrix.os }}, Python ${{ matrix.python-version }})
     runs-on: ${{ matrix.os }}
     steps:

.github/workflows/publish_to_pypi.yml

@@ -53,10 +53,10 @@ jobs:
         uses: actions/checkout@08c6903cd8c0fde910a37f88322edcfb5dd907a8 # v5.0.0
         with:
           persist-credentials: false
-      - name: Set up Python 3.10
+      - name: Set up Python
         uses: actions/setup-python@e797f83bcb11b83ae66e0230d6156d7c80228e7c # v6.0.0
         with:
-          python-version: "3.10"
+          python-version: "3.14"
       - name: Install pypa/build
         run: python -m pip install build
       - name: Build sdist

pyproject.toml

@@ -31,31 +31,31 @@ maintainers = [
 ]
 license-files = ["LICENSE"]
 readme = "README.md"
-requires-python = ">=3.10,<3.14"
+requires-python = ">=3.11,<=3.14"
 classifiers = [
     "Programming Language :: Python :: 3",
-    "Programming Language :: Python :: 3.10",
     "Programming Language :: Python :: 3.11",
     "Programming Language :: Python :: 3.12",
     "Programming Language :: Python :: 3.13",
+    "Programming Language :: Python :: 3.14",
     "Operating System :: OS Independent",
 ]
 dependencies = [
     "attrs>=21.3.0",
     "defusedxml>=0.7.0",
     "imagesize>=1.4.1",
     "lxml<7,>=5.2.0",
-    "numpy<2.3",
+    "numpy>=2.3.5",
     "orjson>=3.10.0",
     "Pillow>=10.3.0",
     "ruamel.yaml>=0.17.0",
-    "shapely>=1.7",
+    "shapely>=2.1.2",
     "typing_extensions>=3.7.4.3",
     "tqdm>=4.67.1",
     "pycocotools>=2.0.4",
     "PyYAML==6.0.3",
-    "pandas>=1.4.0",
-    "pyarrow",
+    "pandas>=2.3.3",
+    "pyarrow>=22.0",
     "json-stream",
     "opencv-python-headless>=4.11.0.86",
 ]
@@ -97,7 +97,7 @@ tf = [
 #]
 
 # PyTorch support
-torch = ["torch", "torchvision"]
+torch = ["torch>=2.9", "torchvision>=0.24"]
 
 # Kaggle download support
 kaggle = ["kaggle"]
@@ -119,15 +119,15 @@ cli = ["tensorboardX>=1.8,!=2.3", "tabulate", "scipy", "matplotlib>=3.3.1"]
 visualizer = ["matplotlib>=3.3.1"]
 
 # NYU Depth Dataset v2 file format
-h5py = ["h5py>=2.10.0"]
+h5py = ["h5py>=3.15.0"]
 
 # BraTS file format
 nibabel = ["nibabel>=3.2.1"]
 
 # AVA dataset
 protobuf = ["protobuf"]
 
-experimental = ["polars>=1.31.0,<2.0.0"]
+experimental = ["polars~=1.35"]
 
 [tool.coverage.run]
 branch = true

src/datumaro/components/annotation.py

@@ -13,8 +13,8 @@
 import attr
 import cv2
 import numpy as np
-import shapely.geometry as sg
 from attr import asdict, attrs, field
+from shapely import Polygon as ShapelyPolygon
 from typing_extensions import Literal
 
 from datumaro.components.media import Image
@@ -899,8 +899,8 @@ def __eq__(self, other):
 
         self_points = self.get_points()
         other_points = other.get_points()
-        self_polygon = sg.Polygon(self_points)
-        other_polygon = sg.Polygon(other_points)
+        self_polygon = ShapelyPolygon(self_points)
+        other_polygon = ShapelyPolygon(other_points)
         # if polygon is not valid, compare points
         if not (self_polygon.is_valid and other_polygon.is_valid):
             return self_points == other_points

src/datumaro/experimental/converters/annotation_converters.py

@@ -93,10 +93,10 @@ def convert(self, df: pl.DataFrame) -> pl.DataFrame:
 
         if field.multi_label or field.is_list:
             mapping_expr = pl.col(input_col).list.eval(
-                pl.element().replace(list(index_mapping.keys()), list(index_mapping.values()), default=None)
+                pl.element().replace_strict(list(index_mapping.keys()), list(index_mapping.values()), default=None)
             )
         else:
-            mapping_expr = pl.col(input_col).replace(
+            mapping_expr = pl.col(input_col).replace_strict(
                 list(index_mapping.keys()), list(index_mapping.values()), default=None
             )
 

src/datumaro/experimental/tiling/tilers.py

@@ -11,8 +11,9 @@
 
 import numpy as np
 import polars as pl
-import shapely.geometry as sg
-import shapely.ops as so
+from shapely import GeometryCollection, MultiPolygon, box, transform
+from shapely import Polygon as ShapelyPolygon
+from shapely.geometry.base import BaseGeometry
 
 from datumaro.experimental.fields import (
     BBoxField,
@@ -326,9 +327,9 @@ def extract_tile(
         return pl.DataFrame(results)
 
 
-def _apply_offset(geom: sg.base.BaseGeometry, offset_x: float, offset_y: float) -> sg.base.BaseGeometry:
+def _apply_offset(geom: BaseGeometry, offset_x: float, offset_y: float) -> BaseGeometry:
     """Apply offset to geometry."""
-    return so.transform(lambda x, y: (x - offset_x, y - offset_y), geom)
+    return transform(geometry=geom, transformation=lambda x, y: (x - offset_x, y - offset_y), interleaved=False)
 
 
 @TilerRegistry.register(PolygonField)
@@ -366,7 +367,7 @@ def tile(self, df: pl.DataFrame, tiles_df: pl.DataFrame, slice_offset: int = 0)
             source_polygons = df[source_idx, column_name]
 
             # Create tile polygon
-            tile_poly = sg.box(
+            tile_poly = box(
                 tile_row["x"],
                 tile_row["y"],
                 tile_row["x"] + tile_row["width"],
@@ -378,13 +379,13 @@ def tile(self, df: pl.DataFrame, tiles_df: pl.DataFrame, slice_offset: int = 0)
             polygon_keeps = []  # Track which polygons to keep
 
             for poly_coords in source_polygons:
-                polygon = sg.Polygon(poly_coords)
+                polygon = ShapelyPolygon(poly_coords)
 
                 # Get intersection and apply offset
                 intersection = polygon.intersection(tile_poly)
 
                 # NOTE: intersection may return a GeometryCollection or MultiPolygon
-                if isinstance(intersection, sg.GeometryCollection | sg.MultiPolygon):
+                if isinstance(intersection, GeometryCollection | MultiPolygon):
                     shapes = [(geom, geom.area) for geom in list(intersection.geoms) if geom.is_valid]
                     if not shapes:
                         tiled_polygons.append(None)  # Placeholder for dropped polygon
@@ -393,7 +394,7 @@ def tile(self, df: pl.DataFrame, tiles_df: pl.DataFrame, slice_offset: int = 0)
 
                     intersection, _ = max(shapes, key=operator.itemgetter(1))
 
-                if not isinstance(intersection, sg.Polygon) or intersection.is_empty or not intersection.is_valid:
+                if not isinstance(intersection, ShapelyPolygon) or intersection.is_empty or not intersection.is_valid:
                     tiled_polygons.append(None)  # Placeholder for dropped polygon
                     polygon_keeps.append(False)
                     continue

src/datumaro/plugins/data_formats/kaggle/base.py

@@ -425,7 +425,7 @@ def _parse_annotations(self, img_file: str, ann_file: str):
             label_name = self._parse_field(object_elem, "name", str, required=True)
 
             bbox_elem = object_elem.find("bndbox")
-            if not bbox_elem:
+            if bbox_elem is None:
                 raise MissingFieldError("bndbox")
 
             xmin = self._parse_field(bbox_elem, "xmin", float)

src/datumaro/plugins/data_formats/mpii/mpii_mat.py

@@ -51,11 +51,11 @@ def _load_items(self, path):
             group_num = 1
 
             image = getattr(item, "image", "")
-            if isinstance(image, spio.matlab.mio5_params.mat_struct):
+            if isinstance(image, spio.matlab.mat_struct):
                 image = getattr(image, "name", "")
 
             anno_values = getattr(item, "annorect", [])
-            if isinstance(anno_values, spio.matlab.mio5_params.mat_struct):
+            if isinstance(anno_values, spio.matlab.mat_struct):
                 anno_values = [anno_values]
 
             for val in anno_values:
@@ -72,12 +72,12 @@ def _load_items(self, path):
                     attributes["scale"] = scale
 
                 objpos = getattr(val, "objpos", None)
-                if isinstance(objpos, spio.matlab.mio5_params.mat_struct):
+                if isinstance(objpos, spio.matlab.mat_struct):
                     attributes["center"] = [getattr(objpos, "x", 0), getattr(objpos, "y", 0)]
 
                 annopoints = getattr(val, "annopoints", None)
-                if isinstance(annopoints, spio.matlab.mio5_params.mat_struct) and not isinstance(
-                    getattr(annopoints, "point"), spio.matlab.mio5_params.mat_struct
+                if isinstance(annopoints, spio.matlab.mat_struct) and not isinstance(
+                    getattr(annopoints, "point"), spio.matlab.mat_struct
                 ):
                     for point in getattr(annopoints, "point"):
                         point_id = getattr(point, "id")

src/datumaro/plugins/data_formats/voc/base.py

@@ -250,7 +250,7 @@ def _parse_attribute(self, object_elem):
     @classmethod
     def _parse_bbox(cls, object_elem):
         bbox_elem = object_elem.find("bndbox")
-        if not bbox_elem:
+        if bbox_elem is None:
             raise MissingFieldError("bndbox")
 
         xmin = cls._parse_field(bbox_elem, "xmin", float)