Merge branch 'main' into doc_order

Author: NicolasHug

.circleci/unittest/linux/scripts/environment.yml

@@ -9,6 +9,7 @@ dependencies:
   - libpng
   - jpeg
   - ca-certificates
+  - h5py
   - pip:
     - future
     - pillow >=5.3.0, !=8.3.*

.circleci/unittest/windows/scripts/environment.yml

@@ -9,6 +9,7 @@ dependencies:
   - libpng
   - jpeg
   - ca-certificates
+  - h5py
   - pip:
     - future
     - pillow >=5.3.0, !=8.3.*

CMakeLists.txt

@@ -4,6 +4,7 @@ set(CMAKE_CXX_STANDARD 14)
 file(STRINGS version.txt TORCHVISION_VERSION)
 
 option(WITH_CUDA "Enable CUDA support" OFF)
+option(USE_PYTHON "Link to Python when building" OFF)
 
 if(WITH_CUDA)
   enable_language(CUDA)
@@ -17,7 +18,10 @@ if(WITH_CUDA)
   endif()
 endif()
 
-find_package(Python3 COMPONENTS Development)
+if (USE_PYTHON)
+  add_definitions(-DUSE_PYTHON)
+  find_package(Python3 REQUIRED COMPONENTS Development)
+endif()
 
 find_package(Torch REQUIRED)
 find_package(PNG REQUIRED)
@@ -76,7 +80,12 @@ FOREACH(DIR ${ALLOW_LISTED})
 ENDFOREACH()
 
 add_library(${PROJECT_NAME} SHARED ${ALL_SOURCES})
-target_link_libraries(${PROJECT_NAME} PRIVATE ${TORCH_LIBRARIES} ${PNG_LIBRARY} ${JPEG_LIBRARIES} Python3::Python)
+target_link_libraries(${PROJECT_NAME} PRIVATE ${TORCH_LIBRARIES} ${PNG_LIBRARY} ${JPEG_LIBRARIES})
+
+if (USE_PYTHON)
+  target_link_libraries(${PROJECT_NAME} PRIVATE Python3::Python)
+endif()
+
 set_target_properties(${PROJECT_NAME} PROPERTIES
   EXPORT_NAME TorchVision
   INSTALL_RPATH ${TORCH_INSTALL_PREFIX}/lib)

README.rst

@@ -157,6 +157,10 @@ so make sure that it is also available to cmake via the ``CMAKE_PREFIX_PATH``.
 
 For an example setup, take a look at ``examples/cpp/hello_world``.
 
+Python linking is disabled by default when compiling TorchVision with CMake, this allows you to run models without any Python 
+dependency. In some special cases where TorchVision's operators are used from Python code, you may need to link to Python. This 
+can be done by passing ``-DUSE_PYTHON=on`` to CMake.
+
 TorchVision Operators
 ---------------------
 In order to get the torchvision operators registered with torch (eg. for the JIT), all you need to do is to ensure that you

cmake/TorchVisionConfig.cmake.in

@@ -28,8 +28,10 @@ include("${CMAKE_CURRENT_LIST_DIR}/${PN}Targets.cmake")
 if(NOT TARGET torch_library)
 find_package(Torch REQUIRED)
 endif()
-if(NOT TARGET Python3::Python)
-find_package(Python3 COMPONENTS Development)
+if (@USE_PYTHON@)
+  if(NOT TARGET Python3::Python)
+    find_package(Python3 COMPONENTS Development)
+  endif()
 endif()
 
 set_target_properties(TorchVision::TorchVision PROPERTIES INTERFACE_INCLUDE_DIRECTORIES "${${PN}_INCLUDE_DIR}" INTERFACE_LINK_LIBRARIES "torch;Python3::Python" )

docs/source/datasets.rst

@@ -41,6 +41,7 @@ You can also create your own datasets using the provided :ref:`base classes <bas
     Country211
     DTD
     EMNIST
+    EuroSAT
     FakeData
     FashionMNIST
     FER2013
@@ -50,6 +51,7 @@ You can also create your own datasets using the provided :ref:`base classes <bas
     FlyingChairs
     FlyingThings3D
     Food101
+    FGVCAircraft
     GTSRB
     HD1K
     HMDB51
@@ -65,13 +67,15 @@ You can also create your own datasets using the provided :ref:`base classes <bas
     MNIST
     Omniglot
     OxfordIIITPet
+    PCAM
     PhotoTour
     Places365
     QMNIST
     SBDataset
     SBU
     SEMEION
     Sintel
+    StanfordCars
     STL10
     SUN397
     SVHN

examples/cpp/hello_world/CMakeLists.txt

@@ -6,6 +6,10 @@ project(hello-world)
 # so there is no need to also add `find_package(Torch)` here.
 find_package(TorchVision REQUIRED)
 
+# This due to LibTorch's version is the one included in the Python
+# package that links to Python.
+find_package(Python3 COMPONENTS Development)
+
 add_executable(hello-world main.cpp)
 
 # We now need to link the TorchVision library to our executable.

setup.py

@@ -201,7 +201,7 @@ def get_extensions():
 
     if sys.platform == "win32":
         define_macros += [("torchvision_EXPORTS", None)]
-
+        define_macros += [("USE_PYTHON", None)]
         extra_compile_args["cxx"].append("/MP")
 
     debug_mode = os.getenv("DEBUG", "0") == "1"
@@ -254,6 +254,9 @@ def get_extensions():
     image_library = []
     image_link_flags = []
 
+    if sys.platform == "win32":
+        image_macros += [("USE_PYTHON", None)]
+
     # Locating libPNG
     libpng = distutils.spawn.find_executable("libpng-config")
     pngfix = distutils.spawn.find_executable("pngfix")

test/datasets_utils.py

@@ -61,6 +61,7 @@ class LazyImporter:
         "requests",
         "scipy.io",
         "scipy.sparse",
+        "h5py",
     )
 
     def __init__(self):

test/test_datasets.py

@@ -2169,6 +2169,27 @@ def inject_fake_data(self, tmpdir, config):
         return num_sequences * (num_examples_per_sequence - 1)
 
 
+class EuroSATTestCase(datasets_utils.ImageDatasetTestCase):
+    DATASET_CLASS = datasets.EuroSAT
+    FEATURE_TYPES = (PIL.Image.Image, int)
+
+    def inject_fake_data(self, tmpdir, config):
+        data_folder = os.path.join(tmpdir, "eurosat", "2750")
+        os.makedirs(data_folder)
+
+        num_examples_per_class = 3
+        classes = ("AnnualCrop", "Forest")
+        for cls in classes:
+            datasets_utils.create_image_folder(
+                root=data_folder,
+                name=cls,
+                file_name_fn=lambda idx: f"{cls}_{idx}.jpg",
+                num_examples=num_examples_per_class,
+            )
+
+        return len(classes) * num_examples_per_class
+
+
 class Food101TestCase(datasets_utils.ImageDatasetTestCase):
     DATASET_CLASS = datasets.Food101
     FEATURE_TYPES = (PIL.Image.Image, int)
@@ -2206,6 +2227,57 @@ def inject_fake_data(self, tmpdir: str, config):
         return len(sampled_classes * n_samples_per_class)
 
 
+class FGVCAircraftTestCase(datasets_utils.ImageDatasetTestCase):
+    DATASET_CLASS = datasets.FGVCAircraft
+    ADDITIONAL_CONFIGS = datasets_utils.combinations_grid(
+        split=("train", "val", "trainval", "test"), annotation_level=("variant", "family", "manufacturer")
+    )
+
+    def inject_fake_data(self, tmpdir: str, config):
+        split = config["split"]
+        annotation_level = config["annotation_level"]
+        annotation_level_to_file = {
+            "variant": "variants.txt",
+            "family": "families.txt",
+            "manufacturer": "manufacturers.txt",
+        }
+
+        root_folder = pathlib.Path(tmpdir) / "fgvc-aircraft-2013b"
+        data_folder = root_folder / "data"
+
+        classes = ["707-320", "Hawk T1", "Tornado"]
+        num_images_per_class = 5
+
+        datasets_utils.create_image_folder(
+            data_folder,
+            "images",
+            file_name_fn=lambda idx: f"{idx}.jpg",
+            num_examples=num_images_per_class * len(classes),
+        )
+
+        annotation_file = data_folder / annotation_level_to_file[annotation_level]
+        with open(annotation_file, "w") as file:
+            file.write("\n".join(classes))
+
+        num_samples_per_class = 4 if split == "trainval" else 2
+        images_classes = []
+        for i in range(len(classes)):
+            images_classes.extend(
+                [
+                    f"{idx} {classes[i]}"
+                    for idx in random.sample(
+                        range(i * num_images_per_class, (i + 1) * num_images_per_class), num_samples_per_class
+                    )
+                ]
+            )
+
+        images_annotation_file = data_folder / f"images_{annotation_level}_{split}.txt"
+        with open(images_annotation_file, "w") as file:
+            file.write("\n".join(images_classes))
+
+        return len(classes * num_samples_per_class)
+
+
 class SUN397TestCase(datasets_utils.ImageDatasetTestCase):
     DATASET_CLASS = datasets.SUN397
 
@@ -2463,6 +2535,50 @@ def _meta_to_split_and_classification_ann(self, meta, idx):
         return (image_id, class_id, species, breed_id)
 
 
+class StanfordCarsTestCase(datasets_utils.ImageDatasetTestCase):
+    DATASET_CLASS = datasets.StanfordCars
+    REQUIRED_PACKAGES = ("scipy",)
+    ADDITIONAL_CONFIGS = datasets_utils.combinations_grid(split=("train", "test"))
+
+    def inject_fake_data(self, tmpdir, config):
+        import scipy.io as io
+        from numpy.core.records import fromarrays
+
+        num_examples = {"train": 5, "test": 7}[config["split"]]
+        num_classes = 3
+        base_folder = pathlib.Path(tmpdir) / "stanford_cars"
+
+        devkit = base_folder / "devkit"
+        devkit.mkdir(parents=True)
+
+        if config["split"] == "train":
+            images_folder_name = "cars_train"
+            annotations_mat_path = devkit / "cars_train_annos.mat"
+        else:
+            images_folder_name = "cars_test"
+            annotations_mat_path = base_folder / "cars_test_annos_withlabels.mat"
+
+        datasets_utils.create_image_folder(
+            root=base_folder,
+            name=images_folder_name,
+            file_name_fn=lambda image_index: f"{image_index:5d}.jpg",
+            num_examples=num_examples,
+        )
+
+        classes = np.random.randint(1, num_classes + 1, num_examples, dtype=np.uint8)
+        fnames = [f"{i:5d}.jpg" for i in range(num_examples)]
+        rec_array = fromarrays(
+            [classes, fnames],
+            names=["class", "fname"],
+        )
+        io.savemat(annotations_mat_path, {"annotations": rec_array})
+
+        random_class_names = ["random_name"] * num_classes
+        io.savemat(devkit / "cars_meta.mat", {"class_names": random_class_names})
+
+        return num_examples
+
+
 class Country211TestCase(datasets_utils.ImageDatasetTestCase):
     DATASET_CLASS = datasets.Country211
 
@@ -2526,5 +2642,28 @@ def inject_fake_data(self, tmpdir: str, config):
         return num_images_per_split[config["split"]]
 
 
+class PCAMTestCase(datasets_utils.ImageDatasetTestCase):
+    DATASET_CLASS = datasets.PCAM
+
+    ADDITIONAL_CONFIGS = datasets_utils.combinations_grid(split=("train", "val", "test"))
+    REQUIRED_PACKAGES = ("h5py",)
+
+    def inject_fake_data(self, tmpdir: str, config):
+        base_folder = pathlib.Path(tmpdir) / "pcam"
+        base_folder.mkdir()
+
+        num_images = {"train": 2, "test": 3, "val": 4}[config["split"]]
+
+        images_file = datasets.PCAM._FILES[config["split"]]["images"][0]
+        with datasets_utils.lazy_importer.h5py.File(str(base_folder / images_file), "w") as f:
+            f["x"] = np.random.randint(0, 256, size=(num_images, 10, 10, 3), dtype=np.uint8)
+
+        targets_file = datasets.PCAM._FILES[config["split"]]["targets"][0]
+        with datasets_utils.lazy_importer.h5py.File(str(base_folder / targets_file), "w") as f:
+            f["y"] = np.random.randint(0, 2, size=(num_images, 1, 1, 1), dtype=np.uint8)
+
+        return num_images
+
+
 if __name__ == "__main__":
     unittest.main()