Merge branch 'main' into models/enhanced_meta

Author: datumbox

docs/source/datasets.rst

@@ -48,6 +48,7 @@ You can also create your own datasets using the provided :ref:`base classes <bas
     FlyingChairs
     FlyingThings3D
     Food101
+    GTSRB
     HD1K
     HMDB51
     ImageNet

test/test_datasets.py

@@ -2275,5 +2275,87 @@ def inject_fake_data(self, tmpdir, config):
         return num_samples
 
 
+class GTSRBTestCase(datasets_utils.ImageDatasetTestCase):
+    DATASET_CLASS = datasets.GTSRB
+    FEATURE_TYPES = (PIL.Image.Image, int)
+
+    ADDITIONAL_CONFIGS = datasets_utils.combinations_grid(train=(True, False))
+
+    def inject_fake_data(self, tmpdir: str, config):
+        root_folder = os.path.join(tmpdir, "GTSRB")
+        os.makedirs(root_folder, exist_ok=True)
+
+        # Train data
+        train_folder = os.path.join(root_folder, "Training")
+        os.makedirs(train_folder, exist_ok=True)
+
+        num_examples = 3
+        classes = ("00000", "00042", "00012")
+        for class_idx in classes:
+            datasets_utils.create_image_folder(
+                train_folder,
+                name=class_idx,
+                file_name_fn=lambda image_idx: f"{class_idx}_{image_idx:05d}.ppm",
+                num_examples=num_examples,
+            )
+
+        total_number_of_examples = num_examples * len(classes)
+        # Test data
+        test_folder = os.path.join(root_folder, "Final_Test", "Images")
+        os.makedirs(test_folder, exist_ok=True)
+
+        with open(os.path.join(root_folder, "GT-final_test.csv"), "w") as csv_file:
+            csv_file.write("Filename;Width;Height;Roi.X1;Roi.Y1;Roi.X2;Roi.Y2;ClassId\n")
+
+            for _ in range(total_number_of_examples):
+                image_file = datasets_utils.create_random_string(5, string.digits) + ".ppm"
+                datasets_utils.create_image_file(test_folder, image_file)
+                row = [
+                    image_file,
+                    torch.randint(1, 100, size=()).item(),
+                    torch.randint(1, 100, size=()).item(),
+                    torch.randint(1, 100, size=()).item(),
+                    torch.randint(1, 100, size=()).item(),
+                    torch.randint(1, 100, size=()).item(),
+                    torch.randint(1, 100, size=()).item(),
+                    torch.randint(0, 43, size=()).item(),
+                ]
+                csv_file.write(";".join(map(str, row)) + "\n")
+
+        return total_number_of_examples
+
+
+class CLEVRClassificationTestCase(datasets_utils.ImageDatasetTestCase):
+    DATASET_CLASS = datasets.CLEVRClassification
+    FEATURE_TYPES = (PIL.Image.Image, (int, type(None)))
+
+    ADDITIONAL_CONFIGS = datasets_utils.combinations_grid(split=("train", "val", "test"))
+
+    def inject_fake_data(self, tmpdir, config):
+        data_folder = pathlib.Path(tmpdir) / "clevr" / "CLEVR_v1.0"
+
+        images_folder = data_folder / "images"
+        image_files = datasets_utils.create_image_folder(
+            images_folder, config["split"], lambda idx: f"CLEVR_{config['split']}_{idx:06d}.png", num_examples=5
+        )
+
+        scenes_folder = data_folder / "scenes"
+        scenes_folder.mkdir()
+        if config["split"] != "test":
+            with open(scenes_folder / f"CLEVR_{config['split']}_scenes.json", "w") as file:
+                json.dump(
+                    dict(
+                        info=dict(),
+                        scenes=[
+                            dict(image_filename=image_file.name, objects=[dict()] * int(torch.randint(10, ())))
+                            for image_file in image_files
+                        ],
+                    ),
+                    file,
+                )
+
+        return len(image_files)
+
+
 if __name__ == "__main__":
     unittest.main()

test/test_image.py

@@ -478,7 +478,8 @@ def test_write_jpeg_reference(img_path, tmpdir):
     assert_equal(torch_bytes, pil_bytes)
 
 
-@pytest.mark.skipif(IS_WINDOWS, reason=("this test fails on windows because PIL uses libjpeg-turbo on windows"))
+# TODO: Remove the skip. See https://github.com/pytorch/vision/issues/5162.
+@pytest.mark.skip("this test fails because PIL uses libjpeg-turbo")
 @pytest.mark.parametrize(
     "img_path",
     [pytest.param(jpeg_path, id=_get_safe_image_name(jpeg_path)) for jpeg_path in get_images(ENCODE_JPEG, ".jpg")],
@@ -497,7 +498,8 @@ def test_encode_jpeg(img_path):
         assert_equal(encoded_jpeg_torch, encoded_jpeg_pil)
 
 
-@pytest.mark.skipif(IS_WINDOWS, reason=("this test fails on windows because PIL uses libjpeg-turbo on windows"))
+# TODO: Remove the skip. See https://github.com/pytorch/vision/issues/5162.
+@pytest.mark.skip("this test fails because PIL uses libjpeg-turbo")
 @pytest.mark.parametrize(
     "img_path",
     [pytest.param(jpeg_path, id=_get_safe_image_name(jpeg_path)) for jpeg_path in get_images(ENCODE_JPEG, ".jpg")],

torchvision/datasets/__init__.py

@@ -3,13 +3,15 @@
 from .celeba import CelebA
 from .cifar import CIFAR10, CIFAR100
 from .cityscapes import Cityscapes
+from .clevr import CLEVRClassification
 from .coco import CocoCaptions, CocoDetection
 from .dtd import DTD
 from .fakedata import FakeData
 from .fer2013 import FER2013
 from .flickr import Flickr8k, Flickr30k
 from .folder import ImageFolder, DatasetFolder
 from .food101 import Food101
+from .gtsrb import GTSRB
 from .hmdb51 import HMDB51
 from .imagenet import ImageNet
 from .inaturalist import INaturalist
@@ -83,4 +85,6 @@
     "Food101",
     "DTD",
     "FER2013",
+    "GTSRB",
+    "CLEVRClassification",
 )

torchvision/datasets/clevr.py

@@ -0,0 +1,88 @@
+import json
+import pathlib
+from typing import Any, Callable, Optional, Tuple, List
+from urllib.parse import urlparse
+
+from PIL import Image
+
+from .utils import download_and_extract_archive, verify_str_arg
+from .vision import VisionDataset
+
+
+class CLEVRClassification(VisionDataset):
+    """`CLEVR <https://cs.stanford.edu/people/jcjohns/clevr/>`_  classification dataset.
+
+    The number of objects in a scene are used as label.
+
+    Args:
+        root (string): Root directory of dataset where directory ``root/clevr`` exists or will be saved to if download is
+            set to True.
+        split (string, optional): The dataset split, supports ``"train"`` (default), ``"val"``, or ``"test"``.
+        transform (callable, optional): A function/transform that takes in an PIL image and returns a transformed
+            version. E.g, ``transforms.RandomCrop``
+        target_transform (callable, optional): A function/transform that takes in them target and transforms it.
+        download (bool, optional): If true, downloads the dataset from the internet and puts it in root directory. If
+            dataset is already downloaded, it is not downloaded again.
+    """
+
+    _URL = "https://dl.fbaipublicfiles.com/clevr/CLEVR_v1.0.zip"
+    _MD5 = "b11922020e72d0cd9154779b2d3d07d2"
+
+    def __init__(
+        self,
+        root: str,
+        split: str = "train",
+        transform: Optional[Callable] = None,
+        target_transform: Optional[Callable] = None,
+        download: bool = True,
+    ) -> None:
+        self._split = verify_str_arg(split, "split", ("train", "val", "test"))
+        super().__init__(root, transform=transform, target_transform=target_transform)
+        self._base_folder = pathlib.Path(self.root) / "clevr"
+        self._data_folder = self._base_folder / pathlib.Path(urlparse(self._URL).path).stem
+
+        if download:
+            self._download()
+
+        if not self._check_exists():
+            raise RuntimeError("Dataset not found or corrupted. You can use download=True to download it")
+
+        self._image_files = sorted(self._data_folder.joinpath("images", self._split).glob("*"))
+
+        self._labels: List[Optional[int]]
+        if self._split != "test":
+            with open(self._data_folder / "scenes" / f"CLEVR_{self._split}_scenes.json") as file:
+                content = json.load(file)
+            num_objects = {scene["image_filename"]: len(scene["objects"]) for scene in content["scenes"]}
+            self._labels = [num_objects[image_file.name] for image_file in self._image_files]
+        else:
+            self._labels = [None] * len(self._image_files)
+
+    def __len__(self) -> int:
+        return len(self._image_files)
+
+    def __getitem__(self, idx: int) -> Tuple[Any, Any]:
+        image_file = self._image_files[idx]
+        label = self._labels[idx]
+
+        image = Image.open(image_file).convert("RGB")
+
+        if self.transform:
+            image = self.transform(image)
+
+        if self.target_transform:
+            label = self.target_transform(label)
+
+        return image, label
+
+    def _check_exists(self) -> bool:
+        return self._data_folder.exists() and self._data_folder.is_dir()
+
+    def _download(self) -> None:
+        if self._check_exists():
+            return
+
+        download_and_extract_archive(self._URL, str(self._base_folder), md5=self._MD5)
+
+    def extra_repr(self) -> str:
+        return f"split={self._split}"

torchvision/datasets/gtsrb.py

@@ -0,0 +1,106 @@
+import csv
+import os
+from typing import Any, Callable, Optional, Tuple
+
+import PIL
+
+from .folder import make_dataset
+from .utils import download_and_extract_archive
+from .vision import VisionDataset
+
+
+class GTSRB(VisionDataset):
+    """`German Traffic Sign Recognition Benchmark (GTSRB) <https://benchmark.ini.rub.de/>`_ Dataset.
+
+    Args:
+        root (string): Root directory of the dataset.
+        train (bool, optional): If True, creates dataset from training set, otherwise
+            creates from test set.
+        transform (callable, optional): A function/transform that  takes in an PIL image and returns a transformed
+            version. E.g, ``transforms.RandomCrop``.
+        target_transform (callable, optional): A function/transform that takes in the target and transforms it.
+        download (bool, optional): If True, downloads the dataset from the internet and
+            puts it in root directory. If dataset is already downloaded, it is not
+            downloaded again.
+    """
+
+    # Ground Truth for the test set
+    _gt_url = "https://sid.erda.dk/public/archives/daaeac0d7ce1152aea9b61d9f1e19370/GTSRB_Final_Test_GT.zip"
+    _gt_csv = "GT-final_test.csv"
+    _gt_md5 = "fe31e9c9270bbcd7b84b7f21a9d9d9e5"
+
+    # URLs for the test and train set
+    _urls = (
+        "https://sid.erda.dk/public/archives/daaeac0d7ce1152aea9b61d9f1e19370/GTSRB_Final_Test_Images.zip",
+        "https://sid.erda.dk/public/archives/daaeac0d7ce1152aea9b61d9f1e19370/GTSRB-Training_fixed.zip",
+    )
+
+    _md5s = ("c7e4e6327067d32654124b0fe9e82185", "513f3c79a4c5141765e10e952eaa2478")
+
+    def __init__(
+        self,
+        root: str,
+        train: bool = True,
+        transform: Optional[Callable] = None,
+        target_transform: Optional[Callable] = None,
+        download: bool = False,
+    ) -> None:
+
+        super().__init__(root, transform=transform, target_transform=target_transform)
+
+        self.root = os.path.expanduser(root)
+
+        self.train = train
+
+        self._base_folder = os.path.join(self.root, type(self).__name__)
+        self._target_folder = os.path.join(self._base_folder, "Training" if self.train else "Final_Test/Images")
+
+        if download:
+            self.download()
+
+        if not self._check_exists():
+            raise RuntimeError("Dataset not found. You can use download=True to download it")
+
+        if train:
+            samples = make_dataset(self._target_folder, extensions=(".ppm",))
+        else:
+            with open(os.path.join(self._base_folder, self._gt_csv)) as csv_file:
+                samples = [
+                    (os.path.join(self._target_folder, row["Filename"]), int(row["ClassId"]))
+                    for row in csv.DictReader(csv_file, delimiter=";", skipinitialspace=True)
+                ]
+
+        self._samples = samples
+        self.transform = transform
+        self.target_transform = target_transform
+
+    def __len__(self) -> int:
+        return len(self._samples)
+
+    def __getitem__(self, index: int) -> Tuple[Any, Any]:
+
+        path, target = self._samples[index]
+        sample = PIL.Image.open(path).convert("RGB")
+
+        if self.transform is not None:
+            sample = self.transform(sample)
+
+        if self.target_transform is not None:
+            target = self.target_transform(target)
+
+        return sample, target
+
+    def _check_exists(self) -> bool:
+        return os.path.exists(self._target_folder) and os.path.isdir(self._target_folder)
+
+    def download(self) -> None:
+        if self._check_exists():
+            return
+
+        download_and_extract_archive(self._urls[self.train], download_root=self.root, md5=self._md5s[self.train])
+
+        if not self.train:
+            # Download Ground Truth for the test set
+            download_and_extract_archive(
+                self._gt_url, download_root=self.root, extract_root=self._base_folder, md5=self._gt_md5
+            )

torchvision/prototype/datasets/_builtin/__init__.py

@@ -1,6 +1,7 @@
 from .caltech import Caltech101, Caltech256
 from .celeba import CelebA
 from .cifar import Cifar10, Cifar100
+from .clevr import CLEVR
 from .coco import Coco
 from .dtd import DTD
 from .fer2013 import FER2013