Fix phototour dataset

[venkatram-dev]

Fixes this issue https://github.com/pytorch/vision/issues/8732

[pytorch-bot[bot]]

:link: Helpful Links

:test_tube: See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/vision/8733

• :page_facing_up: Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

:heavy_exclamation_mark: 1 Active SEVs

There are 1 currently active SEVs. If your PR is affected, please view them below:

• [DomainsOnly] Jobs fail with GLIBC version not found

This comment was automatically generated by Dr. CI and updates every 15 minutes.

[venkatram-dev]

Code to test the changes

import torch
from torchvision.datasets import PhotoTour

def test_phototour_new():
    # Define the root directory where datasets will be stored
    root = "./datasets"

    # List of datasets to test
    datasets = ["trevi", "notredame", "halfdome"]

    for dataset_name in datasets:
        print(f"\nTesting dataset: {dataset_name}")

        # Initialize the dataset
        dataset = PhotoTour(
            root=root,
            name=dataset_name,
            train=True,
            transform=None,  # No need for transforms
            download=True,  # Download the datasets if not already present
        )

        # Check if the dataset has been loaded successfully
        assert len(dataset) > 0, f"Dataset {dataset_name} is empty!"
        print(f"Number of patches in {dataset_name}: {len(dataset)}")

        # Retrieve a sample
        sample = dataset[0]
        print(f"Sample type for {dataset_name}: {type(sample)}")
        if isinstance(sample, torch.Tensor):
            print(f"Sample shape: {sample.shape}")

        # Print the mean and standard deviation of the dataset
        print(f"Mean: {dataset.mean}, Std: {dataset.std}")

        # Access a few samples to verify functionality
        for i in range(min(5, len(dataset))):  # Test first 5 samples
            try:
                data = dataset[i]
                print(f"Sample {i}: {data.shape if isinstance(data, torch.Tensor) else type(data)}")
            except Exception as e:
                print(f"Error accessing sample {i}: {e}")

if __name__ == "__main__":
    test_phototour_new()

output

Testing dataset: trevi
Downloading https://phototour.cs.washington.edu/patches/trevi.zip to ./datasets/trevi.zip
100.0%
Number of patches in trevi: 101120
Sample type for trevi: <class 'torch.Tensor'>
Sample shape: torch.Size([64, 64])
Mean: 0.4832, Std: 0.1913
Sample 0: torch.Size([64, 64])
Sample 1: torch.Size([64, 64])
Sample 2: torch.Size([64, 64])
Sample 3: torch.Size([64, 64])
Sample 4: torch.Size([64, 64])

Testing dataset: notredame
Downloading https://phototour.cs.washington.edu/patches/notredame.zip to ./datasets/notredame.zip
100.0%
Skipping invalid patch at (0, 0) in ./datasets/notredame/patches0000 (Custom).bmp: cannot reshape array of size 12288 into shape (64,64)
Skipping invalid patch at (64, 0) in ./datasets/notredame/patches0000 (Custom).bmp: cannot reshape array of size 12288 into shape (64,64)
Skipping invalid patch at (0, 64) in ./datasets/notredame/patches0000 (Custom).bmp: cannot reshape array of size 12288 into shape (64,64)
Skipping invalid patch at (64, 64) in ./datasets/notredame/patches0000 (Custom).bmp: cannot reshape array of size 12288 into shape (64,64)
Number of patches in notredame: 104192
Sample type for notredame: <class 'torch.Tensor'>
Sample shape: torch.Size([64, 64])
Mean: 0.4757, Std: 0.1931
Sample 0: torch.Size([64, 64])
Sample 1: torch.Size([64, 64])
Sample 2: torch.Size([64, 64])
Sample 3: torch.Size([64, 64])
Sample 4: torch.Size([64, 64])

Testing dataset: halfdome
Downloading https://phototour.cs.washington.edu/patches/halfdome.zip to ./datasets/halfdome.zip
100.0%
Number of patches in halfdome: 107776
Sample type for halfdome: <class 'torch.Tensor'>
Sample shape: torch.Size([64, 64])
Mean: 0.4718, Std: 0.1791
Sample 0: torch.Size([64, 64])
Sample 1: torch.Size([64, 64])
Sample 2: torch.Size([64, 64])
Sample 3: torch.Size([64, 64])
Sample 4: torch.Size([64, 64])

[venkatram-dev]

Pytest for download

pytest test/test_datasets_download.py -vvv -k phototour

                                                                       
test/test_datasets_download.py::test_url_is_accessible[PhotoTour, https://phototour.cs.washington.edu/patches/halfdome.zip] PASSED [ 33%]
test/test_datasets_download.py::test_url_is_accessible[PhotoTour, https://phototour.cs.washington.edu/patches/notredame.zip] PASSED [ 66%]
test/test_datasets_download.py::test_url_is_accessible[PhotoTour, https://phototour.cs.washington.edu/patches/trevi.zip] PASSED [100%]

[venkatram-dev]

pytest for dataset

pytest test/test_datasets.py -k PhotoTour


test/test_datasets.py ......s                                                                                          [100%]

======================================== 6 passed, 1 skipped, 570 deselected in 0.26s ========================================

[kavehsfv]

Thanks venkatram-dev

[ducha-aiki]

That is the different dataset version (earlier one), than it was in torchvision. Here is the one, which we had at torchvision https://cmp.felk.cvut.cz/~mishkdmy/datasets/BrownPhotoTour/

but the issue that I don't have original zips, only the images and labels themselves, so re-packed zips have different hashes.

[ducha-aiki]

Here is the PR with original data, not different scenes https://github.com/pytorch/vision/pull/9002 @NicolasHug