VOCDetection Dataset md5 has changed, no longer will load correctly

[teddykoker]

🐛 Bug

To Reproduce

Steps to reproduce the behavior:

>>>from torchvision.datasets import VOCDetection
>>>ds = VOCDetection(".")
RuntimeError: File not found or corrupted.

Can be fixed by correcting the md5sum:

from torchvision.datasets import VOCDetection
from torchvision.datasets.voc import DATASET_YEAR_DICT
DATASET_YEAR_DICT['2012']['md5'] = "af08689459fe018b209e8d8fb5d4eb2e"
ds = VOCDetection(".", download=True)

Expected behavior

Dataset should download and load without error.

Environment

PyTorch version: 1.6.0
Is debug build: False
CUDA used to build PyTorch: None

OS: Mac OSX 10.15.5 (x86_64)
GCC version: Could not collect
Clang version: 11.0.3 (clang-1103.0.32.62)
CMake version: Could not collect

Python version: 3.8 (64-bit runtime)
Is CUDA available: False
CUDA runtime version: No CUDA
GPU models and configuration: No CUDA
Nvidia driver version: No CUDA
cuDNN version: No CUDA

Versions of relevant libraries:
[conda] torch 1.6.0 pypi_0 pypi
[conda] torchvision 0.7.0 pypi_0 pypi

Additional context

[fmassa]

Hi,

Thanks for opening this issue (and sending a PR)!

I tried to download the VOCDetection dataset now but I couldn't get it to download (it was timing out), so I couldn't verify that the hash has changed yet.

Also, I wonder what happens with users that have downloaded the old tar file, will changing the md5 suddently break for them?

[teddykoker]

Presumably, this change would render the old dataset "corrupt." I believe @vfdev-5 reached out to the organizers to see why the change was made.

Ideally I think the best coarse of action would be to restore the old tar in the organizers website, or perhaps self host the dataset. I think allowing datasets to change could set a bad precedent as people depend on them for reproducible benchmarks (even though it appears to be a small modification of one image in this case.)

Edit: the organizers website (http://host.robots.ox.ac.uk/pascal/VOC/voc2012/) seems to be completely down, not sure whats going on.

[vfdev-5]

Also, I wonder what happens with users that have downloaded the old tar file, will changing the md5 suddently break for them?

@fmassa in my case I have unarchived dataset before data corruption (but no tar file) and I still could create an instance of VOCSegmentation dataset.

import torchvision

from torchvision.datasets import VOCSegmentation
from torchvision.datasets.voc import DATASET_YEAR_DICT

train_ds = VOCSegmentation("/home/data0/")
print(len(train_ds))
print(DATASET_YEAR_DICT["2012"])
> 1464
> {'url': 'http://host.robots.ox.ac.uk/pascal/VOC/voc2012/VOCtrainval_11-May-2012.tar', 'filename': 'VOCtrainval_11-May-2012.tar', 'md5': '6cd6e144f989b92b3379bac3b3de84fd', 'base_dir': 'VOCdevkit/VOC2012'}

[vfdev-5]

Looks like the correct tar with md5 6cd6e144f989b92b3379bac3b3de84fd is back !
@teddykoker could you please recheck from your side.
If everything is OK, I think we can close the issue and the PR. Anyway, @teddykoker thanks for pointing out that !

[teddykoker]

Just verified; the correct tar is back :). Closing the issue and PR. Thanks for the help anyway @vfdev-5 and @fmassa!