Improve ImageNet documentation

[cemde]

📚 The doc issue

The ImageNet documentation is quite short and skips a lot of steps on how to use it for the first time. A better description on how to use it would be nice.

As far as I understand

1. one has to download ...train.tar, ...val.tar and ...test.tar files
2. one has to download the task 1 and 2 devkit
3. one has to unpack these into train, val and test folders respectively.
4. one has to use some shell script to reorganise the folder structure into val/class_tag/image.jpeg. For val set or for all sets?

Suggest a potential alternative/fix

Add instructions:

• Visit https://www.image-net.org/ and download the 2012 train, test and val images for Task 1 and 2.
• In your /data/ImageNet folder create subfolders train, val and test.
• Unpack the image .tar files for each split.
• Copy the following [link here](link here) shell script into your train, val and test folders, respectively and run them. This will reorganize these folders to be compatible with torchvision.dataset.ImageFolder which bases the ImageNet class.

cc @pmeier

[pmeier]

Thanks @cemde for the suggestions. I agree documentation can be improved. However, I don't think you'll have to do any manual work after the download. As long as the archives you downloaded are in the root folder that you pass to the constructor of the dataset, you should be good. Before we hit the constructor of ImageFolder, we call

vision/torchvision/datasets/imagenet.py

Lines 57 to 65 in 928b05c

	def parse_archives(self) -> None:
	if not check_integrity(os.path.join(self.root, META_FILE)):
	parse_devkit_archive(self.root)
	if not os.path.isdir(self.split_folder):
	if self.split == "train":
	parse_train_archive(self.root)
	elif self.split == "val":
	parse_val_archive(self.root)

Meaning all the unpacking and restructuring will be handled for you.

So I guess, the documentation should be something along the lines of

Download the archives manually and place them in the root directory

Do you want to send a PR for that?

[MrinalTyagi]

I would be happy to contribute to this issue incase it is not assigned to anyone :)

[pmeier]

Go for it!

[puhuk]

@MrinalTyagi Are you going on this issue ? If not, can I take this :)

[pmeier]

So far no PR was send. Go ahead if you like.

[pmeier]

Fixed in #7503.