What's the advantage of torchvision.datasets.ImageNet?

[Hou9612]

Hi I use torchvision==0.2.2 now and I'm hesitant to update torchvision. I am worried that updating torchvision may bring some unexpected problems. so I want to know what's the advantage of torchvision.datasets.ImageNet comparing with torchvision.datasets.ImageFolder? Will it be faster when reading and processing ImageNet?

[surgan12]

Have a look here #713 specifically #713 (comment).

[pmeier]

By using the ImageNet dataset you will have the following advantages:

• After the download of the archives you don't need to manually prepare them for usage, i. e. you don't have to move the validation images in the appropriate folders
• The ImageNet object has the functionality of ImageFolder.classes and ImageFolder.class_to_idx implemented under ImageNet.wnids and ImageNet.wnid_to_idx. ImageNet.classes and ImageNet.class_to_idx have to same functionality but refer to the human-readable classes rather than the WordNet IDs.

Other than that the ImageNet dataset has no advantages over the ImageFolder. If you don't need or want the above properties you don't need to upgrade.

[Hou9612]

Have a look here #713 specifically #713 (comment).

Thanks, I'll have a look at it.

[Hou9612]

By using the ImageNet dataset you will have the following advantages:

• After the download of the archives you don't need to manually prepare them for usage, i. e. you don't have to move the validation images in the appropriate folders
• The ImageNet object has the functionality of ImageFolder.classes and ImageFolder.class_to_idx implemented under ImageNet.wnids and ImageNet.wnid_to_idx. ImageNet.classes and ImageNet.class_to_idx have to same functionality but refer to the human-readable classes rather than the WordNet IDs.

Other than that the ImageNet dataset has no advantages over the ImageFolder. If you don't need or want the above properties you don't need to upgrade.

Thanks, I think maybe I don't need to update it.

[fmassa]

Closing following @pmeier and @surgan12 comment