Torchvision datasets supported on dataset_factory

[fsolgui]

Hi, I have been toying with the dataset_factory.py script and i see that different torchvision datasets are already supported (CIFAR10, CIFAR100, places365, imagenet, etc.). I wonder if there is any reason not to include StanfordCars or Flowers102 or any other torchvision dataset for that reason.

Thanks in advance

[rwightman]

I forget why I didn't cover cars & flowers, there was a lack of consistency in some of the tv dataset wrappers, but if they are easy to add open to a PR... pretty sure you can also make those work via tfds (tensorflow datasets), and these hfds (huggingface datasets) might work out of the box (haven't tried but they look like they have the right columns):

• https://huggingface.co/datasets/dpdl-benchmark/oxford_flowers102
• https://huggingface.co/datasets/tanganke/stanford_cars

[rwightman]

Can use like python train.py --dataset hfds/dpdl-benchmark/oxford_flowers102 --num-classes 102 --amp --model resnet34d --opt adamw --lr 1e-4 --pretrained

[fsolgui]

Thanks. I had taken a look into those hfds datasets too. Will also take a look at the tfds datasets.