auto_augment doesn't work in create_loader()?

[Kieran31]

Hello, this is my first time to use timm after reading Chris Hughes's long article. I'm confused about how to apply auto_augment when creating a data loader.

train_loader = create_loader(ImageDataset('train'), (3, 224, 224), batch_size=32, auto_augment="rand-m9-n3-mstd0.5")

n3 stands for 3 augmentations per image. However, next(iter(train_loader))[0].shape returns torch.Size([32, 3, 224, 224]). Why is it not [96, 3, 224, 224]?

[rwightman]

it's 3 augmentations to the same image

[Kieran31]

Could you explain a little more?
my train dataset has 4018 images. the batch size is 32. so len(train_loader) = 126 which is equal to ceil(4018/32). It looks no augmented images in the train_loader. Do you mean they will show up during the training process?

[rwightman]

n3 applies 3 layers of random augmentation to the same image producing one augmented output (so output will also be 32)

if you aren't seeing any augmentation and you are using the exact line as above, you need is_training=True active to apply train time augmentations as per the underlying transform creation fn https://github.com/rwightman/pytorch-image-models/blob/7cdd164d7775c444ee50afa6b825c843038f1b07/timm/data/transforms_factory.py#L167

[Kieran31]

Oh, I see.
If I want to keep the original images, so I need to set num_aug_splits = 2 in the function create_loader() and make train_ds an AugMixDataset?

train_ds = ImageDataset('train')
train_ds = AugMixDataset(train_ds, num_splits=2)
train_loader = create_loader(train_ds, (3, 224, 224), batch_size=32, is_training=True, auto_augment="rand-m9-n3-mstd0.5", num_aug_splits=2)

[rwightman]

@Kieran31 yes, that was the idea behind aug splits, the dataset wrapper + split transformers will run the same images through multiple different augmentations, with the first set being just normalized, and the remaining splits being different runs through random aug ... main purpose was to reproduce the AugMix paper, but it works with RandAugment as well.