Data Augmentation

1

I’m working on a CNN project and I’m trying to improve it. I was told me how I augment my data can greatly influence my results. I was told to look into inception style data augmentation, but I can’t find anything on. Does anyone know what that is and where I can find out about it, as well as in any other augmentation techniques/styles?

This is what I have so far:

train_transform = transforms.Compose([
transforms.Resize(224),
transforms.CenterCrop(224),
transforms.RandomRotation(10), # rotate +/- 10 degrees
transforms.RandomHorizontalFlip(), # reverse 50% of images
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406],
[0.229, 0.224, 0.225])
])

test_transform = transforms.Compose([
transforms.Resize(224),
transforms.CenterCrop(224),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406],
[0.229, 0.224, 0.225])
])

I am using the torchvison transform page on PyTorch, but are there any other augmentation styles or techniques anyone can point me to?

2

I guess you might be looking for RandomResizedCrop, which was used in Inception models.
The original paper might also mention other augmentation techniques.

Besides the torchvision transformations, you could also have a look at kornia.