Basic question about torchvision.transforms

[robd2]

Hello Everyone, basic question about torchvision.transforms.

If my training transform is defined as this:

train_transform = transforms.Compose([
transforms.RandomCrop(60),
transforms.RandomHorizontalFlip(),
transforms.ToTensor()
])

Does this mean that with every epoch I’m actually training with a different data set?

[Nikronic (Nikan Doosti)]

Hello
At every epoch, new transforms will be applied to your data.
Here is a full explanation about what is going on.

Data augmentation in PyTorch

Hello, in any epoch the dataloader will apply a fresh set of random operations “on the fly”. So instead of showing the exact same items at every epoch, you are showing a variant that has been changed in a different way. So after three epochs, you would have seen three random variants of each item in a dataset. That said: I don’t think your counting method works for estimating the number of samples in the augmented set: The flip will double the number of pictures, but the crop has many potentia…

Good luck

[robd2]

Excellent info @Nikronic. That matches what I see during runtime. Any thoughts why the PyTorch team took this approach vs a more static augmentation option? I’m thinking changing the training data on the fly so to say could have both positive and negative effects.

[Nikronic (Nikan Doosti)]

I do not know exactly why this approach is chosen, but it solved my problem even on huge datasets like Places with about 1.8 million images. I augmented it 6x bigger and everything seems fine!

Maybe if you explain your question more specifically, one of PyTorch developers answer you.

Good luck mate