MNIST from torchvision.datasets: are the test images part of the training dataset?

vision
1

Hi,

For a project on adversarial examples (github.com/maxdinech/mnist-attack) I use the MNIST database.

I decided to download the training.pt and test.pt files using torchvision.datasets.MNIST, then to save the first 50.000 values of training.pt as train.pt, and the last 10.000 as val.pt.

But when running performance tests on my trained networks, the accuracy over the test dataset is exactly the same as the accuracy over the val dataset. Are the last 10.000 images of training.pt the same as test.pt ?

If so, could this be changed ? It makes little sense to test the accuracy over already known values.

2

Sorry, I was wrong. An error in my code led me to have the same results on test and val. Hopefully it its not the case.