Hello and thanks for a very interesting repo! The top-level README provides helpful step-by-step instructions for obtaining the data corrected in this repo.
For MNIST, the instructions include two variables with assignments that share the same right-hand side:
test_data = datasets.MNIST(data_dir, train=False, download=True).test_labels.numpy()
test_labels = datasets.MNIST(data_dir, train=False, download=True).test_labels.numpy()
In [6]: np.all(np.equal(test_data, test_labels))
Out[6]: True
(There are warnings from torchvision 0.13.0 about the names changing, but whichever torchvision version is supported by the step-by-step tutorial in the README, it would help to be consistent.)
bash$ sed 's!.*=!!' | while read rhs; do echo $rhs | openssl sha256; done
train_data = datasets.MNIST(data_dir, train=True, download=True).test_data.numpy()
train_labels = datasets.MNIST(data_dir, train=True, download=True).test_data.numpy()
870562877997826fd9627b9eb3890323171ea41841499caec4c8ea1ccddfeea4
870562877997826fd9627b9eb3890323171ea41841499caec4c8ea1ccddfeea4
bash$