PyTorch Lightning tests shouldn't hit network
bfirsh opened this issue · comments
Currently the PyTorch Lightning tests download mnist and train a real model. We shouldn't do that to just test that the callback works -- they should run a fake training process of some kind.
Maybe, some kind of fake data could be generated with NumPy or PyTorch instead of using a "real" dataset?
Yeah this doesn't need to do anything real.
These files are pretty small. Could just include them in the repo. Ideally we'd have a test that didn't actually do anything on real data though. https://github.com/golbin/TensorFlow-MNIST/tree/master/mnist/data
A pytext fixture with dummy generated train data could be nice, as it could be used in future test cases easily.