Does anyone know if Tensorflow normalizes input data by default?
I have grayscale images that values range from roughly 20000-28000. When I normalized the data something odd happened. The network trained for a couple of hundred iterations was doing well in terms of making accurate predictions, but suddenly all predictions went to NaN. Of course, it couldn't recover because TF can't optimize from NaN.
When I didn't normalize the data training went fine and converged.