Does Tensorflow normalize input by default?

Question

Does anyone know if Tensorflow normalizes the default input?

I have grayscale images whose values ​​range from approximately 20,000-28,000. When I normalized the data, something strange happened. The network, trained for a couple of hundred iterations, succeeded in making accurate predictions, but all of a sudden all of the forecasts went on NaN. Of course, he could not recover, because TF could not be optimized from NaN.

When I didn’t normalize, the data training went fine and converged.

Any ideas?

Answer 1

, . NaNs, , , .