Starting to learn Keras after using “pure” Tensorflow, I was reading this guide for auto encoders: https://blog.keras.io/building-autoencoders-in-keras.html (in particular, the section “Let’s build the simplest possible autoencoder”).
If I use the code as shown on the page, it works fine and I am able to reproduce the results. But if I use
tensorflow.keras instead of the standalone
keras (simply by changing the
import lines), the training fails. The loss starts at about 0.69 and never deviates a lot; the output after 50 epochs looks like entirely random noise.
Can someone explain this to me? I thought that the functionality of standalone Keras and TF Keras is supposed to be the same.