#Keras tip: you actually need to specify batch size again when you run model.predict() after model.fit(), otherwise it defaults to a batch size of 32. Depending on your model this could be too small, leading to incredibly slow inference. Common online examples do not seem to emphasize this at all.