I am new to Machine Learning. My understanding is that data normalization before training, reduces complexity and potential errors during gradient decent. I have developed an SLP training model with Python/Tensorflow and have implemented the SLP trained model on micro using 'C' (not using TFLite). The model analyzes 48 features derived from an accelerometer and uses Sigmoid.
The question: if the model is trained/validated using normalized data (tf.keras.utils.normalize()), must the live data also be normalized before feeding into the trained model on micro?
The coefficient weights generated from training the model are understandably different when using normalized verses not normalized data, but the True/False results of running SLP on micro appear the same. If data normalization should be implemented on the micro, why? And if it is necessary, might it be possible to normalize the SLP weights instead?