Hello everyone. I have a structured csv dataset that contains mixed numerical - categorical and numeric_categorical features. 42 features. I used the tensorflow example. I have also a preprosessor pipe line using scikit-learn for normalizing numerical features and onehot encoding categorical features. The problem is that if server (initially) and each client want to have it’s own partition of dataset to train is own model, after preprocessing, the input shape of each model will be different from other client models. So in model.set_weights() it can not use weights of different shape and gives error. How should I structure my code that i can the same shapes in all models? I’m not much experienced in data science and would really be appreciated if you help.
I tried another idea. which is at every set_weight() call, before it , I get the self.model.get_weigths() and new weights and pass them to a function that will update the values of common indexes and keep others untouched in the list of weights . and then call set_weights() with edited list. This works until aggregation step in server happens and gives error that because of incompatible shapes it can not broadcast.