get_best_model not working when adding normalization layer

[Deleted user]

Hi all,

I am trying to train a very simple neural network and tuning it to find the optimal number of units in the hidden layer. When there is no normalization layer, my code works fine but, when I add the normalizer and the normalization layer after the input tensor in the sequential model definition, the following error pops up:

ValueError: Received incompatible tensor with shape (1,) when attempting to restore variable with shape (10,) and name layer_with_weights-1/bias/.ATTRIBUTES/VARIABLE_VALUE.

Find my code below:

import numpy as np

# NN training
from sklearn.model_selection import train_test_split
import tensorflow.keras as tf
import keras_tuner 
import matplotlib.pyplot as plt


class neural_network:
    # Define attributes that are unique to each instance of the class
    def __init__(self, X, y, epochs):
        self.model = tf.Sequential()
        self.history = {} # History object
        self.X_all = X
        self.y_all = y
        self.epochs = epochs

        self.split_data(self.X_all, self.y_all)

        self.units = []
        self.weights = {}
        self.bias = {}

    def build_model(self,units,epochs):
        ''' Define model arquitecture'''
        normalizer = tf.layers.Normalization()
        normalizer.adapt(self.X_all)

        model = tf.Sequential()
        model.add(tf.Input(shape = (self.X_tr.shape[1],))) # Input tensor definition (not a layer!)
        model.add(normalizer)
        model.add(tf.layers.Dense(units, activation = 'relu')) # Hidden layers
        model.add(tf.layers.Dense(units = self.y_tr.shape[1], activation = 'linear')) # Output layer (no ReLU applied here, only weighted sum as per Neur2SP paper. ASK WHY)
        model.compile(loss='mean_squared_error', optimizer='adam', metrics=['mse'])
        model.summary()
        # model.fit(self.X_tr, self.y_tr, epochs=self.epochs, verbose = 0) # verbose: show progress bar or not
        # HOW TO CHOOSE METRICS? https://machinelearningmastery.com/custom-metrics-deep-learning-keras-python/
        return model
    
    def build_hypermodel(self, hp):
        ''' Define search space where tunning the hyperparameters'''
        # hp = keras_tuner.HyperParameters() . Not needed because the call in keras_tunner.RandomSearch already performs that action internally and passes it to the function specified in hypermodel attribute
        units = hp.Int("units", min_value = 5, max_value = 50, step = 5)
        model = self.build_model(units = units, epochs = self.epochs)

        return model
    
    def tune_model(self):
        '''Tune model using random search strategy'''
        tuner = keras_tuner.RandomSearch(
            hypermodel=self.build_hypermodel, # model-building function
            objective="val_mse", # objective to optimize (e.g., metric to be minimized)
        )

        # Add a callback to choose the optimal number of epochs
        # https://www.geeksforgeeks.org/choose-optimal-number-of-epochs-to-train-a-neural-network-in-keras/
        earlystopping = tf.callbacks.EarlyStopping(monitor="val_mse",
                                        mode="min", patience=2, # stop your training as soon as the validation loss stops improving. In this case, it will stop if it doesn't improve in 2 epochs
                                        restore_best_weights=True)

        # tuner.search_space_summary() # print summary of the search space
        tuner.search(self.X_tr, self.y_tr, self.epochs, verbose = 0, 
                    validation_data = (self.X_val, self.y_val),
                    callbacks = [earlystopping]) # search method takes the same arguments that would be passed to the fit method
        tuner.results_summary()
        
        # Get the best model
        best_model = tuner.get_best_models(num_models=1) # Returns a list object but best_model[0] returns a instance of the keras Sequential class
        best_model[0].summary() 
        # Get the best hyperparameters
        best_hps = tuner.get_best_hyperparameters(num_trials = 1)[0] 

To help with reproducibility, the shape of the input variables is (5000,7) and the output shape is (5000,1), being 5000 the total number of samples.

I am new to keras so any guidance on what might be triggering that error would be highly welcome.