Multi Layer Perceptron

This is the simplest use case of DEATF.

In this instance, we require a simple DNN, a Multi Layer Perceptron (MLP). Only restrictions for the evolution have to be established, like maximum number of layers or neurons in the MLP. As is it the simple case, no evalution function has to be used, a predifined one is used (XEntropy). Fashion mnist dataset is used, that is why 28x28 is the input size and 10 the output size.

First of all fashion mnist dataset has to be loaded and preprocessed in order to pass it to the network. In that preprocessing, labels that are integers from 0 to 9, are one hot encoded. That is, 3 turns into a vector [0,0,0,1,0,0,0,0,0,0] with a one in the index 3 (starting to count from 0) and the rest are zeros.

    OHEnc = OneHotEncoder()

    y_train = OHEnc.fit_transform(np.reshape(y_train, (-1, 1))).toarray()
    y_test = OHEnc.fit_transform(np.reshape(y_test, (-1, 1))).toarray()
    y_val = OHEnc.fit_transform(np.reshape(y_val, (-1, 1))).toarray()
    

Then is time to star the evolution. First specifying the desired parameters for evolution and then calling evolve function, evolution will be carried out.

                 x_trains=[x_train], y_trains=[y_train], x_tests=[x_val], y_tests=[y_val], 
                 n_inputs=[[28, 28]], n_outputs=[[10]],
                 population=5, generations=5, batch_size=200, iters=50, 
                 lrate=0.1, cxp=0, mtp=1, seed=0,
                 max_num_layers=10, max_num_neurons=100, max_filter=4, max_stride=3,
                 evol_alg='mu_plus_lambda', sel='tournament', sel_kwargs={'tournsize':3}, 
                 evol_kwargs={}, batch_norm=False, dropout=False)
    
    a = e.evolve()