main.py

import tensorflow as tf
import os

from sklearn import datasets, preprocessing, model_selection

import landscapeviz


def build_model(units=10, l1=0, l2=0, act_fn="relu"):
    model = tf.keras.Sequential([
        tf.keras.layers.Dense(units, activation=act_fn, input_shape=(
            4,), kernel_regularizer=tf.keras.regularizers.l1_l2(l1=l1, l2=l2)),  # input shape required
        tf.keras.layers.Dense(units, activation=act_fn),
        tf.keras.layers.Dense(3, activation=tf.nn.softmax)
    ])

    model.compile("sgd", loss="sparse_categorical_crossentropy", metrics=[
                  'sparse_categorical_accuracy', 'categorical_hinge'])
    return model


def get_data(seed):
    data = datasets.load_iris()
    X_train, X_test, y_train, y_test = model_selection.train_test_split(
        data["data"], data["target"], test_size=0.25, random_state=seed)

    scaler_x = preprocessing.MinMaxScaler(feature_range=(-1, +1)).fit(X_train)
    X_train = scaler_x.transform(X_train)
    X_test = scaler_x.transform(X_test)

    return X_train, X_test, y_train, y_test


if __name__ == "__main__":

    seed = 42
    tf.random.set_seed(seed)
    X_train, X_test, y_train, y_test = get_data(seed)

    checkpoint = tf.keras.callbacks.ModelCheckpoint(
        "./weights/sgd.{epoch:02d}.hdf5", verbose=0, save_weights_only=True, period=1)

    model = build_model()
    model.fit(X_train, y_train, batch_size=32, epochs=60,
              verbose=0, callbacks=[checkpoint])
    landscapeviz.build_mesh(model, (X_train, y_train),
                            grid_length=40, verbose=0, seed=seed)
    landscapeviz.plot_contour(
        key="sparse_categorical_crossentropy", trajectory="./weights")
    landscapeviz.plot_3d(key="sparse_categorical_crossentropy", log=False)