import keras
from keras.layers import Dense
from keras.optimizers import Nadam
from keras.callbacks import ModelCheckpoint, EarlyStopping

import pandas as pd
import numpy as np

# Model parameters:
activation = 'relu'
final_activation = 'softmax'
loss = 'categorical_crossentropy'
batchsize = 100
epochs = 200
lr = 0.001

# Model architecture:
model = keras.Sequential()
model.add(Dense(units=8, input_shape=(8,), activation=activation))
model.add(Dense(units=8, activation=activation))
model.add(Dense(units=4, activation=activation))
model.add(Dense(units=4, activation=activation))
model.add(Dense(units=2, activation=final_activation))

model.compile(optimizer=Nadam(learning_rate=lr), loss=loss,
              metrics=['accuracy', 'AUC'])
model.summary()


# Model checkpoints:
saveModel = ModelCheckpoint('best_model.hdf5',
                            save_best_only=True,
                            monitor='val_loss',
                            mode='min')


# Model training:
model.fit(
            x_train,
            y_train,
            batch_size=batchsize,
            callbacks=[EarlyStopping(verbose=True, patience=20, monitor='val_loss'), saveModel],
            epochs=epochs,
            validation_data=(
                x_val,
                y_val),
            shuffle=True)