class DisplayCallback(tf.keras.callbacks.Callback):
def on_epoch_end(self, epoch, logs=None):
clear_output(wait=True)
show_predictions()
print('\nSample Prediction after epoch {}\n'.format(epoch + 1))
EPOCHS = 3
VAL_SUBSPLITS = 1
VALIDATION_STEPS = info.splits[
'test'].num_examples // BATCH_SIZE // VAL_SUBSPLITS
model_history = model.fit(train,
epochs=EPOCHS,
steps_per_epoch=STEPS_PER_EPOCH,
validation_steps=VALIDATION_STEPS,
validation_data=test,
callbacks=[DisplayCallback()])
loss = model_history.history['loss']
val_loss = model_history.history['val_loss']
epochs = range(EPOCHS)
plt.figure()
plt.plot(epochs, loss, 'r', label='Training loss')
plt.plot(epochs, val_loss, 'bo', label='Validation loss')
plt.title('Training and Validation Loss')
plt.xlabel('Epoch')
plt.ylabel('Loss Value')
plt.ylim([0, 1])
def build_model(data, config, output_length, X_test, input_length, index,
repeats):
print('Run:', str(index), '/', str(repeats))
# start time
start = time.time()
print('Configuration:', config)
# define parameters
epochs, batch_size, n_nodes1, n_nodes2, filter1, filter2, kernel_size = config
# convert data to supervised learning environment
X_train = np.stack([
sliding_window_input(data, input_length, i)
for i in range(len(data) - input_length)
])[:-output_length]
y_train = np.stack([
sliding_window_output(data['Upper Stillwater'], input_length,
output_length, i)
for i in range(len(data) - (input_length + output_length))
])
y_train = y_train.reshape(y_train.shape[0], y_train.shape[1], 1)
print('Feature Training Tensor (samples, timesteps, features):',
X_train.shape)
print('Target Training Tensor (samples, timesteps, features):',
y_train.shape)
# define parameters
n_timesteps, n_features, n_outputs = X_train.shape[1], X_train.shape[
2], y_train.shape[1]
# define residual CNN-LSTM
visible1 = Input(shape=(n_timesteps, n_features))
model = Conv1D(filter1, kernel_size, padding='causal')(visible1)
residual1 = ReLU()(model)
model = Conv1D(filter1, kernel_size, padding='causal')(residual1)
model = Add()([residual1, model])
model = ReLU()(model)
model = Conv1D(filter1, kernel_size, padding='causal')(model)
residual2 = ReLU()(model)
model = Conv1D(filter1, kernel_size, padding='causal')(residual2)
model = Add()([residual2, model])
model = ReLU()(model)
model = MaxPooling1D()(model)
model = Conv1D(filter2, kernel_size, padding='causal')(model)
residual3 = ReLU()(model)
model = Conv1D(filter2, kernel_size, padding='causal')(residual3)
model = Add()([residual3, model])
model = ReLU()(model)
model = Conv1D(filter2, kernel_size, padding='causal')(model)
residual4 = ReLU()(model)
model = Conv1D(filter2, kernel_size, padding='causal')(residual4)
model = Add()([residual4, model])
model = ReLU()(model)
model = MaxPooling1D()(model)
model = Conv1D(n_nodes1, kernel_size, padding='causal')(model)
residual5 = ReLU()(model)
model = Conv1D(n_nodes1, kernel_size, padding='causal')(residual5)
model = Add()([residual5, model])
model = ReLU()(model)
model = Conv1D(n_nodes1, kernel_size, padding='causal')(model)
residual6 = ReLU()(model)
model = Conv1D(n_nodes1, kernel_size, padding='causal')(residual6)
model = Add()([residual6, model])
model = ReLU()(model)
model = MaxPooling1D()(model)
model = Flatten()(model)
model = RepeatVector(n_outputs)(model)
model = LSTM(n_nodes1, activation='relu', return_sequences=True)(model)
model = LSTM(n_nodes1, activation='relu', return_sequences=True)(model)
model = LSTM(n_nodes1, activation='relu', return_sequences=True)(model)
model = LSTM(n_nodes1, activation='relu', return_sequences=True)(model)
dense = TimeDistributed(Dense(n_nodes2, activation='relu'))(model)
output = TimeDistributed(Dense(1))(dense)
model = Model(inputs=visible1, outputs=output)
model.compile(loss='mse', optimizer='adam')
#model.summary()
# fit network
es = EarlyStopping(monitor='loss', mode='min', patience=10)
history = model.fit(X_train,
y_train,
epochs=epochs,
batch_size=batch_size,
verbose=0,
validation_split=0.2,
callbacks=[es])
# test model against hold-out set
y_pred = model.predict(X_test)
print('\n Elapsed time:', round((time.time() - start) / 60, 3), 'minutes')
return y_pred, history
