def get_co2_whole_duration_social_indicators(self, user_features):
with open('cfg/cnn_config.json') as f:
training_config = json.load(f)
train_x, train_y = utils.generate_time_series_df(
self.train_features, self.train_labels,
training_config['time_steps'])
test_x, test_y = utils.generate_time_series_df(
self.test_features, self.test_labels,
training_config['time_steps'])
tss = TimeSeriesSplit()
_, n_features = train_x.shape
model = DeepLearningModel(training_config,
num_features=n_features,
num_outputs=1)
for train_idx, test_idx in tss.split(train_x):
X, X_val = train_x.iloc[train_idx], train_x.iloc[test_idx]
Y, Y_val = train_y.iloc[train_idx], train_y.iloc[test_idx]
features, labels = utils.data_sequence_generator(
X, Y, training_config['time_steps'])
val_f, val_l = utils.data_sequence_generator(
X_val, Y_val, training_config['time_steps'])
model.train_with_validation_provided(features, labels, val_f,
val_l)
co2_pred = model.generate_future_prediction(user_features, test_x,
test_y, self.pred_steps)
co2_data_pred = np.array(co2_pred).reshape(-1, 1)
co2_total_duration = np.concatenate(
(self.co2_data_avlbl.to_numpy().reshape(-1, 1), co2_data_pred))
return co2_total_duration
print(train_labels.shape)
print(test_features.shape)
print(test_labels.shape)
# Train model with 5 fold cross validation
tss = TimeSeriesSplit()
_, n_features = train_features.shape
cnn = DeepLearningModel(training_config,
num_features=n_features,
num_outputs=1)
print(cnn.model.summary())
losses = []
start = time.time()
for train_idx, test_idx in tss.split(train_features):
X, X_val = train_features.iloc[train_idx], train_features.iloc[test_idx]
Y, Y_val = train_labels.iloc[train_idx], train_labels.iloc[test_idx]
features, labels = utils.data_sequence_generator(
X, Y, training_config['time_steps'])
val_f, val_l = utils.data_sequence_generator(X_val, Y_val,
training_config['time_steps'])
h = cnn.train_with_validation_provided(features, labels, val_f, val_l)
losses.append(h.history['loss'])
end = time.time()
print("TRAINING TIME: {}".format(end - start))
# Plot training loss
loss_arr = np.zeros((50, 1))
for loss_per_fold in losses:
for j, loss in enumerate(loss_per_fold):
loss_arr[j] = loss_arr[j] + loss
loss_arr = loss_arr / 5
fig1, ax1 = plt.subplots()
ax1.plot(range(len(loss_arr)), loss_arr, label='Training Loss')
# 4. Evaluate the trained model. Calculate prediction latency. Get prediction MSE, MAE and Soft Accuracy.
test_start = time.time_ns()
co2 = train_y.values.tolist() + test_y.values.tolist()
forecast = naive_forecast(co2, 1)
test_end = time.time_ns()
mse = np.mean(np.square(np.array(co2[1:]) - np.array(forecast)))
mae = np.mean(np.abs(np.array(co2[1:]) - np.array(forecast)))
soft_acc = soft_accuracy(np.array(co2[1:]), np.array(forecast))
print(
"------------------------------------- NAIVE FORECAST ------------------------------------"
)
print("TESTING TIME: {}".format(test_end - test_start))
print("\n\nTesting MSE: {}\nTesting Soft Accuracy: {}\nTesting MAE: {}".
format(mse, soft_acc, mae))
test_f, test_l = utils.data_sequence_generator(
features, labels, training_config_cnn['time_steps'])
test_start = time.time()
model_eval = dnn.model.evaluate(test_f, test_l)
test_end = time.time()
print(
"------------------------------------- DNN MODEL ------------------------------------"
)
print(dnn.model.summary())
print("\n\nPREDICTION LATENCY: {}s".format(test_end - test_start))
print("\n\nMSE: {}\nMAE: {}\nR2: {}\nSoft Accuracy: {}".format(
model_eval[0], model_eval[1], model_eval[2], model_eval[3]))
test_start = time.time()
model_eval = lstm.model.evaluate(test_f, test_l)
test_end = time.time()
print(