def plot_multiple_xy_results(predictions,
y_test,
target_cols,
ind,
folder_name="nn"):
"""
Plotea lo resutltados de la red, cuando es con más de un input
Parameters
----------
predictions : array
predicciones.
real : array
valores reales.
names : list
nombre de las columans target.
folder_name : string, optional
directorio donde se dejaran las carpetas. The default is "nn".
Returns
-------
Plot.
"""
try_create_folder("results")
try_create_folder(f"results/{folder_name}")
for i in range(predictions.shape[1]):
print("Resultados:", target_cols[i], "...")
predi = predictions[:, i]
reali = y_test[:, i]
namei = target_cols[i]
plot_xy_results(predi,
reali,
index=ind,
name=namei,
folder_name=folder_name)
def plot_sequence(predictions, real, fechas, indice, folder_name="nn"):
"""
Plot sequence de la secuecnia
Parameters
----------
predictions : array
predicciones.
real : array
valores reales.
fechas : array
array de fechas.
indice : TYPE
indice de la columna.
Returns
-------
plot de prediciones vs real.
"""
letter_size = 20
new_fechas = []
for fecha in fechas:
fecha = fecha[0:10]
new_fechas.append(fecha)
plt.style.use('dark_background')
fig, ax = plt.subplots(1, figsize=(20, 12))
ax.plot(new_fechas, real, 'gold', linewidth=2)
ax.plot(new_fechas, predictions, 'orangered', linewidth=2)
ax.set_xlabel('Tiempo', fontname="Arial", fontsize=letter_size)
ax.set_ylabel('Predicción vs Real',
fontname="Arial",
fontsize=letter_size + 2)
ax.set_title(f"Predicciones vs real {str(indice)}",
fontname="Arial",
fontsize=letter_size + 10)
ax.legend(['real', 'predicción'],
loc='upper left',
prop={'size': letter_size + 5})
# Tamaño de los ejes
for tick in ax.get_xticklabels():
tick.set_fontsize(letter_size)
for tick in ax.get_yticklabels():
tick.set_fontsize(letter_size)
try_create_folder(f"results/{folder_name}")
plt.xticks(rotation=75)
plt.show()
fig.savefig(f"results/{folder_name}/{indice}_results.png")
def mae_evaluation(predictions,
real,
names,
nn,
folder_name="nn_architectures",
filename="algo"):
"""
Plotea lo resutltados de la red, cuando es con más de un input
Parameters
----------
predictions : array
predicciones.
real : array
valores reales.
names : list
nombre de las columans target.
names : model
tf model.
folder_name : string, optional
directorio donde se dejaran las carpetas. The default is "nn".
Returns
-------
Plot.
"""
try_create_folder("results")
try_create_folder(f"results/{folder_name}")
output = []
for i in range(predictions.shape[1]):
print("Resultados:", names[i], "...")
name = names[i]
predi = predictions[:, i]
reali = real[:, i]
mae = np.abs(predi - reali).mean()
mae = round(mae, 4)
output.append([name, mae, filename])
output = pd.DataFrame(output, columns=["variable", "mae", "architecture"])
output.to_csv(f"results/{folder_name}/{filename}.csv", index=False)
plot_model(nn,
to_file=f"results/{folder_name}/{filename}.png",
show_shapes=True,
show_layer_names=True)
return output
def training_history(history, model_name="NN", filename="NN"):
"""
Según el historial de entrenamiento que hubo plotear el historial
hacía atrás de las variables
Parameters
----------
history : list
lista con errores de validación y training.
model_name : string, optional
nombre del modelo. The default is "Celdas LSTM".
filename : string, optional
nombre del archivo. The default is "LSTM".
Returns
-------
None.
"""
size_training = len(history.history['val_loss'])
fig = plot_instance_training(
history, size_training, model_name,
filename + "_ultimas:" + str(size_training) + "epocas")
fig = plot_instance_training(
history, int(1.5 * size_training / 2), model_name,
filename + "_ultimas:" + str(1.5 * size_training / 2) + "epocas")
# guardar el resultado de entrenamiento de la lstm
print(os.getcwd())
try_create_folder("results")
fig.savefig(f"results/{model_name}_training.png")
fig = plot_instance_training(
history, int(size_training / 2), model_name,
filename + "_ultimas:" + str(size_training / 2) + "epocas")
fig = plot_instance_training(
history, int(size_training / 3), model_name,
filename + "_ultimas:" + str(size_training / 3) + "epocas")
fig = plot_instance_training(
history, int(size_training / 4), model_name,
filename + "_ultimas:" + str(size_training / 4) + "epocas")
print(fig)
validation_split=args.validation_size,
batch_size=args.batch_size,
epochs=args.epochs,
shuffle=False,
verbose=1,
callbacks=callbacks)
# ver resutados de entrenamiento
training_history(history, model_name="cnn", filename="cnn")
# hacer predicciones en el test
predictions = cnn.predict(x_test)
# gráficar los resultados
folder_name = file.replace(".csv", "").replace("featured", "cnn_corr")
# visualizaciones
plot_multiple_xy_results(predictions, y_test, target_cols, indice,
folder_name=folder_name)
# arquitectura usada en string
architecture = get_model_summary(cnn)
print(architecture)
# dataframe de salida
output = mae_evaluation(predictions, y_test, target_cols, cnn,
folder_name=folder_results,
filename=f"{indice}-{folder_name}")
print(output)
path_nn = "results/models"
try_create_folder(path_nn)
cnn.save(path_nn+f"/{indice}_model_cnn.h5")
# features
features = list(df.columns)
features.remove("year")
for col in target_cols:
features.remove(col)
# fechas
date = df[["year"]]
date.reset_index(drop=True, inplace=True)
# targets
targets = df[target_cols]
targets.reset_index(drop=True, inplace=True)
# crear folder featured
try_create_folder(path + "featured")
# eliminar por correlación las varibles y guardar los dataset para probarlos
for thres in range(10, 19):
thres = thres / 20
print("featured dataset para un threshold de:", thres)
# seleccionar solo las carácteristicas
dataset = df[features]
# eliminar por correlación
dataset = selection_by_correlation(dataset, threshold=thres)
dataset = pd.concat([dataset, date], axis=1)
dataset = pd.concat([dataset, targets], axis=1)
print("shape dataset:", dataset.shape)
# guardar el dataset eliminado por corr
thres = str(thres).replace(".", "_")
dataset.to_csv(path + "featured/" + f"featured_{thres}.csv", index=False)