AutoKeras LSTM per estensimetro

Questo e' un tentativo di applicazione di AutoKeras (AutoML ovvero machine learning automatizzato basato su Keras) ai dati del post precedente

La differenza sostanziale e' che la dimensione del train set deve essere un multiplo intero della batch size  (in questo caso 42 e 4200)

# -*- coding: utf-8 -*-

"""Autokeras timeseries_forecaster

Automatically generated by Colaboratory.

Original file is located at

https://colab.research.google.com/drive/1mQHWTqwyRKhShtTrwNXBvPHgOkU4ghsX

"""

!pip install autokeras

import pandas as pd

import tensorflow as tf

import numpy as np

import matplotlib.pyplot as plt

import autokeras as ak

!rm detrend2.csv

!wget http://c1p81.altervista.org/detrend2.zip

!unzip detrend2.zip

dataset=pd.read_csv(r'detrend2.csv', sep=',')

# cancella la prima colonna del tempo

#dataset_c.drop('Data',axis=1,inplace=True)

dataset= dataset_c[1:6001]

dataset.head()

dataset.shape

val_split = int(len(dataset) * 0.7)

data_train = dataset[:val_split]

validation_data = dataset[val_split:]

data_x = data_train[

[

"Est.[mm]"

]

].astype("float64")

data_x_val = validation_data[

[

"Est.[mm]"

]

].astype("float64")

# Data with train data and the unseen data from subsequent time steps.

data_x_test = dataset[

[

"Est.[mm]"

]

].astype("float64")

data_y = data_train["Est.[mm]"].astype("float64")

data_y_val = validation_data["Est.[mm]"].astype("float64")

##print(data_x.shape)

#print(data_y.shape)

#print(data_y)

#la batch size deve essere un divisore

#della serie di train altrimenti genera errore

predict_from = 1

#lunghezza dei dati nel futuro

predict_until = 42

lookback = 30

clf = ak.TimeseriesForecaster(

lookback=lookback,

predict_from=predict_from,

predict_until=predict_until,

max_trials=4,

objective="val_loss",

)

clf.fit(

x=data_x,

y=data_y,

validation_data=(data_x_val, data_y_val),

batch_size=42,

epochs=20

)

# Predizione

predictions = clf.predict(data_x_test)

print(predictions.shape)

print(data_x_val.shape)

print(data_y_val.shape)

loss,acc = clf.evaluate(data_x_val, data_y_val, batch_size=42, verbose=1)

print('Accuracy: %.3f' % acc)

plt.plot(predictions)

plt.title("Estensimetro")

plt.xlabel('')

plt.ylabel('')

plt.show()

model = clf.export_model()

# summarize the loaded model

model.summary()

# save the best performing model to file

model.save('model_sonar.h5')