Keras LSTM网络预测更新

Question

I have been following along with various tutorials from https://machinelearningmastery.com/blog/ using Keras with great success. Recently I have been experimenting with deep LSTM networks for times series sequence prediction problems. The network is training well and producing an acceptable amount of error. In real world application, there is no data point to sub in to produce the next prediction like you can do with the test dataset, so I want to be able to sub in the last prediction.

For the real world application of this network, I need to be able to use my initial data to make prediction t+1, add the result to the input data, predict t+2, add the result to the input data, etc. until I am able to predict a particular amount of time steps into the future.

I initially started with:

start = len(testX)-15
pattern = testX[start]

for i in range(200):
    x = numpy.reshape(testX, (testX.shape[0], testX.shape[1], 1))
    prediction = model.predict(testX[i:i+15], batch_size=batch_size)
    pattern.append(prediction)

print "\nDone."

This method runs into issues with the pattern.append(prediction) line producing the error: AttributeError: 'numpy.ndarray' object has no attribute 'append'. This is not updating the inputs with the last prediction, thus breaking the continuous cycle.

I have not been able to find any ANN models that utilize this type of feed method to be able to predict out in time based on latest data.

Answer 1

Use numpy.append() to append values to the end of an array.

For your code, it should be pattern = numpy.append(pattern,prediction) .

Good luck.