如何在python中使用LSTM将预测值的输出反馈回输入

Question

The inputs here are the 3. The output here (LSTM) is the probabilities that the next x1 input ought to be. Means here I have x1x2 and x3 input values. 1st three inputs LSTM output1 and then next if x1 value = 0 then Lstm output1 is back into as input and predict next output2. If this output 2 value is equal to next x1 value then back into as input and predict output 3. If not equal not to take output 2 as x1 input and take it as mentioned x1 input. The output (Yt) at timestep t depends on the input X1t and on the previous output Yt-1. as a example

x1   x2    x3    predict (output)
100  30    40     120
 0   20    10     130
140  15    30     160

here second value of x1 column is 0, x1 value = 0 then take the value as output 1

x1 = output1

here x1 column 3rd value measured and output 2 value is not equal to x1 3rd value. then take input as measured value. So this is the method that I want to do. But I don't know how to write it. Can anyone helps me to do it? my LSTM code:

fit1 = Sequential ()
fit1.add(LSTM(32, return_sequences=True, activation='relu',input_shape=(3,1)))
fit1.add((LSTM(32, return_sequences=True)))
fit1.add(LSTM(32))
fit1.add(Dense(1))
batchsize = 3
fit1.compile(loss="mean_squared_error",optimizer="adam")
fit1.fit(x_train , y_train , batch_size = batchsize, nb_epoch =10,shuffle=True)
pred1=fit1.predict(x_test)
for i in range(len(x)):
    pred1.append(fit1.predict([x[i,None,:],pred1[i]]))
pred1 = np.asarray(pred1)
pred1 = pred1.reshape(pred1.shape[0],pred1.shape[2])

But this is not working. not having relationship inbetween input and output.

error:

After change the code:

for i in range(len(x)):
pred1 = np.append(pred1, fit1.predict([x[i,None,:],pred1[i]]))
pred1 = np.asarray(pred1)
pred1 = pred1.reshape(pred1.shape[0],pred1.shape[2])

error :

Another example: here I have three inputs data x1,x2,x3 data with time series. I want to predict value in every one hour. (at t+1) I have ay column that I got value in every one hour(t+1). But sometime i measured value after two hours. So in between past time and now time there is a value not measured at t+1. So I will predict value that I didn't measured at t+1. Here x1 value is depending on the output value(y) at t+1 . When I predict the value at that time period take it as y1(t+1), that value should have to read it as x1 value as x1(t) to predict the next output value at t+1 (y2). If I measured that value at t+1 ,if my prediction value == measured value then read it that value as x1(t) to predict the next value at t+1. This is the process that I want to write it as a code. here example in :

time     x1(t)     x2(t)     x3(t)     y(t+1)
6:00:00  120        0         0         110 (I measured it at t+1)

when I predict it using LSTM, if that value is == measured value(t+1) read it as a second input value of x1(t) column. if not equal read the measured value as second input value of x1(t). SO

7:00:00  110       40         10         0  (not measured value at t+1)

Then I predict the value at t+1 =y2 , assume it came 70 then that y2(t+1) value will be the third input of x1 column as x1(t). So

8:00:00  70        0          30         200 (I measured value at t+1)

this is the process that I want to run it using LSTM.

Answer 1

您应该使用pred1 = np.append(pred1, fit1.predict([x[i,None,:],pred1[i]]))而不是pred1.append(fit1.predict([x[i,None,:],pred1[i]])) ， pred1.append(fit1.predict([x[i,None,:],pred1[i]])) 。