Did you notice that the prediction looks like a time shift of the original time series? Is it the doomed pattern of applying this network to a time series?

I m getting below error while executing the program. Please help me to solve it
/recurrent_keras_power.py: line 8: syntax error near unexpected token 1234' ./recurrent_keras_power.py: line 8: np.random.seed(1234)'

Thanks for making your LSTM time series code available; it reported MSE of 0.07. I tried to create a baseline for comparison, simply taking X_t as X_{t-1}. This also gives me MSE of 0.07. Is this odd, or maybe LSTM generalizes better, or my MSE computation is faulty?

Thanks,

import matplotlib.pyplot as plt
import pandas as pd
import numpy as np
import matplotlib.pyplot as plt

df = pd.read_csv('household_power_consumption.txt', sep=';')
df = df[['Global_active_power']]
df = df[df.Global_active_power != '?']
df['G2'] = df['Global_active_power'].shift(1)
print df.head()
df = df.astype(float)
df['err'] = df['G2']-df['Global_active_power']
df['err'] = np.power(df['err'],2)
print df.err.sum() / len(df)
#print np.sqrt(df['err'].sum()) / len(df)

I tried your code. I think it has a very low accuracy.
I try to explain

1. You have to predict the next one value (1 minute). Loss: 0.0854 - val_loss: 0.0721. The error is quite large.
   If you start predicting a few steps forward, based on previously predicted steps, it becomes obvious.
   
   Have you tried to predict a few steps?
2. Using a simple and fast multi-layer Perceptron provides a significant increase in accuracy and performance, for example "dense" with 100-200-1. I test it for my task.

@vict0rsch Could you please delete this issue completely. I will post from my other account the question. 😃

Thank you!

Thanks, very well explained for a beginner like me.

Minor note: I get

The show accuracy argument is deprecated, instead you should pass the accuracy metric to the model at compile time: model.compile(optimizer, loss, metrics=["accuracy"])

I'm going through this and finding a few typos and broken links. Fixing them and including in this PR.

Hi,your original post code is to use 49 dimension X to predict the 50th.how about I want to use 48 dimension to predict 49th and 50th. under such condition : does that mean i just change the output_dime of the last output layer :

model.add(Dense(
output_dim=2))
model.add(Activation("linear"))
Is that right?
Sorry for bothering you again..

Hi Sir,

Can you please help me in solving the below error in keras installation.

I want to install keras via anaconda in ubuntu.

UnsatisfiableError: The following specifications were found to be in conflict:

• keras -> python 2.7* -> openssl 1.0.1*
• python 3.6*
  Use "conda info " to see the dependencies for each package.

Just a few basic questions:

1. Why did you choose this particular network structure? Since you have two LSTM layers, would you get better performance from using the "relu" activation function?
2. Why are you only running 1 epoch? Is that just for testing purposes?
3. How are you measuring errors for forecasting, what is your chosen error metric? Most other results use RMSE, however, is there a better metric? RMSE feels like a necessary but not sufficient condition for measuring the "learning quality" of a neural net.

the tutorial's output is the next one value,but how to use the same time series data to predict the next n steps?
how to design the model to fit it?i have know the 'TimeDistributed()'can make the 3D output,but the output shape is the same as input.shape ,i can only change the output_dim ,so the problem is how to change the time_step params in the output?

Thanks,your tutorial and code very well explained LSTM for a beginner like me,however，when I run it on pycharm(based on python3.6,keras1.0.7),it shows as follows:
File "C:/Users/Guo/Desktop/household_power_consumption/predict.py", line 132, in
run_network()
File "C:/Users/Guo/Desktop/household_power_consumption/predict.py", line 105, in run_network
model = build_model()
File "C:/Users/Guo/Desktop/household_power_consumption/predict.py", line 67, in build_model
return_sequences=True))
TypeError: Expected int32, got <tf.Variable 'lstm_1_W_i:0' shape=(1, 50) dtype=float32_ref> of type 'Variable' instead.

I wonder if it's the version-compatibility,would you give me some advice?
really,thank you again

Could you please update your time-series Readme with a diagram of the architecture? You tried to explain a great detail about what return_sequences do but a simple diagram would be more helpful.

Cheers

Thanks for your examples and explanation regarding use of RNN with a LSTM.

I have a robotics application that takes as input an image and vector of imu measurements. I'm wondering if you have an idea about how to incorporate a time sequence of images, imu, into a network that uses conv2d layers and dense layers as input?

I used 1,000 samples to do a test of your example, but get the following error:
what does "too many indices for array" mean ?

Compilation Time :  0.00799989700317
Train on 418 samples, validate on 23 samples
Epoch 1/1
418/418 [==============================] - 228s - loss: 0.6930 - val_loss: 0.5179
too many indices for array
Training duration (s) :  279.796000004

Hi, the code runs well, thanks.
I got one question that confuse me , why you reshape the X into 3 dimensions : X_train = np.reshape(X_train, (X_train.shape[0], X_train.shape[1], 1)) ?does these represent: nb_samples , timestep and features ?
Many thanks

I would love to recreate that example with the same data that has been used.

However, the dataset is no longer available for download for me.
Is there any way to retrieve it, or can you upload it some place where i can download it?

Thanks :)

This page has a broken link to the recurrent tutorial. It should point here instead.