It can be hard to prepare data when you’re just getting started with deep learning.

Long Short-Term Memory, or LSTM, recurrent neural networks expect three-dimensional input in the Keras Python deep learning library.

If you have a long sequence of thousands of observations in your time series data, you must split your time series into samples and then reshape it for your LSTM model.

In this tutorial, you will discover exactly how to prepare your univariate time series data for an LSTM model in Python with Keras.

Letâ€™s get started.

## How to Prepare Time Series Data

Perhaps the most common question I get is how to prepare time series data for supervised learning.

I have written a few posts on the topic, such as:

- How to Convert a Time Series to a Supervised Learning Problem in Python
- Time Series Forecasting as Supervised Learning

But, these posts don’t help everyone.

I recently got this email:

I have two columns in my data file with 5000 rows, column 1 is time (with 1 hour interval) and column 2 is bits/sec and I am trying to forecast bits/sec. In that case can you please help me to set sample, time step and feature [for LSTMs]?

There are few problems here:

- LSTMs expect 3D input, and it can be challenging to get your head around this the first time.
- LSTMs don’t like sequences of more than 200-400 time steps, so the data will need to be split into samples.

In this tutorial, we will use this question as the basis for showing one way to specifically prepare data for the LSTM network in Keras.

### Need help with Deep Learning for Time Series?

Take my free 7-day email crash course now (with sample code).

Click to sign-up and also get a free PDF Ebook version of the course.

## 1. Load the Data

I assume you know how to load the data as a Pandas Series or DataFrame.

If not, see these posts:

Here, we will mock loading by defining a new dataset in memory with 5,000 time steps.

1 2 3 4 5 6 7 8 9 10 |
from numpy import array # load... data = list() n = 5000 for i in range(n): data.append([i+1, (i+1)*10]) data = array(data) print(data[:5, :]) print(data.shape) |

Running this piece both prints the first 5 rows of data and the shape of the loaded data.

We can see we have 5,000 rows and 2 columns: a standard univariate time series dataset.

1 2 3 4 5 6 |
[[ 1 10] [ 2 20] [ 3 30] [ 4 40] [ 5 50]] (5000, 2) |

## 2. Drop Time

If your time series data is uniform over time and there is no missing values, we can drop the time column.

If not, you may want to look at imputing the missing values, resampling the data to a new time scale, or developing a model that can handle missing values. See posts like:

- How to Handle Missing Timesteps in Sequence Prediction Problems with Python
- How to Handle Missing Data with Python
- How To Resample and Interpolate Your Time Series Data With Python

Here, we just drop the first column:

1 2 3 |
# drop time data = data[:, 1] print(data.shape) |

Now we have an array of 5,000 values.

1 |
(5000,) |

## 3. Split Into Samples

LSTMs need to process samples where each sample is a single time series.

In this case, 5,000 time steps is too long; LSTMs work better with 200-to-400 time steps based on some papers I’ve read. Therefore, we need to split the 5,000 time steps into multiple shorter sub-sequences.

I write more about splitting up long sequences here:

- How to Handle Very Long Sequences with Long Short-Term Memory Recurrent Neural Networks
- How to Prepare Sequence Prediction for Truncated Backpropagation Through Time in Keras

There are many ways to do this, and you may want to explore some depending on your problem.

For example, perhaps you need overlapping sequences, perhaps non-overlapping is good but your model needs state across the sub-sequences and so on.

Here, we will split the 5,000 time steps into 25 sub-sequences of 200 time steps each. Rather than using NumPy or Python tricks, we will do this the old fashioned way so you can see what is going on.

1 2 3 4 5 6 7 8 9 |
# split into samples (e.g. 5000/200 = 25) samples = list() length = 200 # step over the 5,000 in jumps of 200 for i in range(0,n,length): # grab from i to i + 200 sample = data[i:i+length] samples.append(sample) print(len(samples)) |

We now have 25 sub sequences of 200 time steps each.

1 |
25 |

If you’d prefer to do this in a one liner, go for it. I’d love to see what you can come up with.

Post your approach in the comments below.

## 4. Reshape Subsequences

The LSTM needs data with the format of [samples, time steps and features].

Here, we have 25 samples, 200 time steps per sample, and 1 feature.

First, we need to convert our list of arrays into a 2D NumPy array of 25 x 200.

1 2 3 |
# convert list of arrays into 2d array data = array(samples) print(data.shape) |

Running this piece, you should see:

1 |
(25, 200) |

Next, we can use the *reshape()* function to add one additional dimension for our single feature.

1 2 3 4 |
# reshape into [samples, timesteps, features] # expect [25, 200, 1] data = data.reshape((len(samples), length, 1)) print(data.shape) |

And that is it.

The data can now be used as an input (X) to an LSTM model.

1 |
(25, 200, 1) |

## Further Reading

This section provides more resources on the topic if you are looking to go deeper.

### Related Posts

- How to Convert a Time Series to a Supervised Learning Problem in Python
- Time Series Forecasting as Supervised Learning
- How to Load and Explore Time Series Data in Python
- How To Load Machine Learning Data in Python
- How to Handle Missing Timesteps in Sequence Prediction Problems with Python
- How to Handle Missing Data with Python
- How To Resample and Interpolate Your Time Series Data With Python
- How to Handle Very Long Sequences with Long Short-Term Memory Recurrent Neural Networks
- How to Prepare Sequence Prediction for Truncated Backpropagation Through Time in Keras

### API

## Summary

In this tutorial, you discovered how to convert your long univariate time series data into a form that you can use to train an LSTM model in Python.

Did this post help? Do you have any questions?

Let me know in the comments below.

Great article! I wish I had this a couple months ago when I was struggling with doing the same thing for Tensorflow. Glad to see the solution I had mostly aligns with yours.

You mention some papers that discuss optimal sample size. Would you be able to share a link to those? I’m interested to see how the authors arrive at that number.

Thanks.

Perhaps check this post:

https://machinelearningmastery.com/much-training-data-required-machine-learning/

This publication helped me a lot! I really want to thank you for the post. Very simple and straight forward.

I’m happy to hear that!

Hi Jason, thx for sharing.

let say I have a timeseries dataset [1,2,3,4,5,6,7,8] and need to split it with time steps of 4, in your article, the result will be [1,2,3,4], [5,6,7,8]. But in some other articles I’ve read, the result sometime will be is this way: [1,2,3,4], [2,3,4,5],[3,4,5,6],[4,5,6,7],[5,6,7,8].

so what will be the best way to split the samples? thx.

All 3 approaches you have listed are valid, try each and see what works best for your problem.

Is there litterature on the subject? The 3 solutions seem to have a very distinct training time for large datasets. I assume that for the second solution we should keep the memory for the cell, but not for the third, right?

Also, is there a risk that the training overexposed certain timesteps(timestep 5 in the example) in early learning, giving a bigger weight to this data.

BTW great blog and your book on LSTM is the best I found on the subject. thx.

Not really.

I would suggest framing the problem each of the 3 ways and compare them to see what works best for your specific data.

Perhaps this post will help you with reframing the problem:

https://machinelearningmastery.com/reshape-input-data-long-short-term-memory-networks-keras/

When the original univariate time series gets split into a list of subsequences with length as m, with delay between each successive subsequence as d, this forms a new samples of with m dimension input vectors. This is called Takens embedding. When d = m = 4, this is the first case. When d = 1, m = 4, this is the 2nd case. As a matter of fact, any d > 1 is valid and the same goes for m. There are multiple methods available for determine “optimal” values of d and m. Here are some of publications on the subject:

https://arxiv.org/pdf/1605.01571.pdf

https://file.scirp.org/pdf/JMP_2017083015084865.pdf

Interesting, thanks for the refs.

Nice article. One thing I live about Python is list comprehension. One possible one-liner could be

samples = [data[i:i+length] for i in range(0,n, length)]

Nice, thanks.

Went, what you want is called “sliding window”, you could get it in the next code:

from itertools import islice

def window(seq, n=2):

“Returns a sliding window (of width n) over data from the iterable”

” s -> (s0,s1,…s[n-1]), (s1,s2,…,sn), … ”

it = iter(seq)

result = tuple(islice(it, n))

if len(result) == n:

yield result

for elem in it:

result = result[1:] + (elem,)

yield result

Hi Jason! First, I have to say that I really like your posts, they are very helpful.

I’m facing a time series classification problem (two classes) where I have series of around 120-200 time steps and 7 variables each. The problem is that I have only 3000 samples to train. What do you think, Is it feasible a priori to feed a LSTM network or I need more samples?

You mention that LSTM doesn’t work well with more than 200-400 timesteps. What about the number of features? Would you do dimensionality reduction?

Thank you very much in advance!

LSTMs can support multiple features.

It does not sound like enough data.

You could try splitting the sequence up into multiple subsequences to see if that helps?

Hi Jason,

Thank you for this excellent summary, your work is really impressive…I’m especially impressed by how many blog posts you have taken the time to write.

I was wondering why an LSTM network prefers a sequence of 200 – 400 samples, is this due to a memory allocation issue? Or can a longer sequence affect accuracy (I wouldn’t guess this but perhaps it’s possible)?

What role does the batch size play here? Couldn’t this restriction in sequence length be mitigated by selecting a correct batch size?

BR

Staffan

It seems to be a limitation on the training algorithm. I have seen this issue discussed in the literature, but have not pushed hard to better understand it myself.

I’d encourage you to test different configurations on your problem.

Hi jason,

Nice post! a little confused about the “time-steps” parameter. The “time-steps” means the steps span of input data? For example, for univariate problem,and one-step forecasting, i constructed the data with “sliding window”. For each sample,the structure is “t-6,t-5,t-4,t-3,t-2,t-1,t for input(train_x),and t+1 for output(train_y) ” .Using 7 data to forecast to the 8th. i reshaped the input(train_x) as [samples, 7,1]. Is that right?

Learn more about time steps in this post:

https://machinelearningmastery.com/gentle-introduction-backpropagation-time/

I think so.

Hello Jason, sorry for my english. I’m new to neural nework and i am trying to develop a neural network to generate music.

I have many .txt file with a sequence of notes like these

[int(note number), int(time), int(length)]

68 2357 159,

64 2357 260,

…

…

What kind of neural network I have to choose for this purpose?

How can i preprocess this kind of data?

Congratulations for this website and thank you.

For sequence prediction, perhaps RNNs like the LSTM would be a good place to start.

hi

I want to classify classes each class consists of 8_time steps in each time steps 16 features. is this reshape correct

reshape(124,8,1)

I think it would be (?, 6, 16) where “?” is the number of samples, perhaps 124 if I understand your case.

Hello, Jason, thanks for the great work.

I’ve read your articles about organizing the data for LSTM in 3D, but I can not do this with my data, it always presents an error like this:

“Error when checking target: expected dense_363 to have 2 dimensions, but got array with shape (3455, 1, 1)”

My data is organized as follows:

Appetizer:

11,000 lines with 48 columns, each row represents one day and each column represents 0.5h,

The output Y (0, 1) is binary, it represents the occurrence of an event 1 = yes, 0 = no.

So I have X = [0.1, 0.2, 0.3, …, 0.48] Y = [0] or Y = [1]

for more details see my code:

# load data

dataframe = pd.read_csv(‘Parque_A_Classificado_V2.csv’, header=None)

dataset = dataframe.values

# split data to variables train and test

train_size = int(len(dataset) * 0.7)

test_size = len(dataset) – train_size

trainX, trainY = dataset[0:train_size,:48], dataset[train_size:len(dataset),48]

testX, testY = dataset[0:test_size, :48], dataset[test_size:len(dataset), 48]

# reshape input to be [samples, time steps, features]

trainX = trainX.reshape(trainX.shape[0],trainX.shape[1], 1)

testX = testX.reshape(testX.shape[0], testX.shape[1], 1)

trainY = trainY.reshape(trainY.shape[0], 1, 1)

testY = testY.reshape(testY.shape[0], 1, 1)

#criando modelo

model = Sequential()

model.add(LSTM(100, input_shape=(48, 1)))

model.add(Dense(1, activation=’sigmoid’))

# Compile model

model.compile(loss=’binary_crossentropy’, optimizer=’adam’, metrics=[‘acc’])

model.fit(trainX, trainY, validation_data(testX, testY), epochs=1, batch_size=1)

I can not find the error, can you help me?

Maybe this post will make it clearer:

https://machinelearningmastery.com/reshape-input-data-long-short-term-memory-networks-keras/

Many thanks, Jason, your attitude is commendable.

This time I had to run my model.

Glad to hear you worked out your problem.

Hi Jason,

I’m struggling with a problem similar to those described here with a slight difference.

I’m solving a disaggregation problem and so my the dimensions of my output are higher than my input. in order to simplify lets say my original data looks something like this:

X.shape == [1000,1]

Y.shape == [1000,10]

I do some of the input to make things work:

X = X.reshape(1,X.shape[0[,X.shape[1]) #leaving this parameter dependent in case I want to

later use more features

My net looks like this:

model.sequential()

model.add(LSTM(50,batch_input_shape = X.shape, stateful = True)

model.add(Dense(Y.shape[1],activation = ‘relu’) #my output values aren’t between +/-1 so I

chose relu

went with a stateful model because I will most likely have to do batch seperation when running my actual training as I have close to a 10^6 samples

and then I’ve tried both doing the same thing to the Y vector and not touching it, either way I get error (when I reshaped Y I then changed Y.shape[1] to Y.shape[2])

Any thoughts?

Output will be 2D not 3D.

How can I split the 5000-row dataset into train and test portions when I am dividing into samples and reshaping it?

You could split before or after reshaping.

This post will teach you more about how to work with arrays:

https://machinelearningmastery.com/index-slice-reshape-numpy-arrays-machine-learning-python/

Thanks to this article and the one about reshaping input data for LSTM, I understood how to split/reshape the inputs of the LSTM network but I can’t see how to handle labels…

My dataset is 3000 time steps and 9 features. As explained in the article, I split it to get 15 samples of 200 time-steps so my input shape is (15, 200, 9).

My labels are a binary matrix (3000, 6) i.e. I want to solve a 6-class classification problem.

If I feed the labels as is, I’ll get an error “Found 15 input samples and 3000 target samples”.

How to correctly feed the labels to the network? What confuses me is that the targets should be 2D (unlike inputs) so I don’t see how I could split them in the same way as inputs, for example to get a (15, 200, 6) shape…

You will need one label per input sample.

Great blog thank you!

From what I understand you showed how to handle one long time series, but I couldn’t understand what to do with multiple inputs.

For example my input is x1 with dimensions (25, 200, 1)

but I have multiple inputs for my training X = [x1,x2…xn]

How should I shape for model.fit and for the LSTM layers? a 4D tensor?

I explain more here:

https://machinelearningmastery.com/faq/single-faq/how-do-i-prepare-my-data-for-an-lstm

Thank you for the wonderful blog.

Where does the total number of samples to train go in the reshape?

As I understood: (num of subsamples, time stamps, features per timestamp)

Correct: [samples, timesteps, features].

I love all your posts!

Im a bit confused:

I would guess that the number of time steps limits the number of recurrent layers. Since the number of time steps is equivalent to the amount of time steps you run your recurrent neural network. Is this true? If yes how can the memory of the LSTM be larger than the amount of recursions?

And if it isnt larger, why would anybody choose time steps = 1 like you did in some posts?

Thanks.

The time steps and the nodes/layers are unrelated.

Sorry, I fomulated my question badly.

I meant: if I have a sample sequence of lets say 100 time steps, can the memory of the LSTM be greater than these 100 time steps?

Is the memory limited by the amount of time steps given in a sequence?

Thanks for your time. T

The limit for the LSTM seems to be about about 200-400 time steps, from what I have read.

Hi Jason,

Can you please explain what you mean by LSTM does not work well for 200-400 time steps, while you replied to Daniel Salvador that 3000 training samples are not enough?

Does 200-400 mean 200-400 steps ahead prediction?

How many number of training samples you think is fairly enough?

The input data has a number of time steps. The LSTM performance appears to degrade if the number of input time steps is more than about 200-400 steps, according to the literature.

I have not tested this in experiments myself though.

Dear Jason,

Could you help me with this:

I have many phrases, and each phrases contains many words (I have padded so that they are of the same length), and I have trained word embedding for each word. So, in this case, if I want to use LSTM in keras to do some classification task (e.g. each phrase is labeled as 1 or 0, it’s related to the order of words), what will be my input shape for the LSTM layer in this case? Is it like shape (#of phrases, #of words in phrase, # of dimension of word embedding) ? I am a little confused here. Thanks for your help.

Probably: [total phrases, phrase length, 1]

Thanks for your reply. But I’m still confused here.

1. Why it is “1” at last?

2. I think the shape of my input numpy arrary (which will be thrown into Keras sequential model, whose first layer is a LSTM layer) is (#of phrases, #of words in phrase, # of dimension of word embedding). Does is mean that the input shape is (#of words in phrase, # of dimension of word embedding)? Because I want to learn something based on the sequence order between words.

My task is very similar to the task in one of your post. https://machinelearningmastery.com/sequence-classification-lstm-recurrent-neural-networks-python-keras/

In that post, the original input is a vector of words. Then, it will be put into a keras sequential model, however, the first layer is a Word Embedding layer, then followed by the LSTM layer. The output shape of word embedding layer should be a (2D) array, right?. Does that means the input shape of LSTM in this case is 2D rather than 3D? If it’s not, what will be the input shape in that case.

Thanks for your help.

Because you have a 1d sequence of integers to feed into your model, e.g. as input to the embedding.

The word embedding will handle its own dimensionality, don’t reshape data for it.

Hi Jason, could you help me on this?

My dataset has not been collected continuously, but it’s the result of many experiments, each one representing a specific class that I want my LSTM model to learn and predict.

Which is the best strategy to prepare the sequences for the training phase?

Should I concatenate all timeseries available and then use a sliding window to generate the sequences? in this case I may risk to have data of different classes in the same sequence…

Or would it be better to create the sequences separately for each individual class?

Thanks in advance

Perhaps brainstorm 3-5 different ways to frame the prediction problem, then prototype a few. This will help you clarify what the right approach might be.

Hi Jason,

Super post!

You did not do things like the following in your multivariate time seriies of PM 2.5 exmaple at https://machinelearningmastery.com/multivariate-time-series-forecasting-lstms-keras/ .

# split into samples (e.g. 5000/200 = 25)

samples = list()

length = 200

# step over the 5,000 in jumps of 200

for i in range(0,n,length):

# grab from i to i + 200

sample = data[i:i+length]

samples.append(sample)

print(len(samples))

Is it becasue that PM2.5 example assumes overlaping subsequences? Or would you have any other reasons?

For your convenience, you have the following snippets in that PM2.5 example:

# reshape input to be 3D [samples, timesteps, features]

train_X = train_X.reshape((train_X.shape[0], 1, train_X.shape[1]))

test_X = test_X.reshape((test_X.shape[0], 1, test_X.shape[1]))

print(train_X.shape, train_y.shape, test_X.shape, test_y.shape)

(8760, 1, 8) (8760,) (35039, 1, 8) (35039,)

After I change the n_hours=3, i.e., the timesteps, I have the following output in my Spyder:

(8760, 3, 8) (8760,) (35037, 3, 8) (35037,)

Train on 8760 samples, validate on 35037 samples

This means they are overlapping subsequences.

Please let me know if I get it right or not.

Many thanks.

Because I try to keep tutorials simple and focused.

Thanks for your reply. You did not elaborate in this tutorial on when one needs overlapping subsequences, when not. Would you have a tutorial about that, or any tips?

It really depends on the problem, e.g. the composition of the input samples is defined by what you want the model to learn or map to the output sample.

This is awesome (as is your entire series)! I consistently find your articles concise, clear and lucid, so thank you.

A small suggestion about the LSTM series however- you could add a couple of lines about the shaping of Y and the return sequence option. I struggled with it earlier, despite reading all your LSTM articles so it would probably help others!

The return sequence will be a one value per input time step, per node in the layer.

E.g. the layer gets 10 time steps of 1 variable and the layer has 100 nodes, you will get [100, 10, 1].

Would you provide example of the shape of label data? what should be the dimension? can we train on 24 samples and predict the 25th sample?

Here’s an example of making a prediction:

https://machinelearningmastery.com/make-predictions-long-short-term-memory-models-keras/

Hi,

I just wanna say you have awesome articles!

Here is my question:

Let’s say we split the data into a shape

(100, 60, 5)

Meaning 100 samples, each of them looking 60-time steps back and 5 features.

Would I be correct to assume that after we split the data as described, we could now shuffle the 100 samples as we wished and the result would be the same.

So we could apply normal cross-validation which is otherwise not possible with RNNs?

Thanks.

No, you cannot shuffle the samples. You must use walk-forward validation:

https://machinelearningmastery.com/backtest-machine-learning-models-time-series-forecasting/

Hi Jason! First, I have to say that I really like your posts, they are very helpful.

I have some questions about TimeSeries, would you give me some suggestion ?

1. Suppose the data: t1, t2,…t10, I prepare the data by rolling window, the window size is 3, such as [t1,t2,t3] -> t4, Then i trained a LSTM model, I want to know how to predict one time step in future ? for example: predict value on time t20 in future, but the histiry feature [t17, t18, t19] is null.

2. DO i need to prepare my data by rolling window if every timestep has a label? such as binary classify problem:

t1, f11, f12, f13, 1

t2, f21, f22, f23, 0

….

tn, fn1, fn2, fn3, 1

When i train LSTM, i reshape N time samples [N, 3] to [-1, timesteps, 3], N is number of time samples, shape of train data feed to LSTM is [-1, timesteps, 3], but this require N must equal to k * timesteps, for exampe, [60, 3] -> [-1, 12, 3] will be Ok, but [50, 3] -> [-1, 12, 3] will be wrong. I want to know how to process last 2 time sampes, should i pad zeros vector to get a sequence size 12 ?

Thank you very very much.

I have many example, you can get started here:

https://machinelearningmastery.com/start-here/#deep_learning_time_series

Hi, Jason. Thanks for your post.

I have some question. Suppose I have to forecasting the number of people in one region. We divide the region into 3×3 grid, each grid has the value of the current number of people. Then every one hour time interval there are 3×3 matrix, for example 8:00 there are 3×3 matrix, 9:00 there are 3×3 matrix, our goal is to use the previous two time interval (i.e. 10:00 and 11:00) to forecasting the next two interval (i.e. 10:00 and 11:00) numbers of people. How should I to deal this task. Thanks!

You can use a CNN-LSTM or ConvLSTM for read in a matrix time series, then use an encoder-decoder model to output multiple steps.