How to Use the TimeDistributed Layer for Long Short-Term Memory Networks in Python

Long Short-Term Networks or LSTMs are a popular and powerful type of Recurrent Neural Network, or RNN.

They can be quite difficult to configure and apply to arbitrary sequence prediction problems, even with well defined and “easy to use” interfaces like those provided in the Keras deep learning library in Python.

One reason for this difficulty in Keras is the use of the TimeDistributed wrapper layer and the need for some LSTM layers to return sequences rather than single values.

In this tutorial, you will discover different ways to configure LSTM networks for sequence prediction, the role that the TimeDistributed layer plays, and exactly how to use it.

After completing this tutorial, you will know:

  • How to design a one-to-one LSTM for sequence prediction.
  • How to design a many-to-one LSTM for sequence prediction without the TimeDistributed Layer.
  • How to design a many-to-many LSTM for sequence prediction with the TimeDistributed Layer.

Let’s get started.

How to Use the TimeDistributed Layer for Long Short-Term Memory Networks in Python

How to Use the TimeDistributed Layer for Long Short-Term Memory Networks in Python
Photo by jans canon, some rights reserved.

Tutorial Overview

This tutorial is divided into 5 parts; they are:

  1. TimeDistributed Layer
  2. Sequence Learning Problem
  3. One-to-One LSTM for Sequence Prediction
  4. Many-to-One LSTM for Sequence Prediction (without TimeDistributed)
  5. Many-to-Many LSTM for Sequence Prediction (with TimeDistributed)

Environment

This tutorial assumes a Python 2 or Python 3 development environment with SciPy, NumPy, and Pandas installed.

The tutorial also assumes scikit-learn and Keras v2.0+ are installed with either the Theano or TensorFlow backend.

For help setting up your Python environment, see the post:

Need help with LSTMs for Sequence Prediction?

Take my free 7-day email course and discover 6 different LSTM architectures (with sample code).

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

Start Your FREE Mini-Course Now!

TimeDistributed Layer

LSTMs are powerful, but hard to use and hard to configure, especially for beginners.

An added complication is the TimeDistributed Layer (and the former TimeDistributedDense layer) that is cryptically described as a layer wrapper:

This wrapper allows us to apply a layer to every temporal slice of an input.

How and when are you supposed to use this wrapper with LSTMs?

The confusion is compounded when you search through discussions about the wrapper layer on the Keras GitHub issues and StackOverflow.

For example, in the issue “When and How to use TimeDistributedDense,” fchollet (Keras’ author) explains:

TimeDistributedDense applies a same Dense (fully-connected) operation to every timestep of a 3D tensor.

This makes perfect sense if you already understand what the TimeDistributed layer is for and when to use it, but is no help at all to a beginner.

This tutorial aims to clear up confusion around using the TimeDistributed wrapper with LSTMs with worked examples that you can inspect, run, and play with to help your concrete understanding.

Sequence Learning Problem

We will use a simple sequence learning problem to demonstrate the TimeDistributed layer.

In this problem, the sequence [0.0, 0.2, 0.4, 0.6, 0.8] will be given as input one item at a time and must be in turn returned as output, one item at a time.

Think of it as learning a simple echo program. We give 0.0 as input, we expect to see 0.0 as output, repeated for each item in the sequence.

We can generate this sequence directly as follows:

Running this example prints the generated sequence:

The example is configurable and you can play with longer/shorter sequences yourself later if you like. Let me know about your results in the comments.

One-to-One LSTM for Sequence Prediction

Before we dive in, it is important to show that this sequence learning problem can be learned piecewise.

That is, we can reframe the problem into a dataset of input-output pairs for each item in the sequence. Given 0, the network should output 0, given 0.2, the network must output 0.2, and so on.

This is the simplest formulation of the problem and requires the sequence to be split into input-output pairs and for the sequence to be predicted one step at a time and gathered outside of the network.

The input-output pairs are as follows:

The input for LSTMs must be three dimensional. We can reshape the 2D sequence into a 3D sequence with 5 samples, 1 time step, and 1 feature. We will define the output as 5 samples with 1 feature.

We will define the network model as having 1 input with 1 time step. The first hidden layer will be an LSTM with 5 units. The output layer with be a fully-connected layer with 1 output.

The model will be fit with efficient ADAM optimization algorithm and the mean squared error loss function.

The batch size was set to the number of samples in the epoch to avoid having to make the LSTM stateful and manage state resets manually, although this could just as easily be done in order to update weights after each sample is shown to the network.

The complete code listing is provided below:

Running the example first prints the structure of the configured network.

We can see that the LSTM layer has 140 parameters. This is calculated based on the number of inputs (1) and the number of outputs (5 for the 5 units in the hidden layer), as follows:

We can also see that the fully connected layer only has 6 parameters for the number of inputs (5 for the 5 inputs from the previous layer), number of outputs (1 for the 1 neuron in the layer), and the bias.

The network correctly learns the prediction problem.

Many-to-One LSTM for Sequence Prediction (without  TimeDistributed)

In this section, we develop an LSTM to output the sequence all at once, although without the TimeDistributed wrapper layer.

The input for LSTMs must be three dimensional. We can reshape the 2D sequence into a 3D sequence with 1 sample, 5 time steps, and 1 feature. We will define the output as 1 sample with 5 features.

Immediately, you can see that the problem definition must be slightly adjusted to support a network for sequence prediction without a TimeDistributed wrapper. Specifically, output one vector rather build out an output sequence one step at a time. The difference may sound subtle, but it is important to understanding the role of the TimeDistributed wrapper.

We will define the model as having one input with 5 time steps. The first hidden layer will be an LSTM with 5 units. The output layer is afully-connectedd layer with 5 neurons.

Next, we fit the model for only 500 epochs and a batch size of 1 for the single sample in the training dataset.

Putting this all together, the complete code listing is provided below.

Running the example first prints a summary of the configured network.

We can see that the LSTM layer has 140 parameters as in the previous section.

The LSTM units have been crippled and will each output a single value, providing a vector of 5 values as inputs to the fully connected layer. The time dimension or sequence information has been thrown away and collapsed into a vector of 5 values.

We can see that the fully connected output layer has 5 inputs and is expected to output 5 values. We can account for the 30 weights to be learned as follows:

The summary of the network is reported as follows:

The model is fit, printing loss information before finalizing and printing the predicted sequence.

The sequence is reproduced correctly, but as a single piece rather than stepwise through the input data. We may have used a Dense layer as the first hidden layer instead of LSTMs as this usage of LSTMs does not take much advantage of their full capability for sequence learning and processing.

Many-to-Many LSTM for Sequence Prediction (with TimeDistributed)

In this section, we will use the TimeDistributed layer to process the output from the LSTM hidden layer.

There are two key points to remember when using the TimeDistributed wrapper layer:

  • The input must be (at least) 3D. This often means that you will need to configure your last LSTM layer prior to your TimeDistributed wrapped Dense layer to return sequences (e.g. set the “return_sequences” argument to “True”).
  • The output will be 3D. This means that if your TimeDistributed wrapped Dense layer is your output layer and you are predicting a sequence, you will need to resize your y array into a 3D vector.

We can define the shape of the output as having 1 sample, 5 time steps, and 1 feature, just like the input sequence, as follows:

We can define the LSTM hidden layer to return sequences rather than single values by setting the “return_sequences” argument to true.

This has the effect of each LSTM unit returning a sequence of 5 outputs, one for each time step in the input data, instead of single output value as in the previous example.

We also can use the TimeDistributed on the output layer to wrap a fully connected Dense layer with a single output.

The single output value in the output layer is key. It highlights that we intend to output one time step from the sequence for each time step in the input. It just so happens that we will process 5 time steps of the input sequence at a time.

The TimeDistributed achieves this trick by applying the same Dense layer (same weights) to the LSTMs outputs for one time step at a time. In this way, the output layer only needs one connection to each LSTM unit (plus one bias).

For this reason, the number of training epochs needs to be increased to account for the smaller network capacity. I doubled it from 500 to 1000 to match the first one-to-one example.

Putting this together, the full code listing is provided below.

Running the example, we can see the structure of the configured network.

We can see that as in the previous example, we have 140 parameters in the LSTM hidden layer.

The fully connected output layer is a very different story. In fact, it matches the one-to-one example exactly. One neuron that has one weight for each LSTM unit in the previous layer, plus one for the bias input.

This does two important things:

  • Allows the problem to be framed and learned as it was defined, that is one input to one output, keeping the internal process for each time step separate.
  • Simplifies the network by requiring far fewer weights such that only one time step is processed at a time.

The one simpler fully connected layer is applied to each time step in the sequence provided from the previous layer to build up the output sequence.

Again, the network learns the sequence.

We can think of the framing of the problem with time steps and a TimeDistributed layer as a more compact way of implementing the one-to-one network in the first example. It may even be more efficient (space or time wise) at a larger scale.

Further Reading

Below are some resources and discussions on the TimeDistributed layer you may like to dive in into.

Summary

In this tutorial, you discovered how to develop LSTM networks for sequence prediction and the role of the TimeDistributed layer.

Specifically, you learned:

  • How to design a one-to-one LSTM for sequence prediction.
  • How to design a many-to-one LSTM for sequence prediction without the TimeDistributed Layer.
  • How to design a many-to-many LSTM for sequence prediction with the TimeDistributed Layer.

Do you have any questions?
Ask your questions in the comments below and I will do my best to answer them.

Develop LSTMs for Sequence Prediction Today!

Long Short-Term Memory Networks with Python

Develop Your Own LSTM models in Minutes

…with just a few lines of python code

Discover how in my new Ebook:
Long Short-Term Memory Networks with Python

It provides self-study tutorials on topics like:
CNN LSTMs, Encoder-Decoder LSTMs, generative models, data preparation, making predictions and much more…

Finally Bring LSTM Recurrent Neural Networks to
Your Sequence Predictions Projects

Skip the Academics. Just Results.

Click to learn more.


39 Responses to How to Use the TimeDistributed Layer for Long Short-Term Memory Networks in Python

  1. Birkey May 17, 2017 at 7:09 pm #

    Hi, Jason, nice article on TimeDistributed layer!

    Basically, there’re three configurations for X (and thus y):
    1. (5,1,1) – 5 batchs, 1 time step, 1 feature/step – result shape (5,1)
    2. (1,5,1) – 1 batch, 5 time steps, 1 feature/step – result shape (1,5)
    3. (1,1,5) – 1 batch, 1 time step, 5 features/step

    in article, you discussed previous 2 configures.
    I did experiment of config 3, result same shape (1, 5) as 2 does, ’cause X input only 1 batch (which contains 1 sample, which has 5 features.) this config surely lost time information.

    3 differ from 2 in two ways:
    1) how we/model frame the problem: sequence should be framed as multi time steps as 2
    2) different number of LSTM params: config 2 has 140, while config 3 has 220! (big input vector)

    Q:
    in section ‘many to one without TimeDistributed’, with config 2, you said “The time dimension or sequence information has been thrown away and collapsed into a vector of 5 values.” — that surprise me a little bit.
    – does that mean, for seq-to-seq problem, we should always use TimeDistributed?
    – what situation suites config 2 (samples, multi-time-steps, features)?

    • Birkey May 17, 2017 at 7:14 pm #

      I guess for sequence-to-vector problem (predict one target one time step), config 2 is fine. But for sequence-to-sequence problem discussed here, config 2 is not the right choice, go TimeDistributed.

    • Jason Brownlee May 18, 2017 at 8:34 am #

      Very nice, yes I agree.

      Generally, we must model sequences as time steps. BPTT will use the sequence data to estimate the gradient. LSTMs have memory, but we cannot rely on them to remember everything (e.g. sequence length of 1).

      We can configure an MLP or LSTM to output a vector. For an LSTM, if we output a vector of n values for one time step, each output is considered by the LSTM as a feature, not a time step. Thus it is a many-to-one architecture. The vector may contain timesteps, but the LSTM is not outputting time steps, it is outputting features.

      This is no more or less valid, it may require more weights and may give better or worse performance.

      Does that make sense?

  2. Victor Garcia Cazorla May 18, 2017 at 9:48 am #

    Any recommendations when facing a one-to-many problem?

    • Jason Brownlee May 19, 2017 at 8:09 am #

      They often need more training than you think and consider using bidirectional inputs and regularization on input connections.

  3. Phil Ayres July 6, 2017 at 8:19 pm #

    This post is great! Thanks for being about the only person to actually explain simply what the TimeDistributed wrapper is doing.

    I tried it out with audio vocal data to attempt generation of new speech. I’d previously got basic results with a plain Dense layer on the output.

    With the TimeDistributed the network of lstms learned fast. But the result was just to return a rough version of the seed data inputted during generation. This appears to be modelling the equality function, when what I expected was something resembling the sequence following the seed.

    My X input is an array of batches, timesteps, and vocal properties. Just a longer version of your example. My y output for measuring error is effectively the same data, just one timestamp later for each batch (time sequence).

    Since your examples are for equality modelling, it’s hard to tell if I’ve missed a concept. Any thoughts on why this seems to generate equality rather than next timesteps, from my basic description?

    By the way, my original project without TimeDistributed is found at http://babble-rnn.consected.com in case you’re interested in extra context.

    • Jason Brownlee July 9, 2017 at 10:29 am #

      Perhaps you need to fit for longer or require more training data?

      • Phil Ayres July 10, 2017 at 9:01 pm #

        I wondered about that. I think my mistake may be simple…

        Imagine the sequence I was trying to learn was 1,2,3,4,5,6,7,8 (which I’d normalise in the range 0:1). In the standard Keras LSTM example without TimeDistributed I’d have:

        input X[0] = [0,1,2]
        output y[0] = [3]
        X[1] = [1,2,3]
        y[1] = [4]

        So in the TimeDistributed setup I reported above, I tried:

        X[0] = [0,1,2]
        y[0] = [1,2,3]
        X[1] = [1,2,3]
        y[1] = [2,3,4]

        In other words, I was offsetting the intended output by just a single timestep for each batch to be learned.

        But I’m guessing that I should really offset the output to be learned by the full number of timesteps in each batch:

        X[0] = [0,1,2]
        y[0] = [3,4,5]
        X[1] = [1,2,3]
        y[1] = [4,5,6]

        Is the latter example what I should be doing? Intuitively, this would explain why I was learning something close to equality in my first run. But from multiple readings of your code in the post it is not clear to me that this is the case.

        • Jason Brownlee July 11, 2017 at 10:32 am #

          I’m not sure I follow. There are indeed many ways to frame a sequence prediction problem.

          The simplest framing is sequence in => sequence out where either in or out could be one or more time steps.

          Keep one sequence as one sample if possible.

          • Phil Ayres July 12, 2017 at 12:10 am #

            Hmm, I think I have missed something big here. Please humour and let me try once more.

            In the standard LSTM examples on Keras, if I was to learn a long time sequence (for example integers incrementing in the range 1..100000), I would pick a shorter segment of the total sequence to pass to the LSTM (I split my corpus into sub-batches that represent the number of LSTM timesteps), then the output to learn would be just the next item in the sequence. There is no TimeDistributed output, so I get one result to calculate error against.

            input set: 1,2,3
            desired output: 4

            then repeat with other sub-batches in the same way (and Keras scrambles the order), so the next one may be…

            input set: 473, 474, 475
            desired output: 476

            If that makes sense, then allow me to ask simply what the input and output should be for the TimeDistributed setup. Would it be

            (option A)
            input set: 1, 2, 3
            desired output: 2, 3, 4

            (option B)
            input set: 1, 2, 3
            desired output: 4, 5, 6

            (option C)
            something else entirely.

            Am I making more sense now?

            Your example shows input set and desired output being the same, which says to me that the net will just learn the equality function. Again, am I missing something?

            Thanks again for your help.

          • Jason Brownlee July 12, 2017 at 9:47 am #

            Yes, good question.

            Option B.

            In your first example you have a many-to-one time step predictive model. In option B you have a many-to-many time step predictive model. The TimeDistributed wrapper would allow you to use the same Dense layer to output each time step in the output sequence, in this case one output time step per input time step.

            I hope that helps.

  4. zceewhc July 7, 2017 at 10:57 pm #

    Hi Jason, thanks for the great article!

    Having looked through this article and forums online, is it correct to say that if we were to do many-to-one prediction (an input vector with an output value), it will be straightforward and faster to just use the Dense layer?

    In the case where we want to do many-to-many predictions (multiple input vectors/matrices with an output vector/matrix), TimeDistributed layer should be used instead?

    • Jason Brownlee July 9, 2017 at 10:46 am #

      Yes, but in the latter case the dense is wrapped in the timedistributed.

    • Phil Ayres July 12, 2017 at 5:59 pm #

      That does, thank you! For some reason I couldn’t get that from your post, so thanks for taking the time to explain in more detail.

  5. Alexandr Pavel July 12, 2017 at 10:13 am #

    Thanks, that cleared up return_sequences for me, but I still don’t fully understand what TimeDistributed does.

    In the last example (Many-to-Many): If I change TimeDistributed(Dense(1)) to just Dense(1), neither the output shape nor the number of parameters changes and it works just as well. What is the difference between these two options in this case?

    • Jason Brownlee July 13, 2017 at 9:45 am #

      Note the number of weights in the network.

      Without the TimeDistributed wrapper, the Dense is connected to the output from each time step. With the wrapper, the same Dense is applied to each time step.

      It’s a question of how you want to model the problem. Let the Dense combine the time steps and output a vector or process each time step one at a time.

      Does that help?

      • Varuna Bamunusinghe August 31, 2017 at 2:20 am #

        Thanks for the article. I have the same question though… number of weights are same regardless of Dense is wrapped by TimeDistributed or not. So, what is the difference, and where can I see that?

        • Jason Brownlee August 31, 2017 at 6:22 am #

          But we have the same number of weights in a many-to-many model as we did no the one-to-one model.

          A better model design for increased model complexity/capability with the same resources.

          Does that help?

  6. chintan zaveri July 22, 2017 at 4:04 am #

    Hey,
    Thanks for amazing tutorial.
    This shows simple echo program implementation right ?

    I want something like –
    Input(For time period 2012-2013) – 1,2,5,3,6,4,7,8,9,5
    output(For Time Period 2014) – 1,3,4

    The output sequence should be generated based on the input sequence, kindly guide me on that.

    • Jason Brownlee July 22, 2017 at 8:37 am #

      Sounds great.

      What is the issue exactly? You can use code from the blog directly and adapt it for your problem. Where are you having trouble?

  7. ISR July 22, 2017 at 4:24 am #

    How to create a pyramid of LSTMs. i.e. the input to the first node of 2nd layer LSTM will be output at t1 and t2 of first layer LSTM, similrly 2nd node of the 2nd layer will use t3 and t4 from first layer, and so on..

    • Jason Brownlee July 22, 2017 at 8:38 am #

      You mean a stacked LSTM?

    • Sri Harsha Gangisetty August 23, 2017 at 6:45 pm #

      Hmm, that’s an interesting layer configuration, I would go with Tensorflow module directly instead of Keras to create such a model, Keras doesn’t have that functionality I guess.

  8. Ilja August 8, 2017 at 10:17 am #

    Hello! Is the a way to have DIFFERENT length of input and output-timesteps?
    Like, I have series with 100 timesteps in the past and will learn next 10 in the feature?
    TimeDistributed requires equal length.
    If I output return_sequence=false in the last LSTM and Dense with 10 neurones, would it be the same?
    Thanks You!

    • Jason Brownlee August 8, 2017 at 5:10 pm #

      Sorry, I’m not sure I follow, can you restate your question?

      Generally, different numbers of times steps on the input and output are referred to as seq2seq problems and are perhaps best addressed with an encoder-decoder network.

  9. james August 30, 2017 at 9:44 am #

    Is the procedure similar when using SimpleRNN?

  10. Darshan Bagul September 15, 2017 at 4:48 am #

    Hello Jason,

    Nice article. I was wondering if TimeDistributed layer in Keras is analogous to sequence-to-sequence learning module in Tensorflow. If not, could you point out the distinction between the two?
    Thanks.

    • Jason Brownlee September 15, 2017 at 12:17 pm #

      Sorry, I cannot draw this comparison for you as I am not deeply familiar with the TF code.

  11. Eldar M. September 17, 2017 at 8:42 am #

    Hey Jason. I hope I’m understanding this correctly.

    I was trying to model a certain number of days ahead, and found myself frustrated with the fact that I couldn’t just predict one day ahead, then right away use that as part of the sliding input window prior to weights being adjusted – basically I wanted the sliding window to move n days forward using predicted values and only then have gradient descent update weights.

    I think this might be the way to do so, but am unsure if I need to wrap every layer in timedistributed or what exactly to do with that.

    • Jason Brownlee September 18, 2017 at 5:41 am #

      You can do this, but you will need to create the sliding window yourself and call your model recursively.

      Keras will not do this for you with the TimeDistributed layer.

  12. Abuzar September 19, 2017 at 2:02 am #

    Hi Jason,

    Thanks for this post,

    I have an Input sequence and output sequence shape as follows:
    X_shape: (1, 82600, 1)
    Y_shape: (1, 82600, 1)

    When I try to use your code for this input and output I get following error:
    —————————————————————————
    MemoryError Traceback (most recent call last)
    in ()
    60 # create LSTM
    61 model = Sequential()
    —> 62 model.add(LSTM(n_neurons, input_shape=(length, 1), return_sequences=True))
    63 model.add(TimeDistributed(Dense(1)))

    How can I go around this?

    • Abuzar September 19, 2017 at 2:22 am #

      Since my length was 82600, according to the code, nb_neurons = 82600

      I just reduced the number of neurons to 8260 and the compilation was successful

      Since, by default this model is stateless (stateful = True) is not specifically specified, do you think reducing the number of neurons was a right choice or could you suggest some other method.

      Note: In my sequence of length 82600, every 10 numbers are dependent on previous 10 numbers.

      • Jason Brownlee September 19, 2017 at 7:49 am #

        Yes, that was far too many neurons for the first hidden layer. Nice work.

    • Jason Brownlee September 19, 2017 at 7:48 am #

      Perhaps you have too much data to fit into memory.

      Perhaps work with a smaller sample?
      Perhaps try running on a larger computer like AWS?

      • Abuzar September 21, 2017 at 1:58 am #

        Hi Jason, Thanks for responses,

        In extension to the question that Phil Ayres asked:

        My training data shape is (4096,8) that is 4096 rows and each row has 8 features (8 numeric values).

        and the target shape is the same.

        Requirement is one entire row is responsible to predict the next row.
        example:
        input
        [1,2,3,4,5,6,7,8]

        expected
        [100,200,300,400,500,600,700,800]

        How can I use time distributed for this kind of data.
        Can you please provide an example?
        Do I have to call model iteratively, if yes how?

  13. gana October 16, 2017 at 2:22 pm #

    Thank you sir

    For the clarification i have a question that i have a bit of confusion on parameters you have explained above.

    For example:
    3D sequence with 5 samples, 1 time step, and 1 feature. We will define the output as 5 samples with 1 feature.

    X = seq.reshape(5, 1, 1)
    y = seq.reshape(5, 1)

    What are feature and samples in this example?

    Let me have one example assuming we are working on images.
    Then i have 2 classes: class-1 is running man (a sequence with 100 images) and class-2 is walking man (a sequence with 100 images).

    Then:
    – sample = 2 means 1 image from class-1 and 1 image from class-2?
    or just 2 images from class-1 or class-2?
    If batch is set of samples then why we define sample = 5 and again batch =
    5 in this example?
    If sample = 1 then we can define batch = 1. What are difference?

    – time step = 10 means we are taking images (t1-t10) from 100 images of a
    class for prediction? and next time step we are talking images (t2-t11) and
    (t3-t12) etc?

    – is ‘feature’ image dimension or feature map as an output of Conv layer?
    it sounds like a dimension in this example, however cnn says it is output of
    conv layer while we are defining ‘feature’ for input image.
    if my input image has width = 100 and height = 50 then feature = 5000?

Leave a Reply