Stacked Long Short-Term Memory Networks

Gentle introduction to the Stacked LSTM
with example code in Python.

The original LSTM model is comprised of a single hidden LSTM layer followed by a standard feedforward output layer.

The Stacked LSTM is an extension to this model that has multiple hidden LSTM layers where each layer contains multiple memory cells.

In this post, you will discover the Stacked LSTM model architecture.

After completing this tutorial, you will know:

  • The benefit of deep neural network architectures.
  • The Stacked LSTM recurrent neural network architecture.
  • How to implement stacked LSTMs in Python with Keras.

Kick-start your project with my new book Long Short-Term Memory Networks With Python, including step-by-step tutorials and the Python source code files for all examples.

Let’s get started.

Gentle Introduction to Stacked Long Short-Term Memory Networks

Gentle Introduction to Stacked Long Short-Term Memory Networks
Photo by Joost Markerink, some rights reserved.


This post is divided into 3 parts, they are:

  1. Why Increase Depth?
  2. Stacked LSTM Architecture
  3. Implement Stacked LSTMs in Keras

Why Increase Depth?

Stacking LSTM hidden layers makes the model deeper, more accurately earning the description as a deep learning technique.

It is the depth of neural networks that is generally attributed to the success of the approach on a wide range of challenging prediction problems.

[the success of deep neural networks] is commonly attributed to the hierarchy that is introduced due to the several layers. Each layer processes some part of the task we wish to solve, and passes it on to the next. In this sense, the DNN can be seen as a processing pipeline, in which each layer solves a part of the task before passing it on to the next, until finally the last layer provides the output.

Training and Analyzing Deep Recurrent Neural Networks, 2013

Additional hidden layers can be added to a Multilayer Perceptron neural network to make it deeper. The additional hidden layers are understood to recombine the learned representation from prior layers and create new representations at high levels of abstraction. For example, from lines to shapes to objects.

A sufficiently large single hidden layer Multilayer Perceptron can be used to approximate most functions. Increasing the depth of the network provides an alternate solution that requires fewer neurons and trains faster. Ultimately, adding depth it is a type of representational optimization.

Deep learning is built around a hypothesis that a deep, hierarchical model can be exponentially more efficient at representing some functions than a shallow one.

How to Construct Deep Recurrent Neural Networks, 2013.

Need help with LSTMs for Sequence Prediction?

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

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

Stacked LSTM Architecture

The same benefits can be harnessed with LSTMs.

Given that LSTMs operate on sequence data, it means that the addition of layers adds levels of abstraction of input observations over time. In effect, chunking observations over time or representing the problem at different time scales.

… building a deep RNN by stacking multiple recurrent hidden states on top of each other. This approach potentially allows the hidden state at each level to operate at different timescale

How to Construct Deep Recurrent Neural Networks, 2013

Stacked LSTMs or Deep LSTMs were introduced by Graves, et al. in their application of LSTMs to speech recognition, beating a benchmark on a challenging standard problem.

RNNs are inherently deep in time, since their hidden state is a function of all previous hidden states. The question that inspired this paper was whether RNNs could also benefit from depth in space; that is from stacking multiple recurrent hidden layers on top of each other, just as feedforward layers are stacked in conventional deep networks.

Speech Recognition With Deep Recurrent Neural Networks, 2013

In the same work, they found that the depth of the network was more important than the number of memory cells in a given layer to model skill.

Stacked LSTMs are now a stable technique for challenging sequence prediction problems. A Stacked LSTM architecture can be defined as an LSTM model comprised of multiple LSTM layers. An LSTM layer above provides a sequence output rather than a single value output to the LSTM layer below. Specifically, one output per input time step, rather than one output time step for all input time steps.

Stacked Long Short-Term Memory Archiecture

Stacked Long Short-Term Memory Archiecture

Implement Stacked LSTMs in Keras

We can easily create Stacked LSTM models in Keras Python deep learning library

Each LSTMs memory cell requires a 3D input. When an LSTM processes one input sequence of time steps, each memory cell will output a single value for the whole sequence as a 2D array.

We can demonstrate this below with a model that has a single hidden LSTM layer that is also the output layer.

The input sequence has 3 values. Running the example outputs a single value for the input sequence as a 2D array.

To stack LSTM layers, we need to change the configuration of the prior LSTM layer to output a 3D array as input for the subsequent layer.

We can do this by setting the return_sequences argument on the layer to True (defaults to False). This will return one output for each input time step and provide a 3D array.
Below is the same example as above with return_sequences=True.

Running the example outputs a single value for each time step in the input sequence.

Below is an example of defining a two hidden layer Stacked LSTM:

We can continue to add hidden LSTM layers as long as the prior LSTM layer provides a 3D output as input for the subsequent layer; for example, below is a Stacked LSTM with 4 hidden layers.

Further Reading

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


In this post, you discovered the Stacked Long Short-Term Memory network architecture.

Specifically, you learned:

  • The benefit of deep neural network architectures.
  • The Stacked LSTM recurrent neural network architecture.
  • How to implement stacked LSTMs in Python with Keras.

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

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.

See What's Inside

157 Responses to Stacked Long Short-Term Memory Networks

  1. Avatar
    Thabet August 18, 2017 at 11:58 am #

    Thanks alot Jason !
    Your blog is wonderful
    Please keep up the great work
    Best regards/ Thabet

  2. Avatar
    Devakar Kumar Verma August 23, 2017 at 10:48 pm #

    Hi Jason,
    After first stack of LSTM layer, Don’t we need ‘input_shape’ or ‘batch_input_shape’? Need your expert comment.

    • Avatar
      Jason Brownlee August 24, 2017 at 6:43 am #

      No, the input specification is only needed on the first hidden layer.

      • Avatar
        Devakar Kumar Verma August 24, 2017 at 2:14 pm #

        Thanks for your response

  3. Avatar
    Alessandro October 12, 2017 at 3:20 am #

    Can you specify when this approach is needed?

    Wonderful work, thanks!

    • Avatar
      Jason Brownlee October 12, 2017 at 5:35 am #

      Hard question, nice.

      Perhaps generally when you think there may be a hierarchical structure in your sequence data. You can try some stacked LSTMs and see how it impacts model skill.

      Stacked LSTMS will likely need more epochs to complete training, use normal model diagnostics.

      • Avatar
        Lana Lu December 16, 2021 at 6:10 am #

        Hello, could you please explaine more about hierarchical structure? What does this mean?

        • Avatar
          Adrian Tam December 17, 2021 at 7:16 am #

          It just mean stacking different layers to produce the final output. In LSTM, you use one layer of LSTM to convert one sequence to another (e.g. English to French). Then another layer of LSTM to convert that sequence again (e.g. French to German).

  4. Avatar
    DEHUA TANG October 12, 2017 at 7:27 pm #

    Hi Jason, thanks for your work!
    If I have a large size of data, I want to train a Stacked LSTMS about 30 layers.
    Can you tell me if I train a 30 layers Stacked LSTMS, what do I need to pay attention to?
    Why 3 or 4 layers Stacked LSTMS are common?
    would the 30 layers Stacked LSTMS work?

    • Avatar
      Jason Brownlee October 13, 2017 at 5:46 am #

      That is a lot of layers, I have not developed LSTMs that deep myself. I cannot give you good advice.

      Generally, there are diminishing returns beyond 4 layers.

      • Avatar
        Jaime October 18, 2017 at 9:54 am #

        Thank you I think you for your answer, I think that probably there are to much layers and to try to summarize:

        Its necessary and a Dropout and/or Dense(1 LSTM,1Drop, 1 Dense) layer for every LSTM layer in a model or that is almost the same that for example (2 LSTM, 1Drop, 1Dense)

        Thank you in advance

      • Avatar
        Sakshi June 6, 2020 at 10:31 pm #

        Hi Jason,

        I have implemented a LSTM network with 2 layers and on adding 3rd layer, the performance dropped. LSTM is supposed to avoid vanishing gradient problem, then how the performance drop on adding more layers?

        Thank you in advance

  5. Avatar
    mitillo October 16, 2017 at 9:52 pm #

    Hello Jason,

    Like always, very useful article,

    I have a question for you

    I am get use to add LSTM labels in this way:

    layers=shape = [4, seq_len, 1] # feature, window, output
    # neuros=neurons = [128, 128, 32, 1]

    LSTM, Dropout and Dense.

    model.add(LSTM(250, input_shape=(layers[1], layers[0]), return_sequences=True))

    model.add(LSTM(neurons[1], input_shape=(layers[1], layers[0]), return_sequences=True))

    model.add(LSTM(neurons[2], input_shape=(layers[1], layers[0]), return_sequences=False))




    model.compile(loss=’mse’,optimizer=optimizador, metrics=[‘accuracy’])

    There is any difference if I add only LSTM I mean something like that

    model.add(LSTM(250, input_shape=(layers[1], layers[0]), return_sequences=True))

    model.add(LSTM(neurons[1], input_shape=(layers[1], layers[0]), return_sequences=True))

    model.add(LSTM(neurons[2], input_shape=(layers[1], layers[0]), return_sequences=False))


    model.compile(loss=’mse’,optimizer=optimizador, metrics=[‘accuracy’])

    Thank you in advance

  6. Avatar
    Ioudarene November 14, 2017 at 8:36 pm #

    Hello Jason,

    First of all thank you for all your work on your website it is very useful.

    I am implementing in keras a stacked lstm network for solving a many to many sequence problem, I would like to know if for that kind of problem you would still put the parameter return_sequences at the value False for the last lstm layer ?

    Thank you in advance.

  7. Avatar
    Gustav April 27, 2018 at 2:25 am #

    I’m currently working on a LSTM model using just one hidden layer. I was wondering if you know how to best approach whether or not to expand and add more layers.

    Is there a general rule of thumb as to how and why to add more hidden layers and build more deeply when working with time series forecasting?

    Also, how well does adding layers scale with the size of training data. I’m trying to compare training data from 1000 (lower end) to over 100,000 points of training.

    • Avatar
      Jason Brownlee April 27, 2018 at 6:07 am #

      More layers offers more abstraction of the input sequence.

      No good theories or rules of thumb around this that I have seen.

      • Avatar
        Gustav April 27, 2018 at 5:15 pm #


        Then I’ll continue my research and see if I can find anything any correlation between adding more layers and the accuracy of the predictions.

  8. Avatar
    TheRhyno May 10, 2018 at 7:11 am #

    Hi Jason, good post. Would you say that the use of a stacked LSTM is equivalent (or better / worse) in terms of predictive capabilities versus a feedforward network of a few hidden layers which then feeds into a single layer LSTM?

    My line of thinking is if I had a dataset with interesting relationships between the input features themselves, but also as those features change over time, I would expect to get interesting results both from a feedforward network, and an LSTM … does a stacked LSTM get the best of both worlds?

    • Avatar
      Jason Brownlee May 10, 2018 at 8:25 am #

      It can, but it really depends on the specifics of the problem and the relationships being modelled.

      For example, deep encoder-decoder LSTMs do prove very effective at NLP problems with long sequences of text.

  9. Avatar
    Gurgu June 22, 2018 at 11:56 pm #

    Hey there.
    What happens to the computation time needed to train a stacked LSTM? If i have a LSTM with one layer, does a stacked LSTM with m layers need m times as much computation time?

    • Avatar
      Jason Brownlee June 23, 2018 at 6:19 am #

      It depends on your hardware. It is slower, but perhaps not 200% slower.

  10. Avatar
    David Jiménez Barrero July 11, 2018 at 9:51 am #

    Dear Jason,

    Very interesting approach, however I wonder if this is trained using regular backpropagation. I ask this due to the problems backpropagation has when dealing with deep-NN, particularly:

    – Diffusion of gradient problem where gradients that are backpropagated rapidly decrease in magnitude as the depth of the network increases.

    – Appearance of many bad local minima.

    I know there are some approaches for greedy-training layer by layer, is this performed by Keras automatically? Or perhaps is there a maximum network-depth that can be dealt with regular backpropagation?

    Thank you for your help.

    • Avatar
      Jason Brownlee July 11, 2018 at 2:53 pm #

      Yes, but Backprop through time, you can learn more here:

      Yes, Keras does this for us automatically.

      • Avatar
        David Jiménez Barrero July 12, 2018 at 10:10 am #

        Thank you for your reply! However, as BPTT unrolls the network by timesteps, which makes the network seem as an even deeper network as each timestep becomes kind of a new layer; doesn’t this makes the Diffusion of gradient problem worse?

        Additionally, I have a small code question, i.e.:
        In the part where you add a LSTM layer, say “model.add(LSTM(1, return_sequences=True, input_shape=(3,1)))”, the first parameter which you input as 1, defines the number of “units” in the layer. This usually also defines the number of outputs it would have, I wonder if the “return_sequences=True” supersedes this and outputs as many inputs as you have?

        Thank you for your help.

        • Avatar
          Jason Brownlee July 12, 2018 at 3:30 pm #

          The input_shape defines the shape of inputs to the network.

          The 1 refers to the number of units in the first hidden layer which is unrelated to the number of input time steps.

          Return sequences returns a vector for each input time step instead of one vector at the end of the input sequence.

          • Avatar
            robin January 25, 2022 at 7:46 pm #

            Does that mean if there are 100 samples and 60 time steps per sample, return_sequences=True will return 60 vectors of hidden states? But if one hidden state is returned per one sample that goes into the LSTM layer, wouldn’t there be 100 hidden states returned instead of 60?

          • Avatar
            James Carmichael February 2, 2022 at 10:39 am #

            Hi Robin…Please clarify which code listing portion you are considering.

  11. Avatar
    Kamarul July 12, 2018 at 11:51 am #

    Hello Jason,

    Thanks a lot for your work. Your blog really did spark a huge interest in me towards neural network.
    I have read one of your suggested article which is “How to Construct Deep Recurrent Neural Networks”. I wonder whether the novel type of RNN mentioned in that article can be constructed using keras with tensorflow backend.

    Again, good job and please keep up the good work!

    Best regards,

  12. Avatar
    Camilo Macias August 8, 2018 at 10:32 pm #

    Hello Jason!,

    Just as the other people before, first of all thanks for this amazingly helpful blogs and tutorials.

    One question regarding this stacked LSTMs NN.

    I see you seem to always need a Dense layer to give the final output of the stacked network.
    Is that really so? and why is it necessary? does the absence of that Dense layer affect for good or worse the performance of the LSTMs network?

    Thanks again,

    • Avatar
      Jason Brownlee August 9, 2018 at 7:40 am #

      We need something at the output end of the network to make an output that we can interpret as a prediction.

      You could try an LSTM layer, I have never done so.

  13. Avatar
    Jason McDonald September 5, 2018 at 4:39 pm #

    Can stacked LSTM’s learn feed sequence order? For example let’s say I had a random list of a billion numbers that I wanted returned in order. If numbers that are close together in the sort appear at opposite ends of the sequence the LSTM memory may lose track. However if a stack of LSTM’s could learn to rearrange the sequence as it moves up the stack I imagine that could help.

    Currently Im taking a stack of LSTM’s with N outputs and sorting the input sequence between the stacks by one of the output values at each time step. As far as I can tell there is no way to associate gradient with a sorted index so it can only learn to sort through reinforcement (I think).

    • Avatar
      Jason Brownlee September 6, 2018 at 5:32 am #

      Interesting question. I think (gut intuition) it may be too challenging for the LSTM.

  14. Avatar
    ashish September 25, 2018 at 4:23 am #

    Hello, Jason
    I want to use simple or stack lstm network or arrangement of the layer, in parallel. Because of my feature sequence have information from different domain and have difference in morphology.

  15. Avatar
    dulmina September 25, 2018 at 8:04 pm #

    Awesome tutorials,

    I have a question which can be very basic one but i’m struggling with it.

    You said the 1 refers to the number of it means the number of LSTM cells in that layer?

    as far as my knowledge number of lstm cells in first layer is same as to number of time stamps,if that so what 1 actually means?

    Thank you

    • Avatar
      Jason Brownlee September 26, 2018 at 6:15 am #

      No, the number of time steps and the number of LSTM units in the first hidden layer are not related.

      • Avatar
        shiva May 7, 2019 at 8:05 am #

        what is the number of LSTM units in the first hidden layer then? what controls this?

        • Avatar
          Jason Brownlee May 7, 2019 at 2:27 pm #

          The input layer to the model is defined by you and is unrelated to the number of units in the first hidden layer of the network.

  16. Avatar
    dulmina September 26, 2018 at 12:16 pm #

    Oh forget this question.last time i posted it was not shown in the comments area that why i put another one 😀 sorry.

  17. Avatar
    Erik October 19, 2018 at 1:04 am #

    Hi Jason,

    The article states that:

    “building a deep RNN by stacking multiple recurrent hidden states on top of each other. This approach potentially allows the hidden state at each level to operate at different timescale”

    I was wondering what “operate at different timescale” meaning, i.e. the meaning of “timescales” in LSTMs and how stacked LSTM helps with that.


  18. Avatar
    Dmitrijs Balabka October 27, 2018 at 12:45 am #


    Thank you for a very helpful blog.
    I have comment according to Sequential class. Building stacked RNN using keras.layers.RNN + keras.layers.LSTMCell might give better performance (15pct faster):

  19. Avatar
    Saeed October 28, 2018 at 4:05 pm #

    Thanks Jason, your perfect
    I have a question:
    Is it possible to save output of penultimate layer in every epoch in keras?

    • Avatar
      Jason Brownlee October 29, 2018 at 5:54 am #

      Yes, you could have model with two output layers, one is the normal output and one from the second last layer, then run each epoch manually and save the prediction from the second output layer as required.

  20. Avatar
    Hung Nguyen December 3, 2018 at 1:40 pm #

    Hi Jason,
    Is “stacked LSTMs” the same concept with the so-called “Parallel LSTMs”?

  21. Avatar
    Trevor Chandler December 8, 2018 at 4:36 pm #


    This is great work, and very generous to provide to the community.
    Thank you for doing your work in this way!

    Do you have the time and desire to take on some AI work, just to look over some implementations we’ve done, and talk through what is good about it, what may be able to be improved, etc…?

    If not, no problem at all, but if you do, please let me know,

    Thank you for your time,

  22. Avatar
    Adam December 26, 2018 at 8:58 pm #

    Thanks, Jason, It’s so valuable to me. I am a new to LSTM, I have one question:
    Like you mentioned in this blog, It will return one output for each input timestep, And I am here to ensure if I am right in terms of the time step. e.g: The reshaped data is [[[ 0.1][ 0.2][ 0.3]]], The [0.1] represents time step1 input, The [ 0.2] represents time step2 input and The [ 0.3] represents time step3 input. Am I right?

  23. Avatar
    MAK January 2, 2019 at 6:20 am #

    Hello Jason,
    As usual fantastic article.
    I have a question what is the different between stacking LSTM network one on the other to add more dense layers?

    model = Sequential()
    model.add(LSTM(…, return_sequences=True..

    model = Sequential()
    model.add(LSTM(…, return_sequences=True…

    That every dense layers is fully connected.

    • Avatar
      Jason Brownlee January 2, 2019 at 6:43 am #

      Yes, they are different. One is a stacked LSTM and one is an LSTM with lots of dense layers.

      Perhaps I don’t follow your question?

      • Avatar
        MAK January 2, 2019 at 6:28 pm #

        I think my question is not clear

        What is the different of stacked LSTM that each LSTM contains one layer to one LSTM that contains multi Dense layers? from the network structure aspects and accuracy ?
        If I define the dense layer with the same nueron as in the LSTM stack layer

        • Avatar
          Jason Brownlee January 3, 2019 at 6:11 am #

          The difference in the capability between the models really depends on the specific problem.

          Generally, more depth of the RNN can help with complex sequence data. More depth of the fully connected would likely be less beneficial. Try both and compare on your problem.

  24. Avatar
    mk January 17, 2019 at 6:20 pm #

    last LSTM(…) return_sequences=True or false?

  25. Avatar
    Andi January 29, 2019 at 12:35 am #

    Hello Jason,
    Thanks for your great article.

    In case of I have 3 part of data, lets say node a, node b, node c.
    node a is a sequence of a’s position, same goes to node b and node c.
    so, do I need to make 3 of the LSTM layer at once,
    or every node run in different process and own the stacked lstm layer?

    Best Regards

    • Avatar
      Jason Brownlee January 29, 2019 at 6:14 am #

      Perhaps explore a few different framings of your prediction problem and see what works best?

      • Avatar
        Andi January 31, 2019 at 5:02 am #

        Thanks, I did the option 2, and it gave a good result, now I need to optimize.

  26. Avatar
    yh February 8, 2019 at 4:18 pm #

    Hello Jason,

    if the input dimension is the 2D,

    model = Sequential()
    model.add(LSTM(…, return_sequences=True…
    model.add(LSTM(…, return_sequences=True…

    Can be the stacked LSTM?

    • Avatar
      Jason Brownlee February 9, 2019 at 5:54 am #

      No, LSTMs assume each sample has time steps and features.

      • Avatar
        yh February 14, 2019 at 3:01 pm #

        If I reshpae input as ( none , 1, 10) before beginning model

        model.add(LSTM(…, return_sequences=True…
        model.add(LSTM(…, return_sequences=True…
        model.add(Dense(…, input_shape=(None,1)…

        Can this model be stacked?
        I am studying the completed code above this, but why the code maker did put dense input shape none,1.
        All of each layer output shape is 3 dimension

        • Avatar
          Jason Brownlee February 15, 2019 at 7:57 am #

          No, I don’t believe so. You must either change the second LSTM to not return sequences or change the dense part of the model to support sequences via a time distributed wrapper.

  27. Avatar
    Pari February 19, 2019 at 5:52 pm #

    Hello Jason
    your blog helps me every time thanks for that
    I have a question
    I read an example that uses stacked LSTM for seq2seq problem but the input was given in reverse to network.
    please help me why it happens

  28. Avatar
    Gunay February 26, 2019 at 10:41 pm #

    Hi Jason,

    Thanks for the article. I try to understand the main intuition behind on Vanilla and Stacked LSTM. I understand that generally, for the network adding more layer makes it learn difficult tasks better. You have used the sentence like “Given that LSTMs operate on sequence data, it means that the addition of layers adds levels of abstraction of input observations over time. In effect, chunking observations over time or representing the problem at different time scales”. I do not understand here what do you mean exactly about the abstraction of input observations over time. How does it chunk observations over time or representing the problem at different time scales? I would be happy if you could answer this.

    Kind Regards,

    • Avatar
      Jason Brownlee February 27, 2019 at 7:30 am #

      I am hypothesising about what might be going on, we cannot know for sure without an analysis of what each layer has learned.

  29. Avatar
    Srivalya Elluru February 27, 2019 at 8:05 pm #

    Hello Jason Brownlee , before I ask my question I would like to thank you for these amazing tutorials on deep learning using LSTMs.

    Problem I faced:

    My sequence length is 50 and when I gave lstm units to be 100, I started the loss to be nan.
    When I changed the LSTM units to be 50, I stopped getting the loss to be nan for the first epoch but it started to give nan after that.

    Q: Should the number of LSTM units be decided according to the sequence length (time steps for one example) ?

    • Avatar
      Jason Brownlee February 28, 2019 at 6:38 am #

      I don’t think it is related to the number of units.

      Perhaps try scaling your data first?

  30. Avatar
    Hamed May 30, 2019 at 10:52 am #

    Hi Jason,

    Thank you so much for your great tutorials. This website has everything 🙂

    I have a question about stacked LSTM? How can I choose the number of layers for the whole stacked LSTM? and does increasing the layers mean more accuracy and better performance?
    Thank you in advance

  31. Avatar
    Omar June 14, 2019 at 11:05 pm #

    Hello Jason,

    Thank you very much for your articles about RNN and LSTM. Really helped me through the hard times.

    However, I’m still a beginner in this field. Let’s say that I want to make a network with only one LSTM followed by Dense layers.

    Should I use “return_sequences = True” on the LSTM or leave it be (so that the value will be False) ?

    I’m still confused about where to use “return_sequences = True”. Do we only use it when we want to stack LSTM with another LSTM?

    Thank you very much for your assistance.

    • Avatar
      Jason Brownlee June 15, 2019 at 6:35 am #

      Return sequences will return the output of the layer for each input, instead of just the output at the end of the sequence.

      No, return sequences would be false in most cases.

      • Avatar
        Omar August 27, 2019 at 8:00 pm #

        Thank you for replying.

        I have another question. Let’s say we’re going to build stacked LSTM with first LSTM layer contains 200 units and second layer contains 100 units.

        I have a hard time visualizing this. If both layers have the same units, I can simply visualize it like first cell of the first LSTM layer passing hidden states to first cell of the second LSTM layer.

        If the layers have different units, how do I visualize it?

        Thank you very much, you really helped me in understanding Deep Learning in a practical way

        • Avatar
          Jason Brownlee August 28, 2019 at 6:33 am #

          No, all outputs from the first layer are passed to each unit in the second layer. Each unit in a layer is like an independent neural network – they do not communicate with each other within the layer or divide the work.

  32. Avatar
    Sanat June 23, 2019 at 2:18 pm #

    Hi Jason,

    Firstly I wanted to thanks for all your blogs which had almost clarified ML doubts that comes to my mind.

    I am working on Human activity recognition dataset from UCI, when I used a single layer LSTM (96 cells) I got an accuracy of around 91 % for 30 epochs, but when I tried stacking one more LSTM layer on top of that the accuracy decreases tremendously and after certain epochs the train accuracy stays at around 0.16. Is there any reason behind that ?

  33. Avatar
    Jairo June 26, 2019 at 12:19 am #

    Dear Jason,

    When we do the stacked LSTM model as indicated, the last step takes only the state “h_n” from the last layer, or it includes all hidden states “h1”, “h2”, …, “h_n” from all the layers?

    • Avatar
      Jason Brownlee June 26, 2019 at 6:42 am #

      Only from the prior layer.

      • Avatar
        Jairo June 26, 2019 at 7:24 am #

        Thank you, sir. Can I say, then, that the regular stacked LSTM in Keras does not implement the deep Encoder-Decoder architecture (proposed by Sutskever, Vynials e Quoc Le/Google), right? Because in this work (I’m not sure you are familiar with this paper, sorry to talk about a specific case) the hidden state to be decoded is not just the last layer hidden state, but it should be the concatenation of all hidden states from all layers. They justify that saying that since the state should capture all of the sentence, using states from all layers help to preserve the enconding of the entire input sentence.

        Thanks for the information. I’ve searched the Keras API and that was not clear to me.

  34. Avatar
    Rickard July 10, 2019 at 9:03 am #

    Hi Jason!
    This was a very good explanation of stacked LSTM.
    I wonder, have you ever encountered or heard of deep gates? I started learning about LSTM a while back but the only gate architecture I’ve seen so far consist of a single layers. This was surprising since I had assumed that the gates could vary in depth.
    It’d be a great relief if you could answer this question, it’s bugging me constantly and I can’t find any answers on my google searches.

  35. Avatar
    jmc September 12, 2019 at 1:45 am #


    I was wondering what happens with the states, are they passed between different LSTM layers or the only communication between them is the output sequence?

    • Avatar
      Jason Brownlee September 12, 2019 at 5:20 am #

      Typically the states are not passed between layers, only the output of each node at each input time step.

      • Avatar
        jmc September 12, 2019 at 8:05 am #

        So, specifically in Keras if you have this:

        decoded = LSTM(…,return_sequences=True)(inputs, initial_state=[h,c])
        output = LSTM(…)(decoded)

        which states does the 2nd layer uses for the 1st cell?

        If I’m doing an encoder-decoder with multiple LSTM shouldn-t I pass the encoder_states to every LSTM layer in the decoder?

        Like this:
        decoded = LSTM(…,return_sequences=True)(inputs, initial_state=[h,c])
        output = LSTM(…)(decoded, initial_state=[h,c])

        Or may be like this:
        decoded = LSTM(…,return_sequences=True)(inputs, initial_state=[h1,c1)
        output = LSTM(…)(decoded, initial_state=[h2,c2])
        where h1,c1 are from the 1st LSTM layer in the encoder and h2,c2 from the 2nd.

        Thanks for the answers!

  36. Avatar
    Giorgio October 2, 2019 at 4:16 am #

    Hey Jason,

    I have a question. I think I understand how a stacked LSTM layer is connected if there is only one LSTM cell per layer. I’m not sure however, how the LSTM cells would be connected if there are multiple LSTM cells per layer. Does each cell in the first layer return a sequence which is passed to EVERY single cell in the next layer? I tried to implement that in Keras but the second LSTM layer has a suspiciously low amount of parameter :/ .

    I hope I posed the question well. Thanks in advance

    • Avatar
      Jason Brownlee October 2, 2019 at 8:04 am #

      Excellent question.

      The entire output sequence from one layer is provided as input to each node in the next layer.

      Does that make sense?

  37. Avatar
    Shiva November 11, 2019 at 11:33 pm #

    I really appreciate you, Jason.

    I have a stack-GRU for sentiment analysis that always predicts the label of sentiment +1 ane the accuracy is 0.

    in your expriece, what are heuristics to improve the results?


  38. Avatar
    Mahshid Hosseini April 17, 2020 at 6:59 am #

    Thanks for the great tutorial. Is there any way to initialize the second layer LSTM with the final hidden state vector of the first LSTM?

    • Avatar
      Jason Brownlee April 17, 2020 at 7:45 am #

      Perhaps, not sure I have done this before – it does not make sense. I recommend experimenting.

  39. Avatar
    Ceng EMEC April 27, 2020 at 8:05 pm #

    Hi Jason.

    Thanks for the great tutorial.

    What is the difference between DENSE and LSTM layer?

    For example:

    Model1- model.add(LSTM(50, input_shape=(x_train.shape[1], x_train.shape[2])))
    model.add(Dense(16, activation=’relu’))

    Model2- model.add(LSTM(50, return_sequences=True, input_shape=(x_train.shape[1], x_train.shape[2])))
    model.add(LSTM(50, activation=’relu’))

    Also, how does using more than one dense layer and lstm layer benefit the model?

    Thank you very much for your assistance.

  40. Avatar
    Attention June 6, 2020 at 5:31 am #

    Hi there! Would it be beneficial to use attention is such an architecture? What would the new model look like?

  41. Avatar
    Ben July 10, 2020 at 12:48 am #

    Hi Jason,

    Very good post.

    If I have created data out from a physical equation and I know the response is linear (and stochastic).

    There’s really no point in adding more layers, right? How would you go about improving the predictions of data of this type?


  42. Avatar
    ali July 12, 2020 at 4:38 pm # brownlee.thanks a lot for your works in the field of dl
    my question is that whether is their any relation(connection) between units of lstm in a lstm layer?if yes how is the relation.
    second:are lstm units of a lstm layer fully connected with lstm units of next lstm layer?
    third:we are saying lstm has output so why do you add a dense layer to end of network?
    please anwer i’m waiting

    • Avatar
      Jason Brownlee July 13, 2020 at 5:58 am #

      A LSTM layer contains one or more LSTM units.

      Each LSTM unit in one layer receives all of the outputs for each time step provided by all units in the previous layer.

  43. Avatar
    Adriana July 16, 2020 at 6:00 am #

    Hey Jason,

    I don’t usually make comments but your posts have been helping me a lot. I am a Master student at Imperial College of London and I have been developing my master thesis around electricity price predictions. I have been stuck on a few problems when designing my first LSTM architectures and some of the posts you have were very helpful. I also got a couple of your e – books which for now have been quite insightful as well.

    Thank you for all the information you make available online, it is really a life-saver.

    • Avatar
      Jason Brownlee July 16, 2020 at 6:49 am #

      Thank you for your kind words, I’m happy to hear that the tutorials are helpful to your project!

  44. Avatar
    Mike August 7, 2020 at 12:54 am #

    Great tutorial as always! quick question: if I use a stacked model with two or three LSTM, will I have to set the same number of neurons for all of them?


    • Avatar
      Jason Brownlee August 7, 2020 at 6:26 am #


      No, the number of nodes in each layer can vary.

  45. Avatar
    Mohsin Khan August 16, 2020 at 2:21 am #

    Dear Jason, I also need to ask that if I put a BiLSTM layer after a LSTM layer than that model will be recognized as a stacked LSTM model or a hybrid LSTM-BiLSTM model ???

    Thank You

    • Avatar
      Jason Brownlee August 16, 2020 at 5:57 am #

      Perhaps try it and see if it adds value.

      • Avatar
        Mohsin Khan August 16, 2020 at 4:48 pm #

        Dear Jason, I don’t get you. Will it be called a stacked LSTM model or a hybrid LSTM-BiLSTM model ???

        • Avatar
          Jason Brownlee August 17, 2020 at 5:45 am #

          It would be stacked LSTM. I guess you can call it anything you want.

  46. Avatar
    shahzina August 26, 2020 at 5:11 pm #

    Hi Jason,

    How do I build a RNN model where the number of layers can be passed as an argument?

    For example, instead of adding 3 layers, a parameter for number of layers can be passed saying how many layers the model will have?

    Thank you for your time.

    • Avatar
      Jason Brownlee August 27, 2020 at 6:11 am #

      You can write a custom function to do this directly.

  47. Avatar
    Hanna September 4, 2020 at 11:21 pm #

    your blog is wonderful!
    Thank you!

  48. Avatar
    Mayank September 21, 2020 at 5:23 am #

    Can I get code of stacked lstm for the data prediction in matlab

    • Avatar
      Jason Brownlee September 21, 2020 at 8:15 am #

      Sorry, I don’t have any examples for matlab.

  49. Avatar
    Sai October 3, 2020 at 5:00 am #

    Do LSTM layers have any sort of dimension constraints based on the number of cells in the previous layer, like in case of Convolution Layers?

  50. Avatar
    FL. December 15, 2020 at 1:31 am #

    Many thanks for your share.

  51. Avatar
    skyler January 30, 2021 at 9:29 am #

    Hi Jason,
    I am a beginner in ML. I am working to implement a DL model for sound even detection and I have to use three stacked bidirectional GRU and implement a model ( Please note that the attached diagram is part of a bigger model. I need help to implement the stacked BiDirectional GRU layers and upsample the output to get the dimension of (512 x 128), which is the dimension of the input feature.

    • Avatar
      skyler January 30, 2021 at 9:31 am #

      I am trying to implement the model (Fig. 1) from paper (

    • Avatar
      Jason Brownlee January 30, 2021 at 12:34 pm #

      That sounds great.

      Are you having a specific problem with your model?

      • Avatar
        skyler January 30, 2021 at 1:10 pm #

        Hi Jason,
        Yes, I am facing dimensionality issues.
        Actually my feature file dimension (time x frequency) is (512 x x128) . My input layer dimension of the hourglass (HG) is (512, 128, 1). When I input this feature to the hourglass model (as shown in figure 1 of the paper), I get four multi-scale outputs from the hourglass with dimensions :

        (None, 64, 16, 128), (None, 128, 32, 128), (None, 256, 64, 128) and (None, 512, 128, 128). Here 128 (index 3 in each output dimension) is no of channel of HG.
        Now, according to the paper I need to pass each branch (one of the four multi-scale outputs from HG) through the Bi-GRU > upsampling module > Fully Connected layers > Sigmoid.

        Now, the problem is that Bi-GRU accepts 3D input. So I have to use a 1×1 convolution layer to reduce channels from 128 to 1 followed by reshaping the tensor from 4D to 3D.
        It is given that “Since the output’s dimensions of all scales are different, the dimension of the input layer and the hidden layer of these GRU modules are also different. The dimension of input layers and hidden layers are [16, 32, 64, 128] and [32, 32, 64, 64] respectively, and the number of the hidden layer is 3”

        1)should I use return_sequence as True in all three Bi-GRU layers?
        2) As Output of Bi-GRU module is 2D/3D (Depends in return_sequence ) but Upsample2D() layer support 4D tensor. So, how to handle this?

        Best Regards,

        • Avatar
          Jason Brownlee January 31, 2021 at 5:30 am #

          Sorry, I’m not familiar with that model so I cannot give you specific advice.

          Perhaps you can contact the authors directly?
          Perhaps you can prototype a few approaches and use them to learn more about the data and model?
          Perhaps you can find an existing implementation from which to learn?

  52. Avatar
    Thony FdG April 1, 2021 at 4:03 am #

    Hello, I have a big problem and I don’t find the answer.
    I have this model :

    model = Sequential()

    model.add(LSTM(units=100, input_shape=(x_train.shape[1],5), return_sequences=True))

    model.add(LSTM(50, return_sequences=False))


    start = time.time()
    model.compile(loss=’mse’, optimizer=’rmsprop’)

    My question is : How are connected the lstm cell between each other in a layer ? And I have a difference between the first hidden layer and the second (100 cells and 50) how the lstm neurons are connected between each hidden layer (like the first ouput of layer 1 is connected to all input of the second hidden layer and etc) ?

    thanks in advance

    • Avatar
      Jason Brownlee April 1, 2021 at 8:23 am #

      What do you mean exactly?

      E.g. the output from one layer is passed to the next layer, they are not connected directly (e.g. units operate independently in a layer), rather just data is passed between them.

  53. Avatar
    Thony FdG April 2, 2021 at 8:36 pm #

    Here are currently my misunderstandings:

    Using keras LSTM, is “units” the number of LSTM cells in a layer?

    If so, how are they connected to each other in the same layer?

    How is the first layer connected to the second in two cases if the number of units is the same and the number of units is different?

    • Avatar
      Jason Brownlee April 3, 2021 at 5:29 am #

      In Keras, a unit is a cell is a node. They are all the same thing.

      The output of one layer is passed as input to the next layer.

      Each “node” outputs the last value in the sequence unless it is configured to output each value in the sequence (return_sequences=True).

  54. Avatar
    Thony FdG April 4, 2021 at 3:00 am #

    so this representation is not good?

    If not good, how should the correct representation be? please

  55. Avatar
    Lee June 29, 2021 at 1:47 am #

    Hi Jason, I have a question on running a stacked LSTM model on parallel. In my model, I stacked 4 LSTM layers with bidirectional setup. I was trying to use multiple GPUs to accelerate the training process, but the training time does not seem to be affected. Could you help me on the issue of LSTM parallelism with multiple GPUs.

    The structure of my LSTM net looks something like the following:

    model= Sequential()
    model.add(Bidirectional(LSTM(…),input_shape = (…), merge_mode=’sum’))




    regressor.add(Dense(…, activation=’sigmoid’))

    • Avatar
      Jason Brownlee June 29, 2021 at 4:49 am #

      I don’t believe GPUs help much with RNNs like LSTMs.

  56. Avatar
    Rex August 26, 2021 at 6:13 pm #

    Hey Jason, I’ve been working on Demand prediction. The data is time-series.
    I have a problem, but I have no idea how to improve.
    The loss value decreases drastically at the first epoch, then in ten epochs, the loss stops decreasing. It stays almost the same value, just drifts 0.3 ~ -0.3. The performance isn’t bad. But not very good actually. It can get the trend, like peak and valley. However, the value isn’t precise. Do you have any suggestions? I’m just new to LSTM.

    • Avatar
      Adrian Tam August 27, 2021 at 6:05 am #

      Maybe you already reached the potential of your model. In the old time, neural network has only 3 layers (because of computational power is too limited) and the accuracy is not good for many practical use. That’s what it can best do! Only until later we can support building a neural network of more layers, then we see some improvement. I believe your case is similar. Maybe you try a different design of the network?

  57. Avatar
    Michele August 28, 2021 at 5:06 am #

    Hi thanks for sharing knowledge

    I have a qestion.
    in the context of multiple LSTM layers

    is there some general rule for selecting the number of unit for each layer?

    [100, 50, 25] ex 50% reduction or any other %

    [100,100,100] all equal

    [100,200,30] does not matter larger followed bay smaller than larger…

    Any suggestion

    • Avatar
      Adrian Tam August 28, 2021 at 9:45 am #

      No rule here. You need to test it out with your data set to find the suitable one.

  58. Avatar
    Mayowa Osibodu April 25, 2022 at 9:23 pm #

    Great tutorial! Really clarifies things on how to go about stacking LSTMS ????

    • Avatar
      James Carmichael April 26, 2022 at 3:59 am #

      Great feedback Mayowa!

  59. Avatar
    oberlin machine learning December 14, 2022 at 5:24 pm #

    Our group of 4 people literally opened your image in a new tab, enlarged it, and bowed down in utter gratitude. You’ve saved our project, thank you sir.

    • Avatar
      James Carmichael December 15, 2022 at 9:32 am #

      Thank you for your feedback and support!

Leave a Reply