[New Book] Click to get Mastering Digital Art with Stable Diffusion!
Use the offer code 20offearlybird to get 20% off. Hurry, sale ends soon!

Adding a Custom Attention Layer to a Recurrent Neural Network in Keras

Deep learning networks have gained immense popularity in the past few years. The “attention mechanism” is integrated with deep learning networks to improve their performance. Adding an attention component to the network has shown significant improvement in tasks such as machine translation, image recognition, text summarization, and similar applications.

This tutorial shows how to add a custom attention layer to a network built using a recurrent neural network. We’ll illustrate an end-to-end application of time series forecasting using a very simple dataset. The tutorial is designed for anyone looking for a basic understanding of how to add user-defined layers to a deep learning network and use this simple example to build more complex applications.

After completing this tutorial, you will know:

  • Which methods are required to create a custom attention layer in Keras
  • How to incorporate the new layer in a network built with SimpleRNN

Kick-start your project with my book Building Transformer Models with Attention. It provides self-study tutorials with working code to guide you into building a fully-working transformer model that can
translate sentences from one language to another...

Let’s get started.

Adding A Custom Attention Layer To Recurrent Neural Network In Keras <br> Photo by

Adding a custom attention layer to a recurrent neural network in Keras
Photo by Yahya Ehsan, some rights reserved.

Tutorial Overview

This tutorial is divided into three parts; they are:

  • Preparing a simple dataset for time series forecasting
  • How to use a network built via SimpleRNN for time series forecasting
  • Adding a custom attention layer to the SimpleRNN network


It is assumed that you are familiar with the following topics. You can click the links below for an overview.

The Dataset

The focus of this article is to gain a basic understanding of how to build a custom attention layer to a deep learning network. For this purpose, let’s use a very simple example of a Fibonacci sequence, where one number is constructed from the previous two numbers. The first 10 numbers of the sequence are shown below:

0, 1, 1, 2, 3, 5, 8, 13, 21, 34, …

When given the previous ‘t’ numbers, can you get a machine to accurately reconstruct the next number? This would mean discarding all the previous inputs except the last two and performing the correct operation on the last two numbers.

For this tutorial, you’ll construct the training examples from t time steps and use the value at t+1 as the target. For example, if t=3, then the training examples and the corresponding target values would look as follows:

Want to Get Started With Building Transformer Models with Attention?

Take my free 12-day email crash course now (with sample code).

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

The SimpleRNN Network

In this section, you’ll write the basic code to generate the dataset and use a SimpleRNN network to predict the next number of the Fibonacci sequence.

The Import Section

Let’s first write the import section:

Preparing the Dataset

The following function generates a sequence of n Fibonacci numbers (not counting the starting two values). If scale_data is set to True, then it would also use the MinMaxScaler from scikit-learn to scale the values between 0 and 1. Let’s see its output for n=10.

Next, we need a function get_fib_XY() that reformats the sequence into training examples and target values to be used by the Keras input layer. When given time_steps as a parameter, get_fib_XY() constructs each row of the dataset with time_steps number of columns. This function not only constructs the training set and test set from the Fibonacci sequence but also shuffles the training examples and reshapes them to the required TensorFlow format, i.e., total_samples x time_steps x features. Also, the function returns the scaler object that scales the values if scale_data is set to True.

Let’s generate a small training set to see what it looks like. We have set time_steps=3 and total_fib_numbers=12, with approximately 70% of the examples going toward the test points. Note the training and test examples have been shuffled by the permutation() function.

Setting Up the Network

Now let’s set up a small network with two layers. The first one is the SimpleRNN layer, and the second one is the Dense layer. Below is a summary of the model.

Train the Network and Evaluate

The next step is to add code that generates a dataset, trains the network, and evaluates it. This time around, we’ll scale the data between 0 and 1. We don’t need to pass the scale_data parameter as its default value is True.

As output, you’ll see the progress of the training and the following values for the mean square error:

Adding a Custom Attention Layer to the Network

In Keras, it is easy to create a custom layer that implements attention by subclassing the Layer class. The Keras guide lists clear steps for creating a new layer via subclassing. You’ll use those guidelines here. All the weights and biases corresponding to a single layer are encapsulated by this class. You need to write the __init__ method as well as override the following methods:

  • build(): The Keras guide recommends adding weights in this method once the size of the inputs is known. This method “lazily” creates weights. The built-in function add_weight() can be used to add the weights and biases of the attention layer.
  • call(): The call() method implements the mapping of inputs to outputs. It should implement the forward pass during training.

The Call Method for the Attention Layer

The call method of the attention layer has to compute the alignment scores, weights, and context. You can go through the details of these parameters in Stefania’s excellent article on The Attention Mechanism from Scratch. You’ll implement the Bahdanau attention in your call() method.

The good thing about inheriting a layer from the Keras Layer class and adding the weights via the add_weights() method is that weights are automatically tuned. Keras does an equivalent of “reverse engineering” of the operations/computations of the call() method and calculates the gradients during training. It is important to specify trainable=True when adding the weights. You can also add a train_step() method to your custom layer and specify your own method for weight training if needed.

The code below implements the custom attention layer.

RNN Network with Attention Layer

Let’s now add an attention layer to the RNN network you created earlier. The function create_RNN_with_attention() now specifies an RNN layer, an attention layer, and a Dense layer in the network. Make sure to set return_sequences=True when specifying the SimpleRNN. This will return the output of the hidden units for all the previous time steps.

Let’s look at a summary of the model with attention.

Train and Evaluate the Deep Learning Network with Attention

It’s time to train and test your model and see how it performs in predicting the next Fibonacci number of a sequence.

You’ll see the training progress as output and the following:

You can see that even for this simple example, the mean square error on the test set is lower with the attention layer. You can achieve better results with hyper-parameter tuning and model selection. Try this out on more complex problems and by adding more layers to the network. You can also use the scaler object to scale the numbers back to their original values.

You can take this example one step further by using LSTM instead of SimpleRNN, or you can build a network via convolution and pooling layers. You can also change this to an encoder-decoder network if you like.

Consolidated Code

The entire code for this tutorial is pasted below if you would like to try it. Note that your outputs would be different from the ones given in this tutorial because of the stochastic nature of this algorithm.

Further Reading

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





In this tutorial, you discovered how to add a custom attention layer to a deep learning network using Keras.

Specifically, you learned:

  • How to override the Keras Layer class.
  • The method build() is required to add weights to the attention layer.
  • The call() method is required for specifying the mapping of inputs to outputs of the attention layer.
  • How to add a custom attention layer to the deep learning network built using SimpleRNN.

Do you have any questions about RNNs discussed in this post? Ask your questions in the comments below, and I will do my best to answer.

Learn Transformers and Attention!

Building Transformer Models with Attention

Teach your deep learning model to read a sentence

...using transformer models with attention

Discover how in my new Ebook:
Building Transformer Models with Attention

It provides self-study tutorials with working code to guide you into building a fully-working transformer models that can
translate sentences from one language to another...

Give magical power of understanding human language for
Your Projects

See What's Inside

, , , ,

57 Responses to Adding a Custom Attention Layer to a Recurrent Neural Network in Keras

  1. Avatar
    5cc October 13, 2021 at 1:41 pm #

    Hi,I have a question
    I have tried to use LSTM instead of simple_RNN with your help.
    then I found it only has train loss, I can not find the val_loss.
    So how can I monitor the overfitting problem?
    I would like to ask you for help.
    thank you very much!

    • Adrian Tam
      Adrian Tam October 14, 2021 at 3:12 am #

      Likely you didn’t provide validation data when you called fit(), hence no validation has been performed. See this code snippet:

      history = model.fit(X_train, y_train, epochs=200, batch_size=16, validation_data=(X_test,y_test))

      • Avatar
        5cc November 1, 2021 at 1:32 pm #

        Could you tell me how to monitor the overfitting problem with your code?
        Or is it that RNN models with attention mechanism do not need to consider this overfitting problem?
        I don’t see any tuning hyperparameters involved in your example.

        • Avatar
          5cc November 1, 2021 at 2:31 pm #

          I have fixed the problem
          Thank you!

  2. Avatar
    Bhaskar October 18, 2021 at 10:31 pm #

    I am getting the following error : “NameError: name ‘Layer’ is not defined”

    • Adrian Tam
      Adrian Tam October 20, 2021 at 8:50 am #

      Do you have “from keras.layers import Layer”?

  3. Avatar
    Dr. Fouz Sattar October 19, 2021 at 6:55 pm #

    Well structured and well described with clarity.

  4. Avatar
    Ray Huang October 26, 2021 at 12:54 pm #

    Thanks for the clarity explanation and example.
    I have some different result when executing the code above.
    The attention layer’s output should be (None, 2) above.
    However, I get (None, 20, 2) and cause dimensions doesn’t match error.

    The attention layer does output the (None, 2)
    But when it was concatenated to model it becomes (None, 20, 2)
    Could you please tell me what’s the problem?
    Thank you.

    • Adrian Tam
      Adrian Tam October 27, 2021 at 2:58 am #

      It is hard to tell what’s wrong. Can you try to copy over the example code at the end of this post and compare with your version?

      • Avatar
        Ray November 15, 2021 at 6:24 pm #

        Thank you for the reply.

        I just copy the code at the end and execute it on my PC.
        The error is below:

        ValueError: Error when checking target: expected dense_2 to have 3 dimensions, but got array with shape (826, 1)

        It seems that the attention layer return the sequence?

        • Adrian Tam
          Adrian Tam November 16, 2021 at 2:24 am #

          I just verified and don’t see the error. Did you see which line is triggering that?

          • Avatar
            zeeshan November 24, 2021 at 2:54 pm #

            Getting same error. I don’t know but the Dense layer is expecting 3 values, but it is getting 2. I try to use Flatten before Dense, i think it is also not working.
            Below is the line of code.

            ValueError: Error when checking target: expected dense_2 to have 3 dimensions, but got array with shape (826, 1)

            –> 123 model_attention.fit(trainX, trainY, epochs=epochs, batch_size=1, verbose=2)

            My python version is 3.7.3
            Keras 2.3.1
            tensorflow 2.2.0

          • Avatar
            Ray November 29, 2021 at 1:54 pm #

            Thanks for the reply.
            I just use the colab to run the code and get the right result.
            It may caused by environment.
            But I’m not sure which part is wrong.
            Anyway, thank you for the very useful guide on attention layer.
            It’s really helpful.

  5. Avatar
    George Gousios November 22, 2021 at 11:11 am #

    In a seq2seq model trained for time series forecasting and having a 3-stack LSTM encoder plus a similar decoder, would the following approach be reasonable?
    1) Calculate the attention scores after the last encoder LSTM.
    2) Condition the first decoder LSTM with attention outputs (initialize LSTM states from context vector).

    • Avatar
      Vadim Avkhimenia December 2, 2021 at 2:12 am #

      I have the same question…

  6. Avatar
    Zhiwei November 24, 2021 at 2:50 pm #

    I also meet the error “ValueError: Error when checking target: expected dense_9 to have 3 dimensions, but got array with shape (826, 1)” in the line”model_attention.fit(trainX, trainY, epochs=epochs, batch_size=1, verbose=2)”, do you have any suggestion? Thank you.

  7. Avatar
    Zhiwei November 25, 2021 at 6:59 pm #

    After the “model_attention.summary()”, the Output Shape of the attention layer and the dense layer is “(None, 20,2), (None,20,1)”, which is different from what you present in this blog “(None 2) and (None, 1) “, is there something wrong?

  8. Avatar
    zhiqiang November 26, 2021 at 4:57 pm #

    I solved this problem by upgrading the version of Tensorflow from 1.14.0 to 2.7.0

    • Adrian Tam
      Adrian Tam November 29, 2021 at 8:35 am #

      Tensorflow 1.x is too old to use nowadays. The recent tutorials on this blog are all checked against the 2.x version while the older posts may need revision.

  9. Avatar
    Ray November 29, 2021 at 1:50 pm #

    Reply to myself.
    I run the code on colab and the result is the same as the author.
    I think that might be caused by environment. QAQ

    • Adrian Tam
      Adrian Tam December 2, 2021 at 12:27 am #

      Thanks! That’s a good way to check. Usually for python libraries, you can print “libraryname.__version__” to check the version. That helps to identify when one machine report a different result than another.

  10. Avatar
    Nav December 31, 2021 at 4:08 am #

    I ran your code on my machine but it produces different summary for SimpleRNN+Attention network. The output of attention layer in my summary is
    attention_1 (attention) (None,20, 2) 22

    This produces an error for dense layer. I don’t know where is the problem!

    • Avatar
      James Carmichael December 31, 2021 at 10:10 am #

      Hi Nav…Hopefully the issue is now resolved with reinstallation of TensorFlow.


  11. Avatar
    Nav December 31, 2021 at 6:10 am #

    Resolved the issue by upgrading tensorflow.
    Thank you so much for this amazing tutorial!

    • Avatar
      James Carmichael December 31, 2021 at 10:07 am #

      You are very welcome Nav!


  12. Avatar
    Ray January 13, 2022 at 2:04 pm #

    Hi, thank you for the post.

    For the class attention(Layer), why is there only one weight vector for the alignment score (e)?

    In Bahdanau et al’s paper (https://arxiv.org/pdf/1409.0473.pdf), I see that the alighment model has three weights, v_a, W_a, and U_a.

    Furthermore, can you clarify if attention mechanisms are appropriate for non-autoencoder architectures?

    Thanks again.

  13. Avatar
    Keone January 17, 2022 at 7:39 am #

    I noticed Keras has an attention layer implementation. Do you have any plans to use the keras attention layer implementation as an example in one of your blogs? Thank you.


  14. Avatar
    Martin March 4, 2022 at 2:40 am #


    It is my understanding that attention is a way to decrease information in matrixes that are less semantic by multiplying them with a scalar value between 0-1.

    In you case, you’re only left with one time-step (the time_step dimension is lost after attention layer).

    Instead of:
    Layer (type) Output Shape Param #
    input_2 (InputLayer) [(None, 20, 1)] 0
    simple_rnn_2 (SimpleRNN) (None, 20, 2) 8
    attention_1 (attention) (None, 2) 22
    dense_2 (Dense) (None, 1) 3

    shouldn’t the attions yeild these shapes:

    Layer (type) Output Shape Param #
    input_2 (InputLayer) [(None, 20, 1)] 0
    simple_rnn_2 (SimpleRNN) (None, 20, 2) 8
    attention_1 (attention) (None, 20, 2) ???
    dense_2 (Dense) (None, 20, 1) ???

    To make compare with a NLP subject. Shouldn’t you keep the whole sentence and multiply irrelevant words with a scalar value close to 0 instead of just choosing the most semantic word?

    • Avatar
      James Carmichael March 4, 2022 at 2:26 pm #

      Hi Martin…Since semantics are the ultimate goal of NLP, I would recommend choosing the most semantic word.

  15. Avatar
    Debashis March 15, 2022 at 11:38 pm #


    Thanks for the implementation of the custom attention layer.
    Here I want to print the probability values or the alpha.
    If I simply print(alpha), then it’s giving below output.

    Tensor(“attention_1/ExpandDims:0”, shape=(None, 20, 1), dtype=float32)

    But I want the values of alpha. Could you please help to find out the probability values?

  16. Avatar
    yanwei April 24, 2022 at 8:31 pm #

    reply to myself:
    sorry,I saw the two part data was different

  17. Avatar
    Ruan May 6, 2022 at 6:44 pm #

    Hello, ,

    The code runs for me, but the standard recurrent model usually outperforms the
    attention model. Could this be just a version issue or is there a bigger problem?
    My tensorflow is version 2.3.1 and my keras is version 2.4.0

    • Avatar
      James Carmichael May 7, 2022 at 5:49 am #

      Hi Ruan…What are you using as your measures of performance to compare the models?

  18. Avatar
    Arash May 29, 2022 at 7:47 am #

    Hello Dr.Brownlee
    I want to build a mode like

    Input => input_shape => (1, 7, 1) => (batch_size=1, n_steps=7, n_features=1)
    “LSTM” => stateful=True
    “”Attention”” => Error
    “Conv1D” => 128, 3,

    I can make (without “Attention_layer”)



    but when, I add “Attention_layer”, I will have error…

    /usr/local/lib/python3.7/dist-packages/keras/engine/input_spec.py in assert_input_compatibility(input_spec, inputs, layer_name)
    226 ndim = x.shape.rank
    227 if ndim is not None and ndim 228 raise ValueError(f’Input {input_index} of layer “{layer_name}” ‘
    229 ‘is incompatible with the layer: ‘
    230 f’expected min_ndim={spec.min_ndim}, ‘

    ValueError: Input 0 of layer “Conv1D_01” is incompatible with the layer: expected min_ndim=3, found ndim=2. Full shape received: (1, 7)

    How can/must I change “Attention_layer” that I can put it between LSTM & Conv1D?

    Thanks so much.

  19. Avatar
    Arash May 29, 2022 at 7:56 am #

    sorry I forgot something

    I used 2 lstm

    Input => input_shape => (1, 7, 1) => (batch_size=1, n_steps=7, n_features=1)
    “LSTM” => stateful=True, return_sequences=True,
    “LSTM” => stateful=True
    “”Attention”” => Error
    “Conv1D” => 128, 3,

    and, another question
    How can we use “Conv2D”

    Input => input_shape => (1, 7, 1) => (batch_size=1, n_steps=7, n_features=1)
    “LSTM” => stateful=True, return_sequences=True,
    “LSTM” => stateful=True
    “”Attention”” => Error
    “Conv2D” => 128, (3,3),
    “Conv2D” => 64, (3,3),


  20. Avatar
    Arash May 29, 2022 at 10:16 pm #

    Thanks a lot for last guiding.
    But I don’t have problem with ‘Conv1D’ & ‘Conv2D’.

    My problem is;

    I don’t know how can I connect a “custom attention layer” to a Conv1D /Conv2D layer. when “custom attention layer” is before Conv1D /Conv2D layer and before of “custom attention layer” is a LSTM layer .
    Even when I don’t use code of “custom attention layer” which is in this page, and I use “attention layer” that is in keras, like this:

    Input layer => LSTM layer => “custom attention layer” => Conv1D layer => …..

    Again code can not compile without error.
    I watched same my problem in this page,


    But no one seems to have a solution to this problem.

    Thanks for spending your time for reading this comment.

  21. Avatar
    Jack June 8, 2022 at 5:05 pm #


    how to use different evaluation metrics to evaluate the results like MAE and etc?

  22. Avatar
    Mohammed September 3, 2022 at 11:01 am #

    Hello, thank you for the post.
    i try to change this code and make it seq2seq for example but fail do you have any example for seq2seq?

  23. Avatar
    Mohammed September 3, 2022 at 11:16 am #

    i mean make it many to many 🙂

    • Avatar
      James Carmichael September 4, 2022 at 9:52 am #

      Thank you for your feedback Mohammed!

  24. Avatar
    Patrick Ng December 14, 2022 at 8:04 pm #

    A beginner question:
    Inside the build() function of the custom attention layer, it has these two lines:

    self.W = self.add_weight(name=’attention_weight’, shape=(input_shape[-1], 1), initializer=’random_normal’, trainable=True)

    self.b=self.add_weight(name=’attention_bias’, shape=(input_shape[1],1), initializer=’zeros’, trainable=True)

    In the 2nd line, the shape param is (input_shape[1], 1). Why input_shape[1]? (vs. input_shape[-1] in the 1st line)

    My understanding is that for RNN, input_shape = (time_steps, features).
    So input_shape[-1] == input_shape[1]. But why the code wrote it differently among these two lines? What’s the reason behind?

    • Avatar
      James Carmichael December 15, 2022 at 9:30 am #

      Hi Patrick…Your understanding is correct. The alternative notation you mentioned should work as well. Let us know what you find with your implementation.

  25. Avatar
    Patrick December 15, 2022 at 12:52 pm #

    So the difference is not intentional, but inconsistency may create confusion to a learner. Thanks!

    • Avatar
      James Carmichael December 16, 2022 at 8:22 am #

      Thank you for your feedback Patrick!

  26. Avatar
    Markus February 16, 2023 at 8:01 pm #

    Hi there,
    quick feedback: while the idea of a “Bahdanau self attention” is interesting, the training results for both scenarios (with/without attention) are completely meaningless due to the use of Fibonacci numbers. These numbers grow basically exponential, so linearly scaling 1200 of them to the range of 0…1 doesn’t work. Take a look at your scaled trainX, most numbers are like 1E-200, which is 0 when converted to tf.float32.
    Both networks don’t learn anything, the mse values are completely artificial, therefore this tutorial doesn’t show that attention helps. You could take the log before, but then the log Fibonacci sequence is linear, therefore trivial. The Fibonacci numbers simply aren’t useable here.

    • Avatar
      James Carmichael February 17, 2023 at 10:51 am #

      Hi Markus…Thank you for your feedback!

    • Avatar
      Farid T. June 26, 2023 at 10:21 am #

      Came here to give the exact same feedback and write a similar comment. I spent half a day trying to understand what was “wrong” with my code, only to find out eventually that virtually all model predictions were identical and therefore useless. (The models are not learning). After putting some thought into it, I realized you couldn’t possibly expect much better when you convert a range as wide as (1, F_1200) to the (0, 1) range and back. Might be a better idea to use the same dataset as the one in chapter 7 (sunspots) for this chapter. I’m going to try it myself later. Thank you for the book though. I am learning a lot.

  27. Avatar
    Filipa Bento May 14, 2023 at 4:35 am #

    Hey! Thank you for this tutorial.
    What would be the diference bwteen appling the attention mechanism after the input and then the RNN compared to placing the attention after the RNN?

    • Avatar
      James Carmichael May 14, 2023 at 10:08 am #

      Hi Filipa…I would recommend trying both methods. Let us know how your models perform.

  28. Avatar
    kay July 4, 2023 at 9:54 pm #

    Hi, James. thanks for your post. May I ask that what is the difference for the attention layer on this post and the attention layer (Class AttentionDecoder(Recurrent):) on 2017 post : https://machinelearningmastery.com/encoder-decoder-attention-sequence-to-sequence-prediction-keras/
    I am a little confuse that is it the attention in this post is a layer to add between encoder and decoder layer, but for the code in 2017 post, the attention layer is combined with de decoder layer? thanks

  29. Avatar
    J July 21, 2023 at 2:44 pm #

    what is the corresponding Q, K V, in the attention class call method?

  30. Avatar
    Eman Ayad August 15, 2023 at 1:32 am #

    what is the type of this attention?

Leave a Reply