How to Develop a Character-Based Neural Language Model in Keras

A language model predicts the next word in the sequence based on the specific words that have come before it in the sequence.

It is also possible to develop language models at the character level using neural networks. The benefit of character-based language models is their small vocabulary and flexibility in handling any words, punctuation, and other document structure. This comes at the cost of requiring larger models that are slower to train.

Nevertheless, in the field of neural language models, character-based models offer a lot of promise for a general, flexible and powerful approach to language modeling.

In this tutorial, you will discover how to develop a character-based neural language model.

After completing this tutorial, you will know:

  • How to prepare text for character-based language modeling.
  • How to develop a character-based language model using LSTMs.
  • How to use a trained character-based language model to generate text.

Kick-start your project with my new book Deep Learning for Natural Language Processing, including step-by-step tutorials and the Python source code files for all examples.

Let’s get started.

  • Update Feb/2018: Minor update to generation for API change in Keras 2.1.3.
  • Update Feb/2021: Updated final code example to remove redundant line.
How to Develop a Character-Based Neural Language Model in Keras

How to Develop a Character-Based Neural Language Model in Keras
Photo by hedera.baltica, some rights reserved.

Tutorial Overview

This tutorial is divided into 4 parts; they are:

  1. Sing a Song of Sixpence
  2. Data Preparation
  3. Train Language Model
  4. Generate Text

Need help with Deep Learning for Text Data?

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

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

Sing a Song of Sixpence

The nursery rhyme “Sing a Song of Sixpence” is well known in the west.

The first verse is common, but there is also a 4 verse version that we will use to develop our character-based language model.

It is short, so fitting the model will be fast, but not so short that we won’t see anything interesting.

The complete 4 verse version we will use as source text is listed below.

Copy the text and save it in a new file in your current working directory with the file name ‘rhyme.txt‘.

Data Preparation

The first step is to prepare the text data.

We will start by defining the type of language model.

Language Model Design

A language model must be trained on the text, and in the case of a character-based language model, the input and output sequences must be characters.

The number of characters used as input will also define the number of characters that will need to be provided to the model in order to elicit the first predicted character.

After the first character has been generated, it can be appended to the input sequence and used as input for the model to generate the next character.

Longer sequences offer more context for the model to learn what character to output next but take longer to train and impose more burden on seeding the model when generating text.

We will use an arbitrary length of 10 characters for this model.

There is not a lot of text, and 10 characters is a few words.

We can now transform the raw text into a form that our model can learn; specifically, input and output sequences of characters.

Load Text

We must load the text into memory so that we can work with it.

Below is a function named load_doc() that will load a text file given a filename and return the loaded text.

We can call this function with the filename of the nursery rhyme ‘rhyme.txt‘ to load the text into memory. The contents of the file are then printed to screen as a sanity check.

Clean Text

Next, we need to clean the loaded text.

We will not do much to it here. Specifically, we will strip all of the new line characters so that we have one long sequence of characters separated only by white space.

You may want to explore other methods for data cleaning, such as normalizing the case to lowercase or removing punctuation in an effort to reduce the final vocabulary size and develop a smaller and leaner model.

Create Sequences

Now that we have a long list of characters, we can create our input-output sequences used to train the model.

Each input sequence will be 10 characters with one output character, making each sequence 11 characters long.

We can create the sequences by enumerating the characters in the text, starting at the 11th character at index 10.

Running this snippet, we can see that we end up with just under 400 sequences of characters for training our language model.

Save Sequences

Finally, we can save the prepared data to file so that we can load it later when we develop our model.

Below is a function save_doc() that, given a list of strings and a filename, will save the strings to file, one per line.

We can call this function and save our prepared sequences to the filename ‘char_sequences.txt‘ in our current working directory.

Complete Example

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

Run the example to create the ‘char_seqiences.txt‘ file.

Take a look inside you should see something like the following:

We are now ready to train our character-based neural language model.

Train Language Model

In this section, we will develop a neural language model for the prepared sequence data.

The model will read encoded characters and predict the next character in the sequence. A Long Short-Term Memory recurrent neural network hidden layer will be used to learn the context from the input sequence in order to make the predictions.

Load Data

The first step is to load the prepared character sequence data from ‘char_sequences.txt‘.

We can use the same load_doc() function developed in the previous section. Once loaded, we split the text by new line to give a list of sequences ready to be encoded.

Encode Sequences

The sequences of characters must be encoded as integers.

This means that each unique character will be assigned a specific integer value and each sequence of characters will be encoded as a sequence of integers.

We can create the mapping given a sorted set of unique characters in the raw input data. The mapping is a dictionary of character values to integer values.

Next, we can process each sequence of characters one at a time and use the dictionary mapping to look up the integer value for each character.

The result is a list of integer lists.

We need to know the size of the vocabulary later. We can retrieve this as the size of the dictionary mapping.

Running this piece, we can see that there are 38 unique characters in the input sequence data.

Split Inputs and Output

Now that the sequences have been integer encoded, we can separate the columns into input and output sequences of characters.

We can do this using a simple array slice.

Next, we need to one hot encode each character. That is, each character becomes a vector as long as the vocabulary (38 elements) with a 1 marked for the specific character. This provides a more precise input representation for the network. It also provides a clear objective for the network to predict, where a probability distribution over characters can be output by the model and compared to the ideal case of all 0 values with a 1 for the actual next character.

We can use the to_categorical() function in the Keras API to one hot encode the input and output sequences.

We are now ready to fit the model.

Fit Model

The model is defined with an input layer that takes sequences that have 10 time steps and 38 features for the one hot encoded input sequences.

Rather than specify these numbers, we use the second and third dimensions on the X input data. This is so that if we change the length of the sequences or size of the vocabulary, we do not need to change the model definition.

The model has a single LSTM hidden layer with 75 memory cells, chosen with a little trial and error.

The model has a fully connected output layer that outputs one vector with a probability distribution across all characters in the vocabulary. A softmax activation function is used on the output layer to ensure the output has the properties of a probability distribution.

Running this prints a summary of the defined network as a sanity check.

The model is learning a multi-class classification problem, therefore we use the categorical log loss intended for this type of problem. The efficient Adam implementation of gradient descent is used to optimize the model and accuracy is reported at the end of each batch update.

The model is fit for 100 training epochs, again found with a little trial and error.

Save Model

After the model is fit, we save it to file for later use.

The Keras model API provides the save() function that we can use to save the model to a single file, including weights and topology information.

We also save the mapping from characters to integers that we will need to encode any input when using the model and decode any output from the model.

Complete Example

Tying all of this together, the complete code listing for fitting the character-based neural language model is listed below.

Running the example might take one minute.

Note: Your results may vary given the stochastic nature of the algorithm or evaluation procedure, or differences in numerical precision. Consider running the example a few times and compare the average outcome.

You will see that the model learns the problem well, perhaps too well for generating surprising sequences of characters.

At the end of the run, you will have two files saved to the current working directory, specifically model.h5 and mapping.pkl.

Next, we can look at using the learned model.

Generate Text

We will use the learned language model to generate new sequences of text that have the same statistical properties.

Load Model

The first step is to load the model saved to the file ‘model.h5‘.

We can use the load_model() function from the Keras API.

We also need to load the pickled dictionary for mapping characters to integers from the file ‘mapping.pkl‘. We will use the Pickle API to load the object.

We are now ready to use the loaded model.

Generate Characters

We must provide sequences of 10 characters as input to the model in order to start the generation process. We will pick these manually.

A given input sequence will need to be prepared in the same way as preparing the training data for the model.

First, the sequence of characters must be integer encoded using the loaded mapping.

Next, the sequences need to be one hot encoded using the to_categorical() Keras function.

We can then use the model to predict the next character in the sequence.

We use predict_classes() instead of predict() to directly select the integer for the character with the highest probability instead of getting the full probability distribution across the entire set of characters.

We can then decode this integer by looking up the mapping to see the character to which it maps.

This character can then be added to the input sequence. We then need to make sure that the input sequence is 10 characters by truncating the first character from the input sequence text.

We can use the pad_sequences() function from the Keras API that can perform this truncation operation.

Putting all of this together, we can define a new function named generate_seq() for using the loaded model to generate new sequences of text.

Complete Example

Tying all of this together, the complete example for generating text using the fit neural language model is listed below.

Running the example generates three sequences of text.

Note: Your results may vary given the stochastic nature of the algorithm or evaluation procedure, or differences in numerical precision. Consider running the example a few times and compare the average outcome.

The first is a test to see how the model does at starting from the beginning of the rhyme. The second is a test to see how well it does at beginning in the middle of a line. The final example is a test to see how well it does with a sequence of characters never seen before.

We can see that the model did very well with the first two examples, as we would expect. We can also see that the model still generated something for the new text, but it is nonsense.


This section lists some ideas for extending the tutorial that you may wish to explore.

  • Padding. Update the example to provides sequences line by line only and use padding to fill out each sequence to the maximum line length.
  • Sequence Length. Experiment with different sequence lengths and see how they impact the behavior of the model.
  • Tune Model. Experiment with different model configurations, such as the number of memory cells and epochs, and try to develop a better model for fewer resources.

Further Reading

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


In this tutorial, you discovered how to develop a character-based neural language model.

Specifically, you learned:

  • How to prepare text for character-based language modeling.
  • How to develop a character-based language model using LSTMs.
  • How to use a trained character-based language model to generate text.

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

Develop Deep Learning models for Text Data Today!

Deep Learning for Natural Language Processing

Develop Your Own Text models in Minutes

...with just a few lines of python code

Discover how in my new Ebook:
Deep Learning for Natural Language Processing

It provides self-study tutorials on topics like:
Bag-of-Words, Word Embedding, Language Models, Caption Generation, Text Translation and much more...

Finally Bring Deep Learning to your Natural Language Processing Projects

Skip the Academics. Just Results.

See What's Inside

86 Responses to How to Develop a Character-Based Neural Language Model in Keras

  1. Avatar
    Prakash November 7, 2017 at 12:33 am #

    Hi Jason – Thanks for sharing this article. I am in learning phase and when I try to run your program (defining the load_doc function), I am getting error. Is there any package that I need to install before I run the code ?

    Running the first set of lines for loading the doc into memory gives me the following error

    > return text
    Error: unexpected symbol in ” return text”

    • Avatar
      Jason Brownlee November 7, 2017 at 9:51 am #

      It looks like a copy-paste error, ensure you maintain the indenting of the Python code.

  2. Avatar
    Klaas November 8, 2017 at 6:54 am #

    Thanks a lot Jason. One general question. On your blog I read a lot about one hot encoding. From the mnist dataset I get it that it is easy to compare probabilities (e.g if the Number is 2 I want my network to output a 1 on the 3rd row). But when it comes to language huge vocabularies is a one hot encoding not completely inefficient? I mean e.g. 1 Million vocab size and each word a vector with one 1 and 999.999 zeros? I do not really get that.

    • Avatar
      Jason Brownlee November 8, 2017 at 9:31 am #

      Yep, in general we try to reduce the size of the vocab to ensure the model trains in a reasonable time.

      But what is the alternative? Perhaps less crisp word predictions and worse skill?

      • Avatar
        Thomas Shorts February 17, 2021 at 3:28 pm #

        For the “token by word” issue that Klass mentions (having a very wide X data because tokenizing by word yields many more classes than if tokenized by character), would it work to simply feed the model the sequence of embeddings as opposed to a one-hot-encoded sequence where only one of the rows contains the embedding for each word / token?

        You would have to change the input layer around from the example, but would that approach work and alleviate the issues associated with high dimensionality?

  3. Avatar
    Ravi Annaswamy November 10, 2017 at 6:47 am #

    Yet another excellent tutorial, Dr.Jason.

  4. Avatar
    srihari November 10, 2017 at 3:09 pm #


    Can we implement using nltk as helping library to keras, in transforming the text.


  5. Avatar
    Stuart November 11, 2017 at 6:24 am #

    Awesome article. Really appreciate the level of line-by-line detail.

    I think there are some mistakes around this part of the article:

    “Next, the integers need to be one hot encoded using the pad_sequences() Keras function.”

    I think you mean truncated instead of one hot encoded? Also it’s missing the accompanying code snippet.

    • Avatar
      Jason Brownlee November 11, 2017 at 9:28 am #

      Thanks, fixed. I meant the to_categorical() function for one hot encoding.

  6. Avatar
    Antonio November 24, 2017 at 7:40 am #

    Cool, thanks for sharing!

    • Avatar
      Jason Brownlee November 24, 2017 at 9:52 am #

      You’re welcome.

      • Avatar
        Antonio November 24, 2017 at 7:45 pm #

        Quick question, if I may.. if we want to characterize the sequence with some extra input features, how to we prepare the data? Just to illustrate, sticking to the above example, for example I may want to associate to each sequence the name of the Person writing the sequence. This feature may change slightly the prediction of the next character. I thought one option would be to create and train a different model for each Person, but I think this would be quite suboptimal, since majority of the rules learned will be in common to every Person and the data set will be reduced for each model/Person. Another option would be to encode the Person in the input, within each time step. But in this case there is a little bit of redundancy, since the Person input feature will be the same across all the time steps. So, is there a way to provide as input to the model a feature which is independent and unchanged across all the time steps, but which characterize the entire input sequence? Thanks very much

  7. Avatar
    Ethan B January 10, 2018 at 5:05 pm #

    Hi Jason, thanks for the great article!

    I have one question regarding the training phase. I was thinking about using character embeddings, for example fitting a word2vec model on characters which I would then use to train the LSTM, rather than using the one hot encoded characters. Do you think this would give any sort of performance gain? I was going to test this idea myself, but I was curious if you had tried this yourself first, or if you think it is a worthwhile approach.

    • Avatar
      Jason Brownlee January 11, 2018 at 5:49 am #

      I have not tried an embedding of chars, sorry.

    • Avatar
      shm June 15, 2019 at 5:03 am #

      hi Ethan B, have you fit word2vec model on characters??

  8. Avatar
    Al February 2, 2018 at 10:52 am #

    Hi, It’s really a nice tutorial!!
    I have one trouble. When I try to predict using generate_seq, I got this error ValueError: Error when checking : expected lstm_1_input to have shape (408, 37) but got array with shape (10, 37)
    why this would happen? Thanks!!!

    • Avatar
      Riya John February 28, 2018 at 1:54 am #

      Hi, I got a similar error: ValueError: cannot reshape array of size 380 into shape (1,1,10)

      Code worked for me when I commented line: encoded = encoded.reshape(1, encoded.shape[0], encoded.shape[1]) in generate_seq()

      • Avatar
        Franco Arda June 5, 2018 at 4:23 pm #

        @Riiya, thanks for pointing that out.
        @Jason, AI and Riya are right. Code doesn’t run (your blog or book). We need to

        #encoded = encoded.reshape(1, encoded.shape[0], encoded.shape[1])

        to make prediction work.

        • Avatar
          Jason Brownlee June 6, 2018 at 6:36 am #

          Are you able to confirm that you have the latest version of Keras and other libraries installed?

          • Avatar
            Franco Arda June 7, 2018 at 3:32 pm #

            Update now, but issue remains.
            Don’t answer for me – I’m happy that it runs.

          • Avatar
            Kay-Michael Würzner June 21, 2018 at 5:36 pm #

            There is definitely a problem with that reshape line:

            ValueError: cannot reshape array of size 780 into shape (1,1,10)

            Checked the versions of installed libraries. As far as I can see, keras, numpy and tensorflow are needed here.


            Commenting the reshape code indeed helps successfully running the code but given the output it looks like it screws up the actual prediction… Any ideas?

          • Avatar
            Jason Brownlee June 22, 2018 at 6:02 am #

            I wonder if is an issue with the data file, are you able to confirm the raw data for the text matches the post?

    • Avatar
      Franco Arda June 5, 2018 at 4:20 pm #

      indeed, there’s an error ….see below

  9. Avatar
    Neeraj April 7, 2018 at 2:26 pm #

    Hi Jason,

    Can you please help me with an error. I am new at python so many times I dont know how to resolve an error.

    Error 1 :

    mapping = load(open(‘mapping.pkl’, ‘rb’))
    Traceback (most recent call last):

    File “”, line 1, in
    mapping = load(open(‘mapping.pkl’, ‘rb’))

    NameError: name ‘load’ is not defined

    Error 2 :
    from pickle import load

    mapping = load(open(‘mapping.pkl’, ‘rb’))
    Traceback (most recent call last):

    File “”, line 1, in
    mapping = load(open(‘mapping.pkl’, ‘rb’))

    EOFError: Ran out of input


    • Avatar
      Jason Brownlee April 8, 2018 at 6:10 am #

      Perhaps double check that you have coped all of the code from the example?

  10. Avatar
    sagar June 2, 2018 at 1:23 am #

    Hi Jason, this is very helpful and nicely detailed. Thank you for sharing.

    I am working on a problem where I have some 32000 rows of jumbled characters “wewlsfnskfddsl…eredsda” and each row is of length 406. These are hashed, probably. And I need to predict to which class do they belong to? Here class is 1-12 names of books.

    Any suggestions on how I could modify your code above. Would my problem still need text generation? As this is a multi-class classification problem.

    Thank you very much. Looking forward to your advice.

    • Avatar
      Jason Brownlee June 2, 2018 at 6:38 am #

      Sounds like classification. A language mode/text generation would not be helpful.

      I would recommend testing an MLP, CNN and LSTM on the problem. Also look at some of the tutorials on the blog for sentiment classification, they will provide a template. No need for a word embedding either.

    • Avatar
      Koushik June 4, 2018 at 7:52 pm #

      I think many to one RNN model should work for your problem

  11. Avatar
    SM June 4, 2018 at 6:53 pm #

    Hi Jason, thank you for the above suggestions. I am trying to implement a LSTM model and following this post on how to set up. I have a questions about input_shape for LSTM and for dense layer.

    My xtrain is a sequence of numbers as a NumPy array. To give you a background: I have 32514 rows of jumbled characters “wewlsfnskfddsl…” , which I reshaped: X = X.reshape((1,32514,1)) into 3D for in LSTM’s input_shape, taking inspiration from your post “”.

    However, here my y is of 12 classes (0 to 11). It is a multiclass classification problem.

    y = to_categorical(y, num_classes=12)

    How should I define my dense layer? According to keras document, here’s what the input_shape is for dense and LSTM layers:

    Dense: (batch_size, input_dim)
    LSTM: (batch_size, timesteps, input_dim)

    # define model
    model = Sequential()
    model.add(LSTM(75, input_shape=(32514,1)))
    model.add(Dense(input(1), activation=’softmax’))
    # compile model
    model.compile(loss=’categorical_crossentropy’, optimizer=’adam’, metrics=[‘accuracy’])
    # fit model, y, epochs=100, verbose=2)

    # save the model to file‘model.h5’)
    # save the mapping
    dump(mapping, open(‘mapping.pkl’, ‘wb’))

    When I run the above cell, the program seems to be asking to for another input.

    Is this correct? Thank you very much. I love your blog.

  12. Avatar
    Ahmed Sahlol June 11, 2018 at 12:53 am #

    Many thanks Jason for the topic and the detailed explanation, really looks awesome.
    I have a question: Why u assigned the LSTM hidden layer = 75 memory cells?

    • Avatar
      Jason Brownlee June 11, 2018 at 6:09 am #

      The model was configured with a little trial and error.

  13. Avatar
    Christian June 27, 2018 at 12:14 am #

    Hi Jason,

    Do u know if I can use this technique (character) to cluster rare words, like DriverId with Driver, Name, etc ?. Or Vehicle with vehicleId, location, latitude, longitude?,


    • Avatar
      Jason Brownlee June 27, 2018 at 8:19 am #

      Perhaps you can use an embedding for these features?

  14. Avatar
    Ayden M. June 28, 2018 at 10:11 pm #

    Thank you so much for this tutorial. I am getting this error when training my LM, and I can’t figure out how to overcome this:

    IndexError: too many indices for array

    It occurs in line 39: X, y = sequences[:,:-1], sequences[:,-1]

    From what I understand, it doesn’t want to create a 2D vector from a 1D vector. I tried to reshape with numpy but I keep getting errors of similar nature. Do you have any idea how to solve this?

  15. Avatar
    Tuomas V. September 25, 2018 at 12:33 am #


    Thanks for the incredibly easy-to-follow tutorial.

    I’m wondering whether there’s a more memory-friendly way to one-hot encode than using to_categorical?

    The sequences text file I’m working with is nearly 1 Gb in size, my vocabulary size is quite large and thus there are MemoryErrors at:

    sequences = [to_categorical(x, num_classes=vocab_size) for x in X]

    Any help would be appreciated!

    • Avatar
      Jason Brownlee September 25, 2018 at 6:25 am #

      Hmmm. Some thoughts:

      Don’t use one hot encoding, use integer encoding and a word embedding and process the file in chunks via progressive loading.

      • Avatar
        Tuomas V. October 10, 2018 at 10:41 pm #

        Droppped one-hot encoding, switched to fastText word vectors and it has been smooth sailing since. Thanks!

  16. Avatar
    Ayan January 10, 2019 at 4:41 pm #

    Hi Jason..I want to view the embedded representation for each character(embedded using LSTM)..Can you please suggest the lines needed to do so?

    • Avatar
      Jason Brownlee January 11, 2019 at 7:40 am #

      Sorry, I don’t have any examples of character based embeddings – if there is such a thing.

  17. Avatar
    Fengtao Wu January 19, 2019 at 9:21 am #

    There is a error in the complete example at last. In the line 17:

    “encoded = encoded.reshape(1, encoded.shape[0], encoded.shape[1])”

    should be:

    “encoded = encoded.reshape(-1, encoded.shape[1], encoded.shape[2])”

    Could you please check it again?

    • Avatar
      Jason Brownlee January 20, 2019 at 5:38 am #

      Why is that exactly?

    • Avatar
      Dude July 16, 2019 at 4:40 am #

      I’m not sure why but this change causes the code to work from me. I have been having the same error as others.

      • Avatar
        Jason Brownlee July 16, 2019 at 8:23 am #


        Are all libs up to date? Python 3? Keras? TensorFlow?

  18. Avatar
    Mahnaz February 28, 2019 at 10:08 pm #

    I am using the char-level language model to predict the next character given a sequence of previous characters (say 20 characters). My train sequences are almost 1088637. I am using the following model for my training:

    model = Sequential()
    model.add(LSTM(1000, input_shape=(X.shape[1], X.shape[2])))
    model.add(Dense(vocab_size, activation=’softmax’, kernel_initializer=’normal’))

    But I get only 50% of the text correct. It always goes wrong from the middle of the text. Any suggestion? I appreciate your help.

  19. Avatar
    Nauman January 26, 2020 at 11:31 pm #

    Dear is there any tutorial for character-based machine translation?

  20. Avatar
    Efstathios Chatzikyriakidis May 15, 2020 at 2:05 am #

    Hi Jason,

    In your word-based language model examples you presented us:

    1. One-Word-In -> One-Word-Out framing
    2. N-Words-In -> One-Word-Out framing
    3. Line-by-Line framing

    Also for the character-based language model you use here:

    1. N-Char-In -> One-Char-Out
    2. Also suggest the usage of Line-by-line as an experimentation

    However, you can also use a different approach for learning a language model which is:


    Example of input and output:

    Hello Jason!-E-
    -S-Hello Jason!

    This can be done using an LSTM with return_sequences=True where input sequence length equals output sequence length.

    Inputs can be encoded as word embeddings and outputs can be encoded as one-hot.


    • Avatar
      Jason Brownlee May 15, 2020 at 6:04 am #

      Yes, but seq2seq requires an encoder-decoder model, not simply a return sequences from an LSTM.

  21. Avatar
    Shankar May 20, 2020 at 6:13 pm #

    A very basic doubt, sorry since I’m a beginner in this stuff… On what basis do you declare X and Y to be sequences[,:-1] and [:,-1]?

    • Avatar
      Jason Brownlee May 21, 2020 at 6:13 am #

      What do you mean? Perhaps you can elaborate on your doubt?

  22. Avatar
    Pablo June 11, 2020 at 4:55 am #


    Thanks for this! I’m trying to run the example but there is seems to be an error in line
    X, y = sequences[:,:-1], sequences[:,-1]

    It turns out, sequences is first a list of values later converted to an array. This array is not multidimensional and hence the error in the above quoted line. I tried to fix this by replacing that line with

    X, y = sequences[:-1], sequences[-1]

    which makes sense in a single-dimension array, but now throws another error in the following line:

    model.add(LSTM(75, input_shape=(X.shape[1], X.shape[2])))

    I understand X must be of a different shape and that’s why in this last quoted line it is inteded to provide two integers from the two dimensions.

    Any idea how to fix this?

    • Avatar
      Jason Brownlee June 11, 2020 at 6:05 am #

      Thanks. Are you sure, as I believe the example runs as expected without change:

      • Avatar
        Pablo June 11, 2020 at 5:08 pm #

        Hi, thanks for such a quick answer!

        You are right, I didn’t copied the code just right.
        I just did, nonetheless and got this error:

        Traceback (most recent call last):
        File “”, line 37, in
        print(generate_seq(model, mapping, 10, ‘Sing a son’, 20))
        File “”, line 18, in generate_seq
        encoded = encoded.reshape(1, encoded.shape[0], encoded.shape[1])
        ValueError: cannot reshape array of size 380 into shape (1,1,10)

        while generating text. Commenting line 18 makes it work, but I’m not sure it does as expected. Could you confirm it?

        • Avatar
          Jason Brownlee June 12, 2020 at 6:09 am #

          Are you able to confirm that your version of Keras and TensorFlow are up to date?

          • Avatar
            Pablo June 13, 2020 at 2:21 am #

            HI, thanks for anwering back

            I can confirm I have my system up to date. That means tensorflow package at version 2.2.0, which also provides Keras. Is that ok?

          • Avatar
            Jason Brownlee June 13, 2020 at 6:09 am #

            Your versions look good. I will investigate.

            Update: The example works as-is with TensorFlow 2.2 and Keras 2.3.

            Ensure you copied the complete code example.

  23. Avatar
    Pablo June 15, 2020 at 7:26 pm #

    Hello again Jason,

    First of all thanks for all the trouble you are taking with this.

    I just copied again the three complete codes intro empty files and still get the same error posted above:

    Traceback (most recent call last):
    File “”, line 36, in
    print(generate_seq(model, mapping, 10, ‘Sing a son’, 20))
    File “”, line 17, in generate_seq
    encoded = encoded.reshape(1, encoded.shape[0], encoded.shape[1])
    ValueError: cannot reshape array of size 380 into shape (1,1,10)

    Which I fix just commenting that line.

    My python is version 3.8.3, tensorflow 2.2.0, keras 2.3.0-tf. I have everything installed from official Archlinux repos (packages {,python}tensorflow-opt).

    Maybe there is something OS-related messing with this. To be honest, I don’t think this is worth much trouble. I’m able to create the sequences, train the model and generate the text just by modifying that line. There is our discussion here, so anyone can check this comments and find a fix in case they need it.

    Thank you again,

  24. Avatar
    Firas Obeid September 1, 2020 at 3:45 am #

    Is it necessary to one-hot encode my features if my features(characters) lets say are are much larger and expand to 88 for example? Because that will have a tall on my training time and memory…

    • Avatar
      Jason Brownlee September 1, 2020 at 6:38 am #

      Perhaps try it and see, compare to other encodings like an ordinal encoding and an embedding.

      • Avatar
        Firas Obeid September 1, 2020 at 7:47 am #

        Yes most definitely thanks! Iam also trying converting to bytes ‘utf-8’ to simplify embedding look-up and then inverting back to get readable text after prediction.

  25. Avatar
    Stanislav September 13, 2020 at 9:33 am #

    Hi Jason.
    I want to ask about the metrics in this part of the code.

    model.compile(loss=’categorical_crossentropy’, optimizer=’adam’, metrics=[‘accuracy’])

    As far as I understand, these metrics work on sequence comparisons.

    • Avatar
      Stanislav September 13, 2020 at 9:44 am #

      Hi Jason.
      I want to ask about the metrics in this part of the code.

      model.compile(loss=’categorical_crossentropy’, optimizer=’adam’, metrics=[‘accuracy’])

      As far as I understand, these metrics work on sequence comparisons., y, epochs=30, verbose=2, callbacks=[csv_logger])

      As a result, the final dataset is divided into 2 X and Y. In this case, is it learning without a teacher? And “y” is used only for calculating metrics?

      If I misunderstood correctly, please explain why to separate data into “X” and “y”

      And what type of training this is in the end, I do not quite understand?

      Thanks in advance.

      PS. Sorry for my English

    • Avatar
      Jason Brownlee September 14, 2020 at 6:41 am #

      Cross-entropy is minimized so the model predicts the desired output.

      Accuracy is used to evaluate the output categorical variables, e.g. the expected characters. it may or may not be the most appropriate metric in this case. e.g. if we got 100% the model has memorised the input which might not be desirable.

  26. Avatar
    kaur November 30, 2020 at 7:22 pm #

    One epoch out of 100 taking around 180 seconds with loss 1.3362 and accuracy of 0.3520 for biological sequence file having size 1.59 MB.

    How to reduce it ?

  27. Avatar
    Himanshu December 29, 2020 at 2:20 am #

    Hi Jason,

    I am getting this error on execution of the code:

    ValueError: cannot reshape array of size 380 into shape (1,1,10)

    Could you please advise. Thanks

  28. Avatar
    cnsn8 March 3, 2021 at 6:32 am #

    hello, thanks for good information. I would be very happy if you could look. I take the code exactly and run it as it is. However, I get the following error in the section.

    ValueError: Input 0 of layer sequential_15 is incompatible with the layer: expected ndim=3, found ndim=4. Full shape received: [None, 398, 11, 38]

    How can I fix this ?
    thank you.

  29. Avatar
    cnsn8 April 5, 2021 at 8:31 am #

    Thanks again, I fixed it. I have an another question. Is there any source you can recommend that explains the subject of simple controllable text generation with code? Or would you consider making a tutorial on this topic.

Leave a Reply