Your First Deep Learning Project in Python with Keras Step-by-Step

Last Updated on August 16, 2022

Keras is a powerful and easy-to-use free open source Python library for developing and evaluating deep learning models.

It is part of the TensorFlow library and allows you to define and train neural network models in just a few lines of code.

In this tutorial, you will discover how to create your first deep learning neural network model in Python using Keras.

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

Let’s get started.

  • Update Feb/2017: Updated prediction example, so rounding works in Python 2 and 3.
  • Update Mar/2017: Updated example for the latest versions of Keras and TensorFlow.
  • Update Mar/2018: Added alternate link to download the dataset.
  • Update Jul/2019: Expanded and added more useful resources.
  • Update Sep/2019: Updated for Keras v2.2.5 API.
  • Update Oct/2019: Updated for Keras v2.3.0 API and TensorFlow v2.0.0.
  • Update Aug/2020: Updated for Keras v2.4.3 and TensorFlow v2.3.
  • Update Oct/2021: Deprecated predict_class syntax
  • Update Jun/2022: Updated to modern TensorFlow syntax
Tour of Deep Learning Algorithms

Develop your first neural network in Python with Keras step-by-step
Photo by Phil Whitehouse, some rights reserved.

Keras Tutorial Overview

There is not a lot of code required, but we will go over it slowly so that you will know how to create your own models in the future.

The steps you will learn in this tutorial are as follows:

  1. Load Data
  2. Define Keras Model
  3. Compile Keras Model
  4. Fit Keras Model
  5. Evaluate Keras Model
  6. Tie It All Together
  7. Make Predictions

This Keras tutorial makes a few assumptions. You will need to have:

  1. Python 2 or 3 installed and configured
  2. SciPy (including NumPy) installed and configured
  3. Keras and a backend (Theano or TensorFlow) installed and configured

If you need help with your environment, see the tutorial:

Create a new file called and type or copy-and-paste the code into the file as you go.

Need help with Deep Learning in Python?

Take my free 2-week email course and discover MLPs, CNNs and LSTMs (with code).

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

1. Load Data

The first step is to define the functions and classes you intend to use in this tutorial.

You will use the NumPy library to load your dataset and two classes from the Keras library to define your model.

The imports required are listed below.

You can now load our dataset.

In this Keras tutorial, you will use the Pima Indians onset of diabetes dataset. This is a standard machine learning dataset from the UCI Machine Learning repository. It describes patient medical record data for Pima Indians and whether they had an onset of diabetes within five years.

As such, it is a binary classification problem (onset of diabetes as 1 or not as 0). All of the input variables that describe each patient are numerical. This makes it easy to use directly with neural networks that expect numerical input and output values and is an ideal choice for our first neural network in Keras.

The dataset is available here:

Download the dataset and place it in your local working directory, the same location as your Python file.

Save it with the filename:

Take a look inside the file; you should see rows of data like the following:

You can now load the file as a matrix of numbers using the NumPy function loadtxt().

There are eight input variables and one output variable (the last column). You will be learning a model to map rows of input variables (X) to an output variable (y), which is often summarized as y = f(X).

The variables can be summarized as follows:

Input Variables (X):

  1. Number of times pregnant
  2. Plasma glucose concentration at 2 hours in an oral glucose tolerance test
  3. Diastolic blood pressure (mm Hg)
  4. Triceps skin fold thickness (mm)
  5. 2-hour serum insulin (mu U/ml)
  6. Body mass index (weight in kg/(height in m)^2)
  7. Diabetes pedigree function
  8. Age (years)

Output Variables (y):

  1. Class variable (0 or 1)

Once the CSV file is loaded into memory, you can split the columns of data into input and output variables.

The data will be stored in a 2D array where the first dimension is rows and the second dimension is columns, e.g., [rows, columns].

You can split the array into two arrays by selecting subsets of columns using the standard NumPy slice operator or “:”. You can select the first eight columns from index 0 to index 7 via the slice 0:8. We can then select the output column (the 9th variable) via index 8.

You are now ready to define your neural network model.

Note: The dataset has nine columns, and the range 0:8 will select columns from 0 to 7, stopping before index 8. If this is new to you, then you can learn more about array slicing and ranges in this post:

2. Define Keras Model

Models in Keras are defined as a sequence of layers.

We create a Sequential model and add layers one at a time until we are happy with our network architecture.

The first thing to get right is to ensure the input layer has the correct number of input features. This can be specified when creating the first layer with the input_shape argument and setting it to (8,) for presenting the eight input variables as a vector.

How do we know the number of layers and their types?

This is a tricky question. There are heuristics that you can use, and often the best network structure is found through a process of trial and error experimentation (I explain more about this here). Generally, you need a network large enough to capture the structure of the problem.

In this example, let’s use a fully-connected network structure with three layers.

Fully connected layers are defined using the Dense class. You can specify the number of neurons or nodes in the layer as the first argument and the activation function using the activation argument.

Also, you will use the rectified linear unit activation function referred to as ReLU on the first two layers and the Sigmoid function in the output layer.

It used to be the case that Sigmoid and Tanh activation functions were preferred for all layers. These days, better performance is achieved using the ReLU activation function. Using a sigmoid on the output layer ensures your network output is between 0 and 1 and is easy to map to either a probability of class 1 or snap to a hard classification of either class with a default threshold of 0.5.

You can piece it all together by adding each layer:

  • The model expects rows of data with 8 variables (the input_shape=(8,) argument).
  • The first hidden layer has 12 nodes and uses the relu activation function.
  • The second hidden layer has 8 nodes and uses the relu activation function.
  • The output layer has one node and uses the sigmoid activation function.

Note:  The most confusing thing here is that the shape of the input to the model is defined as an argument on the first hidden layer. This means that the line of code that adds the first Dense layer is doing two things, defining the input or visible layer and the first hidden layer.

3. Compile Keras Model

Now that the model is defined, you can compile it.

Compiling the model uses the efficient numerical libraries under the covers (the so-called backend) such as Theano or TensorFlow. The backend automatically chooses the best way to represent the network for training and making predictions to run on your hardware, such as CPU, GPU, or even distributed.

When compiling, you must specify some additional properties required when training the network. Remember training a network means finding the best set of weights to map inputs to outputs in your dataset.

You must specify the loss function to use to evaluate a set of weights, the optimizer used to search through different weights for the network, and any optional metrics you want to collect and report during training.

In this case, use cross entropy as the loss argument. This loss is for a binary classification problems and is defined in Keras as “binary_crossentropy“. You can learn more about choosing loss functions based on your problem here:

We will define the optimizer as the efficient stochastic gradient descent algorithm “adam“. This is a popular version of gradient descent because it automatically tunes itself and gives good results in a wide range of problems. To learn more about the Adam version of stochastic gradient descent, see the post:

Finally, because it is a classification problem, you will collect and report the classification accuracy defined via the metrics argument.

4. Fit Keras Model

You have defined your model and compiled it to get ready for efficient computation.

Now it is time to execute the model on some data.

You can train or fit your model on your loaded data by calling the fit() function on the model.

Training occurs over epochs, and each epoch is split into batches.

  • Epoch: One pass through all of the rows in the training dataset
  • Batch: One or more samples considered by the model within an epoch before weights are updated

One epoch comprises one or more batches, based on the chosen batch size, and the model is fit for many epochs. For more on the difference between epochs and batches, see the post:

The training process will run for a fixed number of epochs (iterations) through the dataset that you must specify using the epochs argument. You must also set the number of dataset rows that are considered before the model weights are updated within each epoch, called the batch size, and set using the batch_size argument.

This problem will run for a small number of epochs (150) and use a relatively small batch size of 10.

These configurations can be chosen experimentally by trial and error. You want to train the model enough so that it learns a good (or good enough) mapping of rows of input data to the output classification. The model will always have some error, but the amount of error will level out after some point for a given model configuration. This is called model convergence.

This is where the work happens on your CPU or GPU.

No GPU is required for this example, but if you’re interested in how to run large models on GPU hardware cheaply in the cloud, see this post:

5. Evaluate Keras Model

You have trained our neural network on the entire dataset, and you can evaluate the performance of the network on the same dataset.

This will only give you an idea of how well you have modeled the dataset (e.g., train accuracy), but no idea of how well the algorithm might perform on new data. This was done for simplicity, but ideally, you could separate your data into train and test datasets for training and evaluation of your model.

You can evaluate your model on your training dataset using the evaluate() function and pass it the same input and output used to train the model.

This will generate a prediction for each input and output pair and collect scores, including the average loss and any metrics you have configured, such as accuracy.

The evaluate() function will return a list with two values. The first will be the loss of the model on the dataset, and the second will be the accuracy of the model on the dataset. You are only interested in reporting the accuracy so ignore the loss value.

6. Tie It All Together

You have just seen how you can easily create your first neural network model in Keras.

Let’s tie it all together into a complete code example.

You can copy all the code into your Python file and save it as “” in the same directory as your data file “pima-indians-diabetes.csv“. You can then run the Python file as a script from your command line (command prompt) as follows:

Running this example, you should see a message for each of the 150 epochs, printing the loss and accuracy, followed by the final evaluation of the trained model on the training dataset.

It takes about 10 seconds to execute on my workstation running on the CPU.

Ideally, you would like the loss to go to zero and the accuracy to go to 1.0 (e.g., 100%). This is not possible for any but the most trivial machine learning problems. Instead, you will always have some error in your model. The goal is to choose a model configuration and training configuration that achieve the lowest loss and highest accuracy possible for a given dataset.

Note: If you try running this example in an IPython or Jupyter notebook, you may get an error.

The reason is the output progress bars during training. You can easily turn these off by setting verbose=0 in the call to the fit() and evaluate() functions; for example:

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.

What score did you get?
Post your results in the comments below.

Neural networks are stochastic algorithms, meaning that the same algorithm on the same data can train a different model with different skill each time the code is run. This is a feature, not a bug. You can learn more about this in the post:

The variance in the performance of the model means that to get a reasonable approximation of how well your model is performing, you may need to fit it many times and calculate the average of the accuracy scores. For more on this approach to evaluating neural networks, see the post:

For example, below are the accuracy scores from re-running the example five times:

You can see that all accuracy scores are around 77%, and the average is 76.924%.

7. Make Predictions

The number one question I get asked is:

“After I train my model, how can I use it to make predictions on new data?”

Great question.

You can adapt the above example and use it to generate predictions on the training dataset, pretending it is a new dataset you have not seen before.

Making predictions is as easy as calling the predict() function on the model. You are using a sigmoid activation function on the output layer, so the predictions will be a probability in the range between 0 and 1. You can easily convert them into a crisp binary prediction for this classification task by rounding them.

For example:

Alternately, you can convert the probability into 0 or 1 to predict crisp classes directly; for example:

The complete example below makes predictions for each example in the dataset, then prints the input data, predicted class, and expected class for the first five examples in the dataset.

Running the example does not show the progress bar as before, as the verbose argument has been set to 0.

After the model is fit, predictions are made for all examples in the dataset, and the input rows and predicted class value for the first five examples is printed and compared to the expected class value.

You can see that most rows are correctly predicted. In fact, you can expect about 76.9% of the rows to be correctly predicted based on your estimated performance of the model in the previous section.

If you would like to know more about how to make predictions with Keras models, see the post:

Keras Tutorial Summary

In this post, you discovered how to create your first neural network model using the powerful Keras Python library for deep learning.

Specifically, you learned the six key steps in using Keras to create a neural network or deep learning model step-by-step, including:

  1. How to load data
  2. How to define a neural network in Keras
  3. How to compile a Keras model using the efficient numerical backend
  4. How to train a model on data
  5. How to evaluate a model on data
  6. How to make predictions with the model

Do you have any questions about Keras or about this tutorial?
Ask your question in the comments, and I will do my best to answer.

Keras Tutorial Extensions

Well done, you have successfully developed your first neural network using the Keras deep learning library in Python.

This section provides some extensions to this tutorial that you might want to explore.

  • Tune the Model. Change the configuration of the model or training process and see if you can improve the performance of the model, e.g., achieve better than 76% accuracy.
  • Save the Model. Update the tutorial to save the model to a file, then load it later and use it to make predictions (see this tutorial).
  • Summarize the Model. Update the tutorial to summarize the model and create a plot of model layers (see this tutorial).
  • Separate, Train, and Test Datasets. Split the loaded dataset into a training and test set (split based on rows) and use one set to train the model and the other set to estimate the performance of the model on new data.
  • Plot Learning Curves. The fit() function returns a history object that summarizes the loss and accuracy at the end of each epoch. Create line plots of this data, called learning curves (see this tutorial).
  • Learn a New Dataset. Update the tutorial to use a different tabular dataset, perhaps from the UCI Machine Learning Repository.
  • Use Functional API. Update the tutorial to use the Keras Functional API for defining the model (see this tutorial).

Further Reading

Are you looking for some more Deep Learning tutorials with Python and Keras?

Take a look at some of these:

Related Tutorials



How did you go? Do you have any questions about deep learning?
Post your questions in the comments below, and I will do my best to help.

1,162 Responses to Your First Deep Learning Project in Python with Keras Step-by-Step

  1. Saurav May 27, 2016 at 11:08 pm #

    The input layer doesn’t have any activation function, but still activation=”relu” is mentioned in the first layer of the model. Why?

    • Jason Brownlee May 28, 2016 at 6:32 am #

      Hi Saurav,

      The first layer in the network here is technically a hidden layer, hence it has an activation function.

      • sam Johnson December 21, 2016 at 2:44 am #

        Why have you made it a hidden layer though? the input layer is not usually represented as a hidden layer?

        • Jason Brownlee December 21, 2016 at 8:41 am #

          Hi sam,

          Note this line:

          It does a few things.

          • It defines the input layer as having 8 inputs.
          • It defines a hidden layer with 12 neurons, connected to the input layer that use relu activation function.
          • It initializes all weights using a sample of uniform random numbers.

          Does that help?

          • Pavidevi May 17, 2017 at 2:31 am #

            Hi Jason,

            U have used two different activation functions so how can we know which activation function fit the model?

          • Jason Brownlee May 17, 2017 at 8:38 am #

            Sorry, I don’t understand the question.

          • Marco Cheung August 23, 2017 at 12:51 am #

            Hi Jason,

            I am interested in deep learning and machine learning. You mentioned “It defines a hidden layer with 12 neurons, connected to the input layer that use relu activation function.” I wonder how can we determine the number of neurons in order to achieve a high accuracy rate of the model?

            Thanks a lot!!!

          • Jason Brownlee August 23, 2017 at 6:55 am #

            Use trial and error. We cannot specify the “best” number of neurons analytically. We must test.

          • Ramzan Shahid November 10, 2017 at 4:32 am #

            Sir, thanks for your tutorial. Would you like to make tutorial on stock Data Prediction through Neural Network Model and training this on any stock data. If you have on this so please share the link. Thanks

          • Jason Brownlee November 10, 2017 at 10:39 am #

            I am reticent to post tutorials on stock market prediction given the random walk hypothesis of security prices:

          • Dhara Bhavsar August 28, 2019 at 9:54 pm #


            I would like to know more about activation function. How it is working? How many activation functions? Using different activation function How much affect the output of the model?

            I would like to also know about the Hidden Layer. How the size of the hidden layer affect the model?

          • Jason Brownlee August 29, 2019 at 6:09 am #

            In this tutorial, we use relu in the hidden layers, learn more here:

            The size of the layer impacts the capacity of the model, learn more here:

      • dhani June 28, 2018 at 2:44 am #

        hi how use cnn for pixel classification on mhd images

        • Jason Brownlee June 28, 2018 at 6:22 am #

          What is pixel classification? What are mhd images?

      • Tanmay Kulkarni February 11, 2020 at 5:50 am #

        Hello! I want to know if there’s a way to know the values of all weights after each updation?

    • BlackBookKeeper August 18, 2018 at 10:15 pm #

      runfile(‘C:/Users/Owner/Documents/’, wdir=’C:/Users/Owner/Documents’)
      Traceback (most recent call last):

      File “”, line 1, in
      runfile(‘C:/Users/Owner/Documents/’, wdir=’C:/Users/Owner/Documents’)

      File “C:\Users\Owner\Anaconda3\lib\site-packages\spyder\utils\site\”, line 705, in runfile
      execfile(filename, namespace)

      File “C:\Users\Owner\Anaconda3\lib\site-packages\spyder\utils\site\”, line 102, in execfile
      exec(compile(, filename, ‘exec’), namespace)

      File “C:/Users/Owner/Documents/”, line 13, in
      model.add(Dense(12, input_dim=8, activation=’relu’))

      File “C:\Users\Owner\Anaconda3\lib\site-packages\keras\engine\”, line 160, in add + ‘_input’)

      File “C:\Users\Owner\Anaconda3\lib\site-packages\keras\engine\”, line 177, in Input

      File “C:\Users\Owner\Anaconda3\lib\site-packages\keras\legacy\”, line 91, in wrapper
      return func(*args, **kwargs)

      File “C:\Users\Owner\Anaconda3\lib\site-packages\keras\engine\”, line 86, in __init__

      File “C:\Users\Owner\Anaconda3\lib\site-packages\keras\backend\”, line 515, in placeholder
      x = tf.placeholder(dtype, shape=shape, name=name)

      File “C:\Users\Owner\AppData\Roaming\Python\Python36\site-packages\tensorflow\python\ops\”, line 1530, in placeholder
      return gen_array_ops._placeholder(dtype=dtype, shape=shape, name=name)

      File “C:\Users\Owner\AppData\Roaming\Python\Python36\site-packages\tensorflow\python\ops\”, line 1954, in _placeholder

      File “C:\Users\Owner\AppData\Roaming\Python\Python36\site-packages\tensorflow\python\framework\”, line 767, in apply_op

      File “C:\Users\Owner\AppData\Roaming\Python\Python36\site-packages\tensorflow\python\framework\”, line 2508, in create_op

      File “C:\Users\Owner\AppData\Roaming\Python\Python36\site-packages\tensorflow\python\framework\”, line 1894, in set_shapes_for_outputs

      File “C:\Users\Owner\AppData\Roaming\Python\Python36\site-packages\tensorflow\python\framework\”, line 443, in set_shape
      self._shape = self._shape.merge_with(shape)

      File “C:\Users\Owner\AppData\Roaming\Python\Python36\site-packages\tensorflow\python\framework\”, line 550, in merge_with
      stop = key.stop

      File “C:\Users\Owner\AppData\Roaming\Python\Python36\site-packages\tensorflow\python\framework\”, line 798, in as_shape
      “””Returns this shape as a TensorShapeProto.”””

      File “C:\Users\Owner\AppData\Roaming\Python\Python36\site-packages\tensorflow\python\framework\”, line 431, in __init__
      size for one or more dimension. e.g. TensorShape([None, 256])

      File “C:\Users\Owner\AppData\Roaming\Python\Python36\site-packages\tensorflow\python\framework\”, line 376, in as_dimension
      other = as_dimension(other)

      File “C:\Users\Owner\AppData\Roaming\Python\Python36\site-packages\tensorflow\python\framework\”, line 32, in __init__
      if value is None:

      TypeError: int() argument must be a string, a bytes-like object or a number, not ‘TensorShapeProto’

      this error occurs when {model.add(Dense(12, input_dim=8, activation=’relu’))} this command is run

      any help?

    • Penchalaiah December 8, 2019 at 6:24 pm #

      Fantastic tutorial. The explanation is simple and precise. Thanks a lot

    • Loc June 29, 2022 at 1:00 pm #

      great arttist

  2. Geoff May 29, 2016 at 6:18 am #

    Can you explain how to implement weight regularization into the layers?

  3. KWC June 14, 2016 at 12:08 pm #

    Import statements if others need them:

    from keras.models import Sequential
    from keras.layers import Dense, Activation

  4. Aakash Nain June 29, 2016 at 6:00 pm #

    If there are 8 inputs for the first layer then why we have taken them as ’12’ in the following line :

    model.add(Dense(12, input_dim=8, init=’uniform’, activation=’relu’))

    • Jason Brownlee June 30, 2016 at 6:47 am #

      Hi Aakash.

      The input layer is defined by the input_dim parameter, here set to 8.

      The first hidden layer has 12 neurons.

  5. Joshua July 2, 2016 at 12:04 am #

    I ran your program and i have an error:
    ValueError: could not convert string to float:
    what could be the reason for this, and how may I solve it.
    great post by the way.

    • Jason Brownlee July 2, 2016 at 6:20 am #

      It might be a copy-paste error. Perhaps try to copy and run the whole example listed in section 6?

    • KeyChy July 3, 2019 at 5:45 pm #

      Maybe when you set all parameters in an extra column in your *.csv file. Than you schould replace the delimiter from , to ; like:
      dataset = numpy.loadtxt(“pima-indians-diabetes.csv”, delimiter=”;”)
      This solved the Problem for me.

  6. cheikh brahim July 5, 2016 at 7:40 pm #

    thank you for your simple and useful example.

  7. Nikhil Thakur July 6, 2016 at 6:39 pm #

    Hello Sir, I am trying to use Keras for NLP , specifically sentence classification. I have given the model building part below. It’s taking quite a lot time to execute. I am using Pycharm IDE.

    batch_size = 32
    nb_filter = 250
    filter_length = 3
    nb_epoch = 2
    pool_length = 2
    output_dim = 5
    hidden_dims = 250

    # Build the model

    model1 = Sequential()

    model1.add(Convolution1D(nb_filter, filter_length ,activation=’relu’,border_mode=’valid’,
    input_shape=(len(embb_weights),dim), weights=[embb_weights]))



    model1.add(Dense(output_dim, activation=’sigmoid’))

    sgd = SGD(lr=0.1, decay=1e-6, momentum=0.9, nesterov=True)


  8. Andre Norman July 15, 2016 at 10:40 am #

    Hi Jason, thanks for the awesome example. Given that the accuracy of this model is 79.56%. From here on, what steps would you take to improve the accuracy?

    Given my nascent understanding of Machine Learning, my initial approach would have been:

    Implement forward propagation, then compute the cost function, then implement back propagation, use gradient checking to evaluate my network (disable after use), then use gradient descent.

    However, this approach seems arduous compared to using Keras. Thanks for your response.

    • Jason Brownlee July 15, 2016 at 10:52 am #

      Hi Andre, indeed Keras makes working with neural nets so much easier. Fun even!

      We may be maxing out on this problem, but here is some general advice for lifting performance.
      – data prep – try lots of different views of the problem and see which is best at exposing the structure of the problem to the learning algorithm (data transforms, feature engineering, etc.)
      – algorithm selection – try lots of algorithms and see which one or few are best on the problem (try on all views)
      – algorithm tuning – tune well performing algorithms to get the most out of them (grid search or random search hyperparameter tuning)
      – ensembles – combine predictions from multiple algorithms (stacking, boosting, bagging, etc.)

      For neural nets, there are a lot of things to tune, I think there are big gains in trying different network topologies (layers and number of neurons per layer) in concert with training epochs and learning rate (bigger nets need more training).

      I hope that helps as a start.

      • Andre Norman July 18, 2016 at 7:19 am #

        Awesome! Thanks Jason =)

    • quentin August 7, 2017 at 8:41 pm #

      Some interesting stuff here

  9. Romilly Cocking July 21, 2016 at 12:31 am #

    Hi Jason, it’s a great example but if anyone runs it in an IPython/Jupyter notebook they are likely to encounter an I/O error when running the fit step. This is due to a known bug in IPython.

    The solution is to set verbose=0 like this

    # Fit the model, Y, nb_epoch=40, batch_size=10, verbose=0)

  10. Anirban July 23, 2016 at 10:20 pm #

    Great example. Have a query though. How do I now give a input and get the output (0 or 1). Can you pls give the cmd for that.

    • Jason Brownlee July 24, 2016 at 6:53 am #

      You can call model.predict() to get predictions and round on each value to snap to a binary value.

      For example, below is a complete example showing you how to round the predictions and print them to console.

      • Debanjan March 27, 2017 at 12:04 pm #

        Hi, Why you are not using any test set? You are predicting from the training set , I think.

        • Jason Brownlee March 28, 2017 at 8:19 am #

          Correct, it is just an example to get you started with Keras.

      • David June 26, 2017 at 12:24 am #

        Jason, I’m not quite understanding how the predicted values ([1.0, 0.0, 1.0, 0.0, 1.0,…) map to the real world problem. For instance, what does that first “1.0” in the results indicate?

        I get that it’s a prediction of ‘true’ for diabetes…but to which patient is it predicting that—the first in the list? So then the second result, “0.0,” is the prediction for the second patient/row in the dataset?

        • Jason Brownlee June 26, 2017 at 6:08 am #

          Remember the original file has 0 and 1 values in the final class column where 0 is no onset of diabetes and 1 is an onset of diabetes.

          We are predicting new values in this column.

          We are making predictions for special rows, we pass in their medical info and predict the onset of diabetes. We just happen to do this for a number of rows at a time.

          • ami July 16, 2018 at 4:30 pm #

            hello jason

            i am getting this error while calculating the predictions.

            #calculate predictions

            predictions = model.predict(X)

            #round predictions

            rounded = [round(x) for x in predictions]


            TypeError Traceback (most recent call last)
            in ()
            2 predictions = model.predict(X)
            3 #round predictions
            —-> 4 rounded = [round(x) for x in predictions]
            5 print(rounded)

            in (.0)
            2 predictions = model.predict(X)
            3 #round predictions
            —-> 4 rounded = [round(x) for x in predictions]
            5 print(rounded)

            TypeError: type numpy.ndarray doesn’t define __round__ method

          • Jason Brownlee July 17, 2018 at 6:09 am #

            Try removing the call to round().

      • Rachel June 28, 2017 at 8:28 pm #

        Hi Jason,
        Can I ask why you use the same data X you fit the model to do the prediction?

        # Fit the model, Y, epochs = 150, batch_size = 10, verbose = 2)

        # calculate predictions
        predictions = model.predict(X)


        • Jason Brownlee June 29, 2017 at 6:34 am #

          It is all I have at hand. X means data matrix.

          Replace X in predict() with Xprime or whatever you like.

      • jitendra March 27, 2018 at 7:20 pm #

        hii, how will i feed the input (8,125,96,0,0,0.0,0.232,54) to get our output.

        predictions = model.predict(X)
        i mean insead of X i want to get output of 8,125,96,0,0,0.0,0.232,54.

        • Jason Brownlee March 28, 2018 at 6:24 am #

          Wrap your input in an array, n-columns with one row, then pass that to the model.

          Does that help?

          • Roman October 5, 2018 at 11:22 pm #

            Hello, trying to use predictions on similar neural network but keep getting errors that input dimension has other shape.

            Can you say how array must look on exampled neural network?

          • Jason Brownlee October 6, 2018 at 5:45 am #

            For an MLP, data must be organized into a 2d array of samples x features

  11. Anirban July 23, 2016 at 10:52 pm #

    I am not able to get to the last epoch. Getting error before that:
    Epoch 11/150
    390/768 [==============>……………]Traceback (most recent call last):.6921

    ValueError: I/O operation on closed file

    I could resolve this by varying the epoch and batch size.

    Now to predict a unknown value, i loaded a new dataset and used predict cmd as below :
    dataset_test = numpy.loadtxt(“pima-indians-diabetes_test.csv”,delimiter=”,”) –has only one row

    X = dataset_test[:,0:8]

    But I am getting error :
    X = dataset_test[:,0:8]

    IndexError: too many indices for array

    Can you help pls.


    • Jason Brownlee July 24, 2016 at 6:55 am #

      I see problems like this when you run from a notebook or from an IDE.

      Consider running examples from the console to ensure they work.

      Consider tuning off verbose output (verbose=0 in the call to fit()) to disable the progress bar.

  12. David Kluszczynski July 28, 2016 at 12:42 am #

    Hi Jason!
    Loved the tutorial! I have a question however.
    Is there a way to save the weights to a file after the model is trained for uses, such as kaggle?

  13. Alex Hopper July 29, 2016 at 5:45 am #

    Hey, Jason! Thank you for the awesome tutorial! I’ve use your tutorial to learn about CNN. I have one question for you… Supposing I want to use Keras to classicate images and I have 3 or more classes to classify, How could my algorithm know about this classes? You know, I have to code what is a cat, a dog and a horse. Is there any way to code this? I’ve tried it:

    target_names = [‘class 0(Cats)’, ‘class 1(Dogs)’, ‘class 2(Horse)’]
    print(classification_report(np.argmax(Y_test,axis=1), y_pred,target_names=target_names))

    But my results are not classifying correctly.

    precision recall f1-score support
    class 0(Cat) 0.00 0.00 0.00 17
    class 1(Dog) 0.00 0.00 0.00 14
    class 2(Horse) 0.99 1.00 0.99 2526

    avg / total 0.98 0.99 0.98 2557

  14. Anonymouse August 2, 2016 at 11:28 pm #

    This was really useful, thank you

    I’m using keras (with CNNs) for sentiment classification of documents and I’d like to improve the performance, but I’m completely at a loss when it comes to tuning the parameters in a non-arbitrary way. Could you maybe point me somewhere that will help me go about this in a more systematic fashion? There must be some heuristics or rules-of-thumb that could guide me.

    • Jason Brownlee August 3, 2016 at 8:09 am #

      I have a tutorial coming out soon (next week) that provide lots of examples of tuning the hyperparameters of a neural network in Keras, but limited to MLPs.

      For CNNs, I would advise tuning the number of repeating layers (conv + max pool), the number of filters in repeating block, and the number and size of dense layers at the predicting part of your network. Also consider using some fixed layers from pre-trained models as the start of your network (e.g. VGG) and try just training some input and output layers around it for your problem.

      I hope that helps as a start.

  15. Shopon August 14, 2016 at 5:04 pm #

    Hello Jason , My Accuracy is : 0.0104 , but yours is 0.7879 and my loss is : -9.5414 . Is there any problem with the dataset ? I downloaded the dataset from a different site .

    • Jason Brownlee August 15, 2016 at 12:36 pm #

      I think there might be something wrong with your implementation or your dataset. Your numbers are way out.

  16. mohamed August 15, 2016 at 9:30 am #

    after training, how i can use the trained model on new sample

    • Jason Brownlee August 15, 2016 at 12:36 pm #

      You can call model.predict()

      See an above comment for a specific code example.

  17. Omachi Okolo August 16, 2016 at 10:21 pm #

    Hi Jason,
    i’m a student conducting a research on how to use artificial neural network to predict the business viability of potential software projects.
    I intend to use python as a programming language. The application of ANN fascinates me but i’m new to machine learning and python. Can you help suggest how to go about this.
    Many thanks

  18. Agni August 17, 2016 at 6:23 am #

    Dear Jeson, this is a great tutorial for beginners. It will satisfy the need of many students who are looking for the initial help. But I have a question. Could you please light on a few things: i) how to test the trained model using test dataset (i.e., loading of test dataset and applied the model and suppose the test file name is test.csv) ii) print the accuracy obtained on test dataset iii) the o/p has more than 2 class (suppose 4-class classification problem).
    Please show the whole program to overcome any confusion.
    Thanks a lot.

  19. Doron Vetlzer August 17, 2016 at 9:29 am #

    I am trying to build a Neural Network with some recursive connections but not a full recursive layer, how do I do this in Keras?

    • Doron Vetlzer August 17, 2016 at 9:31 am #

      I could print a diagram of the network but what I want Basically is that each neuron in the current time frame to know only its own previous output and not the output of all the neurons in the output layer.

    • Jason Brownlee August 17, 2016 at 10:04 am #

      I don’t know off hand Doron.

      • Doron Veltzer August 23, 2016 at 2:28 am #

        Thanks for replying though, have a good day.

  20. sairam August 30, 2016 at 8:49 am #

    Hello Jason,

    This is a great tutorial . Thanks for sharing.

    I am having a dataset of 100 finger prints and i want to extract minutiae of 100 finger prints using python ( Keras). Can you please advise where to start? I am really confused.

    • Jason Brownlee August 31, 2016 at 8:43 am #

      If your fingerprints are images, you may want to consider using convolutional neural networks (CNNs) that are much better at working image data.

      See this tutorial on digit recognition for a start:

      • padmashri July 6, 2017 at 10:12 pm #

        Hi Jason
        Thanks for this great tutorial, i am new to machine learning i went through your basic tutorial on keras and also handwritten-digit-recognition. I would like to understand how i can train a set of image data, for eg. the set of image data can be some thing like square, circle, pyramid.
        pl. let me know how the input data needs to fed to the program and how we need to export the model.

  21. CM September 1, 2016 at 4:23 pm #

    Hi Jason,

    Thanks for the great article. But I had 1 query.

    Are there any inbuilt functions in keras that can give me the feature importance for the ANN model?

    If not, can you suggest a technique I can use to extract variable importance from the loss function? I am considering an approach similar to that used in RF which involves permuting the values of the selected variable and calculating the relative increase in loss.


  22. Kamal September 7, 2016 at 2:09 am #

    Dear Jason, I am new to Deep learning. Being a novice, I am asking you a technical question which may seem silly. My question is that- can we use features (for example length of the sentence etc.) of a sentence while classifying a sentence ( suppose the o/p are +ve sentence and -ve sentence) using deep neural network?

    • Jason Brownlee September 7, 2016 at 10:27 am #

      Great question Kamal, yes you can. I would encourage you to include all such features and see which give you a bump in performance.

  23. Saurabh September 11, 2016 at 12:42 pm #

    Hi, How would I use this on a dataset that has multiple outputs? For example a dataset with output A and B where A could be 0 or 1 and B could be 3 or 4 ?

  24. Tom_P September 17, 2016 at 1:47 pm #

    Hi Jason,
    The tutorial looks really good but unfortunately I keep getting an error when importing Dense from keras.layers, I get the error : AttributeError: module ‘theano’ has no attribute ‘gof’
    I have tried reinstalling Theano but it has not fixed the issue.

    Best wishes

    • Jason Brownlee September 18, 2016 at 7:57 am #

      Hi Tom, sorry to hear that. I have not seen this problem before.

      Have you searched google? I can see a few posts and it might be related to your version of scipy or similar.

      Let me know how you go.

  25. shudhan September 21, 2016 at 5:54 pm #

    Hey Jason,

    Can you please make a tutorial on how to add additional train data into the already trained model? This will be helpful for the bigger data sets. I read that warm start is used for random forest. But not sure how to implement as algorithm. A generalised version of how to implement would be good. Thank You!

    • Jason Brownlee September 22, 2016 at 8:08 am #

      Great question Shudhan!

      Yes, you could save your weights, load them later into a new network topology and start training on new data again.

      I’ll work out an example in coming weeks, time permitting.

  26. Joanna September 22, 2016 at 1:09 am #

    Hi Jason,
    first of all congratulations for this amazing work that you have done!
    Here is my question:
    What about if my .csv file includes also both nominal and numerical attributes?
    Should I change my nominal values to numerical?

    Thank you in advance

  27. ATM October 2, 2016 at 5:47 am #

    A small bug:-
    Line 25 : rounded = [round(x) for x in predictions]

    should have numpy.round instead, for the code to run!
    Great tutorial, regardless. The best i’ve seen for intro to ANN in python. Thanks!

    • Jason Brownlee October 2, 2016 at 8:20 am #

      Perhaps it’s your version of Python or environment?

      In Python 2.7 the round() function is built-in.

      • AC January 14, 2017 at 2:11 am #

        If there is comment for python3, should be better.
        #use unmpy.round instead, if using python3,

  28. Ash October 9, 2016 at 1:36 am #

    This is simple to grasp! Great post! How can we perform dropout in keras?

  29. Homagni Saha October 14, 2016 at 4:15 am #

    Hello Jason,
    You are using model.predict in the end to predict the results. Is it possible to save the model somewhere in the harddisk and transfer it to another machine(turtlebot running on ROS for my instance) and then use the model directly on turtlebot to predict the results?
    Please tell me how
    Thanking you
    Homagni Saha

  30. Rimi October 16, 2016 at 8:21 pm #

    Hi Jason,
    I implemented you code to begin with. But I am getting an accuracy of 45.18% with the same parameters and everything.
    Cant figure out why.

    • Jason Brownlee October 17, 2016 at 10:29 am #

      There does sound like a problem there Rimi.

      Confirm the code and data match exactly.

  31. Ankit October 26, 2016 at 8:12 pm #

    Hi Jason,
    I am little confused with first layer parameters. You said that first layer has 12 neurons and expects 8 input variables.

    Why there is a difference between number of neurons, input_dim for first layer.


    • Jason Brownlee October 27, 2016 at 7:45 am #

      Hi Ankit,

      The problem has 8 input variables and the first hidden layer has 12 neurons. Inputs are the columns of data, these are fixed. The Hidden layers in general are whatever we design based on whatever capacity we think we need to represent the complexity of the problem. In this case, we have chosen 12 neurons for the first hidden layer.

      I hope that is clearer.

  32. Tom October 27, 2016 at 3:04 am #

    I have a data , IRIS like data but with more colmuns.
    I want to use MLP and DBN/CNNClassifier (or any other Deep Learning classificaiton algorithm) on my data to see how correctly it does classified into 6 groups.

    Previously using DEEP LEARNING FOR J, today first time see KERAS.
    does KERAS has examples (code examples) of DL Classification algorithms?


    • Jason Brownlee October 27, 2016 at 7:48 am #

      Yes Tom, the example in this post is an example of a neural network (deep learning) applied to a classification problem.

  33. Rumesa October 30, 2016 at 1:57 am #

    I have installed theano but it gives me the error of it mendatory to install both packages? because tensorflow is not supported on wndows.the only way to get it on windows is to install virtual machine

    • Jason Brownlee October 30, 2016 at 8:57 am #

      Keras will work just fine with Theano.

      Just install Theano, and configure Keras to use the Theano backend.

      More information about configuring the Keras backend here:

      • Rumesa October 31, 2016 at 4:36 am #

        hey jason I have run your code but got the following error.Although I have aready installed theano me out.I just stuck.

        Using TensorFlow backend.
        Traceback (most recent call last):
        File “C:\Users\pc\Desktop\”, line 2, in
        from keras.models import Sequential
        File “C:\Users\pc\Anaconda3\lib\site-packages\keras\”, line 2, in
        from . import backend
        File “C:\Users\pc\Anaconda3\lib\site-packages\keras\backend\”, line 64, in
        from .tensorflow_backend import *
        File “C:\Users\pc\Anaconda3\lib\site-packages\keras\backend\”, line 1, in
        import tensorflow as tf
        ImportError: No module named ‘tensorflow’

        • Jason Brownlee October 31, 2016 at 5:34 am #

          Change the backend used by Keras from TensorFlow to Theano.

          You can do this either by using the command line switch or changing the Keras config file.

          See the link I posted in the previous post for instructions.

    • Maria January 6, 2017 at 1:05 pm #

      Hello Rumesa!
      Have you solved your problem? I have the same one. Everywhere is the same answer with keras.json file or envirinment variable but it doesn’t work. Can you tell me what have worked for you?

      • Jason Brownlee January 7, 2017 at 8:20 am #


        Maybe there is an issue with the latest version and a tight coupling to tensorflow? I have not seen this myself.

        Perhaps it might be worth testing prior versions of Keras, such as 1.1.0?

        Try this:

  34. Alexon November 1, 2016 at 6:54 am #

    Hi Jason,

    First off, thanks so much for creating these resources, I have been keeping an eye on your newsletter for a while now, and I finally have the free time to start learning more about it myself, so your work has been really appreciated.

    My question is: How can I set/get the weights of each hidden node?

    I am planning to create several arrays randomized weights, then use a genetic algorithm to see which weight array performs the best and improve over generations. How would be the best way to go about this, and if I use a “relu” activation function, am I right in thinking these randomly generated weights should be between 0 and 0.05?

    Many thanks for your help 🙂

    • Jason Brownlee November 1, 2016 at 8:05 am #

      Thanks Alexon,

      You can get and set the weights from a network.

      You can learn more about how to do this in the context of saving the weights to file here:

      I hope that helps as a start, I’d love to hear how you go.

      • Alexon November 6, 2016 at 6:36 am #

        Thats great, thanks for pointing me in the right direction.
        I’d be happy to let you know how it goes, but might take a while as this is very much a “when I can find the time” project between jobs 🙂


  35. Arnaldo Gunzi November 2, 2016 at 10:17 pm #

    Nice introduction, thanks!

  36. Abbey November 14, 2016 at 11:05 pm #

    Good day

    I have a question, how can I represent a character as a vector that could be an input for the neural network to predict the word meaning and trained using LSTM

    For instance, I have bf to predict boy friend or best friend and similarly I have 2mor to predict tomorrow. I need to encode all the input as a character represented as vector, so that it can be train with RNN/LSTM to predict the output.

    Thank you.

    Kind Regards

    • Jason Brownlee November 15, 2016 at 7:54 am #

      Hi Abbey, You can map characters to integers to get integer vectors.

      • Abbey November 15, 2016 at 6:17 pm #

        Thank you Jason, if i map characters to integers value to get vectors using English Alphabets, numbers and special characters

        The question is how will LSTM predict the character. Please example in more details for me.


        • Jason Brownlee November 16, 2016 at 9:27 am #

          Hi Abbey,

          If your output values are also characters, you can map them onto integers, and reverse the mapping to convert the predictions back to text.

          • Abbey November 16, 2016 at 8:39 pm #

            The output value of the characters encoding will be text

      • Abbey November 15, 2016 at 6:22 pm #

        Thank you, Jason, if I map characters to integers value to get vectors representation of the informal text using English Alphabets, numbers and special characters

        The question is how will LSTM predict the character or words that have close meaning to the input value. Please example in more details for me. I understand how RNN/LSTM work based on your tutorial example but the logic in designing processing is what I am stress with.


  37. Ammar November 27, 2016 at 10:35 am #

    hi Jason,
    i am trying to implement CNN one dimention on my data. so, i bluit my network.
    the issue is:
    def train_model(model, X_train, y_train, X_test, y_test):
    X_train = X_train.reshape(-1, 1, 41)
    X_test = X_test.reshape(-1, 1, 41)

    numpy.random.seed(seed), y_train, validation_data=(X_test, y_test), nb_epoch=100, batch_size=64)
    # Final evaluation of the model
    scores = model.evaluate(X_test, y_test, verbose=0)
    print(“Accuracy: %.2f%%” % (scores[1] * 100))
    this method above does not work and does not give me any error message.
    could you help me with this please?

    • Jason Brownlee November 28, 2016 at 8:40 am #

      Hi Ammar, I’m surprised that there is no error message.

      Perhaps run from the command line and add some print() statements to see exactly where it stops.

  38. KK November 28, 2016 at 6:55 pm #

    Hi Jason
    Great work. I have another doubt. How can we apply this to text mining. I have a csv file containing review document and label. I want to apply classify the documents based on the text available. Can U do this favor.

    • Jason Brownlee November 29, 2016 at 8:48 am #

      I would recommend converting the chars to ints and then using an Embedding layer.

  39. Alex M November 30, 2016 at 10:52 pm #

    Mr Jason, this is great tutorial but I am stack with some errors.

    First I can’t load data set correctly, tried to correct error but can’t make it. ( FileNotFoundError: [Errno 2] No such file or directory: ‘pima-indians-diabetes.csv’ ).

    Second: While trying to evaluate the model it says (X is not defined) May be this is because uploading failed.


    • Jason Brownlee December 1, 2016 at 7:29 am #

      You need to download the file and place it in your current working directory Alex.

      Does that help?

  40. Alex M December 1, 2016 at 6:45 pm #

    Sir, it is now successful….

  41. Bappaditya December 2, 2016 at 7:35 pm #

    Hi Jason,

    First of all a special thanks to you for providing such a great tutorial. I am very new to machine learning and truly speaking i had no background in data science. The concept of ML overwhelmed me and now i have a desire to be an expert of this field. I need your advice to start from a scratch. Also i am a PhD student in Computer Engineering ( computer hardware )and i want to apply it as a tool for fault detection and testing for ICs.Can you provide me some references on this field?

  42. Alex M December 3, 2016 at 8:00 pm #

    Well as usual in our daily coding life errors happen, now I have this error how can I correct it? Thanks!

    ” —————————————————————————
    NoBackendError Traceback (most recent call last)
    in ()
    16 import librosa.display
    17 audio_path = (‘/Users/MA/Python Notebook/OK.mp3’)
    —> 18 y, sr = librosa.load(audio_path)

    C:\Users\MA\Anaconda3\lib\site-packages\librosa\core\ in load(path, sr, mono, offset, duration, dtype)
    108 y = []
    –> 109 with audioread.audio_open(os.path.realpath(path)) as input_file:
    110 sr_native = input_file.samplerate
    111 n_channels = input_file.channels

    C:\Users\MA\Anaconda3\lib\site-packages\audioread\ in audio_open(path)
    113 # All backends failed!
    –> 114 raise NoBackendError()


    That is the error I am getting just when trying to load a song into librosa…
    Thanks!! @Jason Brownlee

    • Jason Brownlee December 4, 2016 at 5:30 am #

      Sorry, this looks like an issue with your librosa library, not a machine learning issue. I can’t give you expert advice, sorry.

  43. Alex M December 4, 2016 at 10:30 pm #

    Thanks I have managed to correct the error…

    Happy Sunday to you all……

    • Jason Brownlee December 5, 2016 at 6:49 am #

      Glad to hear it Alex.

    • ayush June 19, 2018 at 3:27 am #

      how did you solved the problem?

  44. Lei December 4, 2016 at 10:52 pm #

    Hi, Jason, thank you for your amazing examples.
    I run the same code on my laptop. But I did not get the same results. What could be the possible reasons?
    I am using windows 8.1 64bit+eclipse+anaconda 4.2+theano 0.9.4+CUDA7.5
    I got results like follows.

    … …
    Epoch 145/150

    10/768 […………………………] – ETA: 0s – loss: 0.3634 – acc: 0.8000
    80/768 [==>………………………] – ETA: 0s – loss: 0.4066 – acc: 0.7750
    150/768 [====>…………………….] – ETA: 0s – loss: 0.4059 – acc: 0.8067
    220/768 [=======>………………….] – ETA: 0s – loss: 0.4047 – acc: 0.8091
    300/768 [==========>……………….] – ETA: 0s – loss: 0.4498 – acc: 0.7867
    380/768 [=============>…………….] – ETA: 0s – loss: 0.4595 – acc: 0.7895
    450/768 [================>………….] – ETA: 0s – loss: 0.4568 – acc: 0.7911
    510/768 [==================>………..] – ETA: 0s – loss: 0.4553 – acc: 0.7882
    580/768 [=====================>……..] – ETA: 0s – loss: 0.4677 – acc: 0.7776
    660/768 [========================>…..] – ETA: 0s – loss: 0.4697 – acc: 0.7788
    740/768 [===========================>..] – ETA: 0s – loss: 0.4611 – acc: 0.7838
    768/768 [==============================] – 0s – loss: 0.4614 – acc: 0.7799
    Epoch 146/150

    10/768 […………………………] – ETA: 0s – loss: 0.3846 – acc: 0.8000
    90/768 [==>………………………] – ETA: 0s – loss: 0.5079 – acc: 0.7444
    170/768 [=====>……………………] – ETA: 0s – loss: 0.4500 – acc: 0.7882
    250/768 [========>…………………] – ETA: 0s – loss: 0.4594 – acc: 0.7840
    330/768 [===========>………………] – ETA: 0s – loss: 0.4574 – acc: 0.7818
    400/768 [==============>……………] – ETA: 0s – loss: 0.4563 – acc: 0.7775
    470/768 [=================>…………] – ETA: 0s – loss: 0.4654 – acc: 0.7723
    540/768 [====================>………] – ETA: 0s – loss: 0.4537 – acc: 0.7870
    620/768 [=======================>……] – ETA: 0s – loss: 0.4615 – acc: 0.7806
    690/768 [=========================>….] – ETA: 0s – loss: 0.4631 – acc: 0.7739
    750/768 [============================>.] – ETA: 0s – loss: 0.4649 – acc: 0.7733
    768/768 [==============================] – 0s – loss: 0.4636 – acc: 0.7734
    Epoch 147/150

    10/768 […………………………] – ETA: 0s – loss: 0.3561 – acc: 0.9000
    90/768 [==>………………………] – ETA: 0s – loss: 0.4167 – acc: 0.8556
    170/768 [=====>……………………] – ETA: 0s – loss: 0.4824 – acc: 0.8059
    250/768 [========>…………………] – ETA: 0s – loss: 0.4534 – acc: 0.8080
    330/768 [===========>………………] – ETA: 0s – loss: 0.4679 – acc: 0.7848
    400/768 [==============>……………] – ETA: 0s – loss: 0.4590 – acc: 0.7950
    460/768 [================>………….] – ETA: 0s – loss: 0.4619 – acc: 0.7913
    530/768 [===================>……….] – ETA: 0s – loss: 0.4562 – acc: 0.7868
    600/768 [======================>…….] – ETA: 0s – loss: 0.4497 – acc: 0.7883
    680/768 [=========================>….] – ETA: 0s – loss: 0.4525 – acc: 0.7853
    760/768 [============================>.] – ETA: 0s – loss: 0.4568 – acc: 0.7803
    768/768 [==============================] – 0s – loss: 0.4561 – acc: 0.7812
    Epoch 148/150

    10/768 […………………………] – ETA: 0s – loss: 0.4183 – acc: 0.9000
    80/768 [==>………………………] – ETA: 0s – loss: 0.3674 – acc: 0.8750
    160/768 [=====>……………………] – ETA: 0s – loss: 0.4340 – acc: 0.8250
    240/768 [========>…………………] – ETA: 0s – loss: 0.4799 – acc: 0.7583
    320/768 [===========>………………] – ETA: 0s – loss: 0.4648 – acc: 0.7719
    400/768 [==============>……………] – ETA: 0s – loss: 0.4596 – acc: 0.7775
    470/768 [=================>…………] – ETA: 0s – loss: 0.4475 – acc: 0.7809
    540/768 [====================>………] – ETA: 0s – loss: 0.4545 – acc: 0.7778
    620/768 [=======================>……] – ETA: 0s – loss: 0.4590 – acc: 0.7742
    690/768 [=========================>….] – ETA: 0s – loss: 0.4769 – acc: 0.7652
    760/768 [============================>.] – ETA: 0s – loss: 0.4748 – acc: 0.7658
    768/768 [==============================] – 0s – loss: 0.4734 – acc: 0.7669
    Epoch 149/150

    10/768 […………………………] – ETA: 0s – loss: 0.3043 – acc: 0.9000
    90/768 [==>………………………] – ETA: 0s – loss: 0.4913 – acc: 0.7111
    170/768 [=====>……………………] – ETA: 0s – loss: 0.4779 – acc: 0.7588
    250/768 [========>…………………] – ETA: 0s – loss: 0.4794 – acc: 0.7640
    320/768 [===========>………………] – ETA: 0s – loss: 0.4957 – acc: 0.7562
    370/768 [=============>…………….] – ETA: 0s – loss: 0.4891 – acc: 0.7703
    450/768 [================>………….] – ETA: 0s – loss: 0.4737 – acc: 0.7867
    520/768 [===================>……….] – ETA: 0s – loss: 0.4675 – acc: 0.7865
    600/768 [======================>…….] – ETA: 0s – loss: 0.4668 – acc: 0.7833
    680/768 [=========================>….] – ETA: 0s – loss: 0.4677 – acc: 0.7809
    760/768 [============================>.] – ETA: 0s – loss: 0.4648 – acc: 0.7803
    768/768 [==============================] – 0s – loss: 0.4625 – acc: 0.7826
    Epoch 150/150

    10/768 […………………………] – ETA: 0s – loss: 0.2751 – acc: 1.0000
    100/768 [==>………………………] – ETA: 0s – loss: 0.4501 – acc: 0.8100
    170/768 [=====>……………………] – ETA: 0s – loss: 0.4588 – acc: 0.8059
    250/768 [========>…………………] – ETA: 0s – loss: 0.4299 – acc: 0.8200
    310/768 [===========>………………] – ETA: 0s – loss: 0.4298 – acc: 0.8129
    380/768 [=============>…………….] – ETA: 0s – loss: 0.4365 – acc: 0.8053
    460/768 [================>………….] – ETA: 0s – loss: 0.4469 – acc: 0.7957
    540/768 [====================>………] – ETA: 0s – loss: 0.4436 – acc: 0.8000
    620/768 [=======================>……] – ETA: 0s – loss: 0.4570 – acc: 0.7871
    690/768 [=========================>….] – ETA: 0s – loss: 0.4664 – acc: 0.7783
    760/768 [============================>.] – ETA: 0s – loss: 0.4617 – acc: 0.7789
    768/768 [==============================] – 0s – loss: 0.4638 – acc: 0.7773

    32/768 [>………………………..] – ETA: 0s
    448/768 [================>………….] – ETA: 0sacc: 79.69%

  45. Nanya December 10, 2016 at 2:55 pm #

    Hello Jason Brownlee,Thx for sharing~
    I’m new in deep learning.And I am wondering can what you dicussed here:”Keras” be used to build a CNN in tensorflow and train some csv fiels for classification.May be this is a stupid question,but waiting for you reply.I’m working on my graduation project for Word sense disambiguation with cnn,and just can’t move on.Hope for your heip~Bese wishes!

    • Jason Brownlee December 11, 2016 at 5:22 am #

      Sorry Nanya, I’m not sure I understand your question. Are you able to rephrase it?

  46. Anon December 16, 2016 at 12:51 am #

    I’ve just installed Anaconda with Keras and am using python 3.5.
    It seems there’s an error with the rounding using Py3 as opposed to Py2. I think it’s because of this change:

    I removed the rounding and just used print(predictions) and it seemed to work outputting floats instead.

    Does this look correct?

    Epoch 150/150
    0s – loss: 0.4593 – acc: 0.7839
    [[ 0.79361773]
    [ 0.10443526]
    [ 0.90862554]
    [ 0.33652252]
    [ 0.63745886]
    [ 0.11704451]]

  47. Florin Claudiu Mihalache December 19, 2016 at 2:37 am #

    Hi Jason Brownlee
    I tried to modified your exemple for my problem (Letter Recognition ,
    My data set look like (T,2,8,3,5,1,8,13,0,6,6,10,8,0,8,0,8) .I try to split the data in input and ouput like this :

    X = dataset[:,1:17]
    Y = dataset[:,0]
    but a have some error (something related that strings are not recognized) .
    I tried to modified each letter whit the ASCII code (A became 65 and so on).The string error disappeared.
    The program compiles now but the output look like this :

    17445/20000 [=========================>….] – ETA: 0s – loss: -1219.4768 – acc:0.0000e+00
    17605/20000 [=========================>….] – ETA: 0s – loss: -1219.4706 – acc:0.0000e+00
    17730/20000 [=========================>….] – ETA: 0s – loss: -1219.4566 – acc:0.0000e+00
    17890/20000 [=========================>….] – ETA: 0s – loss: -1219.4071 – acc:0.0000e+00
    18050/20000 [==========================>…] – ETA: 0s – loss: -1219.4599 – acc:0.0000e+00
    18175/20000 [==========================>…] – ETA: 0s – loss: -1219.3972 – acc:0.0000e+00
    18335/20000 [==========================>…] – ETA: 0s – loss: -1219.4642 – acc:0.0000e+00
    18495/20000 [==========================>…] – ETA: 0s – loss: -1219.5032 – acc:0.0000e+00
    18620/20000 [==========================>…] – ETA: 0s – loss: -1219.4391 – acc:0.0000e+00
    18780/20000 [===========================>..] – ETA: 0s – loss: -1219.5652 – acc:0.0000e+00
    18940/20000 [===========================>..] – ETA: 0s – loss: -1219.5520 – acc:0.0000e+00
    19080/20000 [===========================>..] – ETA: 0s – loss: -1219.5381 – acc:0.0000e+00
    19225/20000 [===========================>..] – ETA: 0s – loss: -1219.5182 – acc:0.0000e+00
    19385/20000 [============================>.] – ETA: 0s – loss: -1219.6742 – acc:0.0000e+00
    19535/20000 [============================>.] – ETA: 0s – loss: -1219.7030 – acc:0.0000e+00
    19670/20000 [============================>.] – ETA: 0s – loss: -1219.7634 – acc:0.0000e+00
    19830/20000 [============================>.] – ETA: 0s – loss: -1219.8336 – acc:0.0000e+00
    19990/20000 [============================>.] – ETA: 0s – loss: -1219.8532 – acc:0.0000e+00
    20000/20000 [==============================] – 1s – loss: -1219.8594 – acc: 0.0000e+00
    18880/20000 [===========================>..] – ETA: 0sacc: 0.00%

    I do not understand why. Can you please help me

    • Anon December 26, 2016 at 6:44 am #

      What version of Python are you running?

  48. karishma sharma December 22, 2016 at 10:03 am #

    Hi Jason,

    Since the epoch is set to 150 and batch size is 10, does the training algorithm pick 10 training examples at random in each iteration, given that we had only 768 total in X. Or does it sample randomly after it has finished covering all.


    • Jason Brownlee December 23, 2016 at 5:27 am #

      Good question,

      It iterates over the dataset 150 times and within one epoch it works through 10 rows at a time before doing an update to the weights. The patterns are shuffled before each epoch.

      I hope that helps.

  49. Kaustuv January 9, 2017 at 4:57 am #

    Hi Jason
    Thanks a lot for this blog. It really helps me to start learning deep learning which was in a planning state for last few months. Your simple enrich blogs are awsome. No questions from my side before completing all tutorials.
    One question regarding availability of your book. How can I buy those books from India ?

  50. Stephen Wilson January 15, 2017 at 4:00 pm #

    Hi Jason, firstly your work here is a fantastic resource and I am very thankful for the effort you put in.
    I am a slightly-better-than-beginner at python and an absolute novice at ML, I wonder if you could help me classify my problem and find an angle to work at it from.

    My data is thus:
    Column Names: 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, Result
    Values: 4, 4, 6, 6, 3, 2, 5, 5, 0, 0, 0, 0, 0, 0, 0, 4

    I want to find the percentage chance of each Column Names category being the Result based off the configuration of all the values present from 1-15. Then if need be compare the configuration of Values with another row of values to find the same, Resulting in the total needed calculation as:

    Column Names: 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, Result
    Values: 4, 4, 6, 6, 3, 2, 5, 5, 0, 0, 0, 0, 0, 0, 0, 4
    Values2: 7, 3, 5, 1, 4, 8, 6, 2, 9, 9, 9, 9, 9, 9, 9

    I apologize if my explanation is not clear, and appreciate any help you can give me thank you.

  51. Rohit January 16, 2017 at 10:37 pm #

    Thanks Jason for such a nice and concise example.

    Just wanted to ask if it is possible to save this model in a file and port it to may be an Android or iOS device? If so, what are the libraries available for the same?



    • Jason Brownlee January 17, 2017 at 7:38 am #

      Thanks Rohit,

      Here’s an example of saving a Keras model to file:

      I don’t know about running Keras on an Android or iOS device. Let me know how you go.

      • zaheer khan June 16, 2017 at 7:17 pm #

        Dear Jason, Thanks for sharing this article.
        I am novice to the deep learning, and my apology if my question is not clear. my question is could we call all that functions and program from any .php,.aspx, or .html webpage. i mean i load the variables and other files selection from user interface and then make them input to this functions.

        will be waiting for your kind reply.
        thanks in advance.

        • Jason Brownlee June 17, 2017 at 7:25 am #

          Perhaps, this sounds like a systems design question, not really machine learning.

          I would suggest you gather requirements, assess risks like any software engineering project.

  52. Hsiang January 18, 2017 at 3:35 pm #

    Hi, Jason

    Thank you for your blog! It is wonderful!

    I used tensorflow as backend, and implemented the procedures using Jupyter.
    I did “source activate tensorflow” -> “ipython notebook”.
    I can successfully use Keras and import tensorflow.

    However, it seems that such environment doesn’t support pandas and sklearn.
    Do you have any way to incorporate pandas, sklearn and keras?
    (I wish to use sklearn to revisit the classification problem and compare the accuracy with the deep learning method. But I also wish to put the works together in the same interface.)


    • Jason Brownlee January 19, 2017 at 7:24 am #

      Sorry, I do not use notebooks myself. I cannot offer you good advice.

      • Hsiang January 19, 2017 at 12:53 pm #

        Thanks, Jason!
        Actually the problem is not on notebooks. Even I used the terminal mode, i.e. doing “source activate tensorflow” only. It failed to import sklearn. Does that mean tensorflow library is not compatible with sklearn? Thanks again!

        • Jason Brownlee January 20, 2017 at 10:17 am #

          Sorry Hsiang, I don’t have experience using sklearn and tensorflow with virtual environments.

          • Hsiang January 21, 2017 at 12:46 am #

            Thank you!

          • Jason Brownlee January 21, 2017 at 10:34 am #

            You’re welcome Hsiang.

  53. keshav bansal January 24, 2017 at 12:45 am #

    hello sir,
    A very informative post indeed . I know my question is a very trivial one but can you please show me how to predict on a explicitly mentioned data tuple say v=[6,148,72,35,0,33.6,0.627,50]
    thanks for the tutorial anyway

    • Jason Brownlee January 24, 2017 at 11:04 am #

      Hi keshav,

      You can make predictions by calling model.predict()

  54. CATRINA WEBB January 25, 2017 at 9:06 am #

    When I rerun the file (without predictions) does it reset the model and weights?

  55. Ericson January 30, 2017 at 8:04 pm #

    excuse me sir, i wanna ask you a question about this paragraph”dataset = numpy.loadtxt(“pima-indians-diabetes.csv”,delimiter=’,’)”, i used the mac and downloaded the dataset,then i exchanged the text into csv file. Running the program

    ,hen i got:{Python 2.7.13 (v2.7.13:a06454b1afa1, Dec 17 2016, 12:39:47)
    [GCC 4.2.1 (Apple Inc. build 5666) (dot 3)] on darwin
    Type “copyright”, “credits” or “license()” for more information.
    ============ RESTART: /Users/luowenbin/Documents/ ============
    Using TensorFlow backend.

    Traceback (most recent call last):
    File “/Users/luowenbin/Documents/”, line 9, in
    dataset = numpy.loadtxt(“pima-indians-diabetes.csv”,delimiter=’,’)
    File “/Library/Frameworks/Python.framework/Versions/2.7/lib/python2.7/site-packages/numpy/lib/”, line 985, in loadtxt
    items = [conv(val) for (conv, val) in zip(converters, vals)]
    File “/Library/Frameworks/Python.framework/Versions/2.7/lib/python2.7/site-packages/numpy/lib/”, line 687, in floatconv
    return float(x)
    ValueError: could not convert string to float: book
    >>> }
    How can i solve this problem? give me a hand thank you!

    • Jason Brownlee February 1, 2017 at 10:22 am #

      Hi Ericson,

      Confirm that the contents of “pima-indians-diabetes.csv” meet your expectation of a list of CSV lines.

  56. Sukhpal February 7, 2017 at 9:00 pm #

    excuse me sir,when i run this code for my data set ,I encounter this problem…please help me finding solution to this problem
    runfile(‘C:/Users/sukhpal/.spyder/’, wdir=’C:/Users/sukhpal/.spyder’)
    Using TensorFlow backend.
    Traceback (most recent call last):

    File “”, line 1, in
    runfile(‘C:/Users/sukhpal/.spyder/’, wdir=’C:/Users/sukhpal/.spyder’)

    File “C:\Users\sukhpal\Anaconda2\lib\site-packages\spyder\utils\site\”, line 866, in runfile
    execfile(filename, namespace)

    File “C:\Users\sukhpal\Anaconda2\lib\site-packages\spyder\utils\site\”, line 87, in execfile
    exec(compile(scripttext, filename, ‘exec’), glob, loc)

    File “C:/Users/sukhpal/.spyder/”, line 1, in
    from keras.models import Sequential

    File “C:\Users\sukhpal\Anaconda2\lib\site-packages\keras\”, line 2, in
    from . import backend

    File “C:\Users\sukhpal\Anaconda2\lib\site-packages\keras\backend\”, line 67, in
    from .tensorflow_backend import *

    File “C:\Users\sukhpal\Anaconda2\lib\site-packages\keras\backend\”, line 1, in
    import tensorflow as tf

    ImportError: No module named tensorflow

    • Jason Brownlee February 8, 2017 at 9:34 am #

      This is a change with the most recent version of tensorflow, I will investigate and change the example.

      For now, consider installing and using an older version of tensorflow.

  57. Will February 14, 2017 at 5:33 am #

    Great tutorial! Amazing amount of work you’ve put in and great marketing skills (I also have an email list, ebooks and sequence, etc). I ran this in Jupyter notebook… I noticed the 144th epoch (acc .7982) had more accuracy than at 150. Why is that?

    P.S. i did this for the print: print(numpy.round(predictions))
    It seems to avoid a list of arrays which when printing includes the dtype (messy)

  58. Sukhpal February 14, 2017 at 3:50 pm #

    Please help me to find out this error
    runfile(‘C:/Users/sukhpal/.spyder/’, wdir=’C:/Users/sukhpal/.spyder’)ERROR: execution aborted

    • Jason Brownlee February 15, 2017 at 11:32 am #

      I’m not sure Sukhpal.

      Consider getting code working from the command line, I don’t use IDEs myself.

  59. Kamal February 14, 2017 at 5:15 pm #

    please help me to find this error find this error
    Epoch 194/195
    195/195 [==============================] – 0s – loss: 0.2692 – acc: 0.8667
    Epoch 195/195
    195/195 [==============================] – 0s – loss: 0.2586 – acc: 0.8667
    195/195 [==============================] – 0s
    Traceback (most recent call last):

  60. Kamal February 15, 2017 at 3:24 pm #

    sir when i run the code on my data set
    then it doesnot show overall accuracy although it shows the accuracy and loss for the whole iterations

    • Jason Brownlee February 16, 2017 at 11:06 am #

      I’m not sure I understand your question Kamal, please you could restate it?

  61. Val February 15, 2017 at 9:00 pm #

    Hi Jason, im just starting deep learning in python using keras and theano. I have followed the installation instructions without a hitch. Tested some examples but when i run this one line by line i get a lot of exceptions and errors once i run the “,Y, nb_epochs=150, batch_size=10”

  62. CrisH February 17, 2017 at 8:12 pm #

    Hi, how do I know what number to use for random.seed() ? I mean you use 7, is there any reason for that? Also is it enough to use it only once, in the beginning of the code?

  63. kk February 18, 2017 at 1:53 am #

    am new to deep learning and found this great tutorial. keep it up and look forward!!

  64. Iqra Ameer February 21, 2017 at 5:20 am #

    HI, I have a problem in execution the above example as it. It seems that it’s not running properly and stops at Using TensorFlow backend.

    Epoch 147/150
    768/768 [==============================] – 0s – loss: 0.4709 – acc: 0.7878
    Epoch 148/150
    768/768 [==============================] – 0s – loss: 0.4690 – acc: 0.7812
    Epoch 149/150
    768/768 [==============================] – 0s – loss: 0.4711 – acc: 0.7721
    Epoch 150/150
    768/768 [==============================] – 0s – loss: 0.4731 – acc: 0.7747
    32/768 [>………………………..] – ETA: 0sacc: 76.43%

    I am new in this field, could you please guide me about this error.
    I also executed on another data set, it stops with the same behavior.

    • Jason Brownlee February 21, 2017 at 9:39 am #

      What is the error exactly? The example hangs?

      Maybe try the Theano backend and see if that makes a difference. Also make sure all of your libraries are up to date.

  65. Iqra Ameer February 22, 2017 at 5:47 am #

    Dear Jason,
    Thank you so much for your valuable suggestions. I tried Theano backend and also updated all my libraries, but again it hanged at:

    768/768 [==============================] – 0s – loss: 0.4656 – acc: 0.7799
    Epoch 149/150
    768/768 [==============================] – 0s – loss: 0.4589 – acc: 0.7826
    Epoch 150/150
    768/768 [==============================] – 0s – loss: 0.4611 – acc: 0.7773
    32/768 [>………………………..] – ETA: 0sacc: 78.91%

    • Jason Brownlee February 22, 2017 at 10:05 am #

      I’m sorry to hear that, I have not seen this issue before.

      Perhaps a RAM issue or a CPU overheating issue? Are you able to try different hardware?

    • frd March 8, 2017 at 2:50 am #


      Were you able to find a solution for that?

      I’m having exactly the same problem

      ( … )
      Epoch 149/150
      768/768 [==============================] – 0s – loss: 0.4593 – acc: 0.7773
      Epoch 150/150
      768/768 [==============================] – 0s – loss: 0.4586 – acc: 0.7891
      32/768 [>………………………..] – ETA: 0sacc: 76.69%

  66. Bhanu February 23, 2017 at 1:51 pm #

    Hello sir,
    i want to ask wether we can convert this code to deep learning wid increasing number of layers..

    • Jason Brownlee February 24, 2017 at 10:12 am #

      Sure you can increase the number of layers, try it and see.

  67. Ananya Mohapatra February 28, 2017 at 6:40 pm #

    hello sir,
    could you please tell me how do i determine the no.of neurons in each layer, because i am using a different datset and am unable to know the no.of neurons in each layer

    • Jason Brownlee March 1, 2017 at 8:33 am #

      Hi Ananya, great question.

      Sorry, there is no good theory on how to configure a neural net.

      You can configure the number of neurons in a layer by trial and error. Also consider tuning the number of epochs and batch size at the same time.

      • Ananya Mohapatra March 1, 2017 at 4:42 pm #

        thank you so much sir. It worked ! 🙂

  68. Jayant Sahewal February 28, 2017 at 8:11 pm #

    Hi Jason,

    really helpful blog. I have a question about how much time does it take to converge?

    I have a dataset with around 4000 records, 3 input columns and 1 output column. I came up with the following model

    def create_model(dropout_rate=0.0, weight_constraint=0, learning_rate=0.001, activation=’linear’):
    # create model
    model = Sequential()
    model.add(Dense(6, input_dim=3, init=’uniform’, activation=activation, W_constraint=maxnorm(weight_constraint)))
    model.add(Dense(1, init=’uniform’, activation=’sigmoid’))
    # Optimizer
    optimizer = Adam(lr=learning_rate)
    # Compile model
    model.compile(loss=’binary_crossentropy’, optimizer=optimizer, metrics=[‘accuracy’])
    return model

    # create model
    model = KerasRegressor(build_fn=create_model, verbose=0)
    # define the grid search parameters
    batch_size = [10]
    epochs = [100]
    weight_constraint = [3]
    dropout_rate = [0.9]
    learning_rate = [0.01]
    activation = [‘linear’]
    param_grid = dict(batch_size=batch_size, nb_epoch=epochs, dropout_rate=dropout_rate, \
    weight_constraint=weight_constraint, learning_rate=learning_rate, activation=activation)
    grid = GridSearchCV(estimator=model, param_grid=param_grid, n_jobs=-1, cv=5)
    grid_result =, Y_train)

    I have a 32 core machine with 64 GB RAM and it does not converge even in more than an hour. I can see all the cores busy, so it is using all the cores for training. However, if I change the input neurons to 3 then it converges in around 2 minutes.

    Keras version: 1.1.1
    Tensorflow version: 0.10.0rc0
    theano version:

    It’s using Tensorflow backend. Can you help me understand what is going on or point me in the right direction? Do you think switching to theano will help?


  69. Animesh Mohanty March 1, 2017 at 9:21 pm #

    hello sir,
    could you please tell me how can i plot the results of the code on a graph . I made a few adjustments to the code so as to run it on a different dataset.

    • Jason Brownlee March 2, 2017 at 8:16 am #

      What do you want to plot exactly Animesh?

      • Animesh Mohanty March 2, 2017 at 4:56 pm #

        Accuracy vs no.of neurons in the input layer and the no.of neurons in the hidden layer

  70. param March 2, 2017 at 12:15 am #

    sir can u plz explain
    the different attributes used in this statement
    print(“%s: %.2f%%” % (model.metrics_names[1], scores[1]*100))

  71. Vijin K P March 2, 2017 at 4:01 am #

    Hi Jason,

    It was an awesome post. Could you please tell me how to we decide the following in a DNN 1. number of neurons in the hidden layers
    2. number of hidden layers


  72. Bogdan March 2, 2017 at 11:48 pm #

    Jason, Please tell me about these lines in your code:

    seed = 7

    What do they do? And why do they do it?

    One more question is why do you call the last section Bonus:Make a prediction?
    I thought this what ANN was created for. What the point if your network’s output is just what you have already know?

    • Jason Brownlee March 3, 2017 at 7:44 am #

      They seed the random number generator so that it produces the same sequence of random numbers each time the code is run. This is to ensure you get the same result as me.

      I’m not convinced it works with Keras though.

      More on randomness in machine learning here:

      I was showing how to build and evaluate the model in this tutorial. The part about standalone prediction was an add-on.

  73. Sounak sahoo March 3, 2017 at 7:39 pm #

    what exactly is the work of “seed” in the neural network code? what does it do?

    • Jason Brownlee March 6, 2017 at 10:44 am #

      Seed refers to seeding the random number generator so that the same sequence of random numbers is generated each time the example is run.

      The aim is to make the examples 100% reproducible, but this is hard with symbolic math libs like Theano and TensorFlow backends.

      For more on randomness in machine learning, see this post:

  74. Priya Sundari March 3, 2017 at 10:19 pm #

    hello sir
    could you plz tell me what is the role of optimizer and binary_crossentropy exactly? it is written that optimizer is used to search through the weights of the network which weights are we talking about exactly?

  75. Bogdan March 3, 2017 at 10:23 pm #

    If I am not mistaken, those lines I commented about used when we write

    init = ‘uniform’


  76. Bogdan March 3, 2017 at 10:44 pm #

    Could you explain in more details what is the batch size?

    • Jason Brownlee March 6, 2017 at 10:50 am #

      Hi Bogdan,

      Batch size is how many patterns to show to the network before the weights are updated with the accumulated errors. The smaller the batch, the faster the learning, but also the more noisy the learning (higher variance).

      Try exploring different batch sizes and see the effect on the train and test performance over each epoch.

  77. Mohammad March 7, 2017 at 6:50 am #

    Dear Jason
    Firstly, thanks for your great tutorials.
    I am trying to classify computer networks packets using first 500 bytes of every packet to identify its protocol. I am trying to use 1d convolution. for simpler task,I just want to do binary classification and then tackle multilabel classification for 10 protocols. Here is my code but the accuracy which is like .63. how can I improve the performance? should I Use RNNs?
    activation=’relu’,subsample_length=1, input_shape=(500, 1)))
    optimizer=’adam’,metrics=[‘accuracy’]), y_train,
    #x2= get_activations(model, 0,xprim )
    #score = model.evaluate(t, y_test, show_accuracy = True, verbose = 0)

  78. Damiano March 7, 2017 at 10:13 pm #

    Hi Jason, thank you so much for this awesome tutorial. I have just started with python and machine learning.
    I am joking with the code doing few changes, for example i have changed..


    # create model
    model = Sequential()
    model.add(Dense(250, input_dim=8, init=’uniform’, activation=’relu’))
    model.add(Dense(200, init=’uniform’, activation=’relu’))
    model.add(Dense(200, init=’uniform’, activation=’relu’))
    model.add(Dense(1, init=’uniform’, activation=’sigmoid’))

    and this:, Y, nb_epoch=250, batch_size=10)

    then i would like to pass some arrays for prediction so…

    new_input = numpy.array([[3,88,58,11,54,24.8,267,22],[6,92,92,0,0,19.9,188,28], [10,101,76,48,180,32.9,171,63], [2,122,70,27,0,36.8,0.34,27], [5,121,72,23,112,26.2,245,30]])

    predictions = model.predict(new_input)
    print predictions # [1.0, 1.0, 1.0, 0.0, 1.0]

    is this correct? In this example i used the same series of training (that have 0 class), but i am getting wrong results. Only one array is correctly predicted.

    Thank you so much!

  79. ANJI March 13, 2017 at 8:48 pm #

    hello sir,
    could you please tell me to rectify my error below it is raised while model is training:

    ValueError: Error when checking model input: expected convolution2d_input_1 to have 4 dimensions, but got array with shape (68, 28, 28).

  80. Rimjhim March 14, 2017 at 8:21 pm #

    I want a neural that can predict sin values. Further from a given data set i need to determine the function(for example if the data is of tan or cos, then how to determine that data is of tan only or cos only)

    Thanks in advance

  81. Sudarshan March 15, 2017 at 11:19 pm #

    Keras just updated to Keras 2.0. I have an updated version of this code here:

  82. subhasish March 16, 2017 at 5:09 pm #

    hello sir,
    can we use PSO (particle swarm optimisation) in this? if so can you tell how?

    • Jason Brownlee March 17, 2017 at 8:25 am #

      Sorry, I don’t have an example of PSO for fitting neural network weights.

  83. Ananya Mohapatra March 16, 2017 at 10:03 pm #

    hello sir,
    what type of neural network is used in this code? as there are 3 types of Neural network that are… feedforward, radial basis function and recurrent neurak network.

    • Jason Brownlee March 17, 2017 at 8:28 am #

      A multilayer perceptron (MLP) neural network. A classic type from the 1980s.

  84. Diego March 17, 2017 at 3:58 am #

    got this error while compiling..

    sigmoid_cross_entropy_with_logits() got an unexpected keyword argument ‘labels’

    • Jason Brownlee March 17, 2017 at 8:30 am #

      Perhaps confirm that your libraries are all up to date (Keras, Theano or TensorFlow)?

  85. Rohan March 20, 2017 at 5:20 am #

    Hi Jason!

    I am trying to use two odd frames of a video to predict the even one. Thus I need to give two images as input to the network and get one image as output. Can you help me with the syntax for the first model.add()? I have X_train of dimension (190, 2, 240, 320, 3) where 190 are the number of odd pairs, 2 are the two odd images, and (240,320,3) are the (height, width, depth) of each image.

  86. Herli Menezes March 21, 2017 at 8:33 am #

    Hello, Jason,
    Thanks for your good tutorial. However i found some issues:
    Warnings like these:

    1 – Warning (from warnings module):
    File “/usr/lib/python2.7/site-packages/keras/legacy/”, line 86
    call to the Keras 2 API: ' + signature)
    UserWarning: Update your
    Dense call to the Keras 2 API: Dense(12, activation=”relu”, kernel_initializer=”uniform”, input_dim=8)

    2 - Warning (from warnings module):
    File "/usr/lib/python2.7/site-packages/keras/legacy/", line 86
    call to the Keras 2 API: ‘ + signature)
    UserWarning: Update your Dense call to the Keras 2 API: Dense(8, activation="relu", kernel_initializer="uniform")

    3 – Warning (from warnings module):
    File “/usr/lib/python2.7/site-packages/keras/legacy/”, line 86
    call to the Keras 2 API: ' + signature)
    UserWarning: Update your
    Dense call to the Keras 2 API: Dense(1, activation=”sigmoid”, kernel_initializer=”uniform”)

    3 - Warning (from warnings module):
    File "/usr/lib/python2.7/site-packages/keras/", line 826
    nb_epoch argument in fit '
    UserWarning: The
    nb_epoch argument in fit has been renamed epochs`.

    I think these are due to some package update..

    But, the output of predictions was an array of zeros…
    such as: [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, ….0.0]

    I am running in a Linux Machine, Fedora 24,
    Python 2.7.13 (default, Jan 12 2017, 17:59:37)
    [GCC 6.3.1 20161221 (Red Hat 6.3.1-1)] on linux2


    Thank you!

    • Jason Brownlee March 21, 2017 at 8:45 am #

      These look like warnings related to the recent Keras 2.0 release.

      They look like just warning and that you can still run the example.

      I do not know why you are getting all zeros. I will investigate.

  87. Ananya Mohapatra March 21, 2017 at 6:21 pm #

    hello sir,
    can you please help me build a recurrent neural network with the above given dataset. i am having a bit trouble in building the layers…

    • Jason Brownlee March 22, 2017 at 7:56 am #

      Hi Ananya ,

      The Pima Indian diabetes dataset is a binary classification problem. It is not appropriate for a Recurrent Neural Network as there is no sequence information to learn.

      • Ananya Mohapatra March 22, 2017 at 8:04 pm #

        sir so could you tell on which type of dataset would the recurrent neural network accurately work? i have the dataset of EEG signals of epileptic patients…will recurrent network work on this?

        • Jason Brownlee March 23, 2017 at 8:49 am #

          It may if it is regular enough.

          LSTMs are excellent at sequence problems that have regularity or clear signals to detect.

  88. Shane March 22, 2017 at 5:18 am #

    Hi Jason, I have a quick question related to an error I am receiving when running the code in the tutorial…

    When I run

    # Compile model
    model.compile(loss='binary_crossentropy', optimizer='adam', metrics=['accuracy'])

    Python returns the following error:

    sigmoid_cross_entropy_with_logits() got an unexpected keyword argument ‘labels’

    • Jason Brownlee March 22, 2017 at 8:09 am #

      Sorry, I have not seen this error Shane.

      Perhaps check that your environment is up to date with the latest versions of the deep learning libraries?

  89. Tejes March 24, 2017 at 1:04 am #

    Hi Jason,
    Thanks for this awesome post.
    I ran your code with tensorflow back end, just out of curiosity. The accuracy returned was different every time I ran the code. That didn’t happen with Theano. Can you tell me why?

    Thanks in advance!

  90. Saurabh Bhagvatula March 27, 2017 at 9:49 pm #

    Hi Jason,
    I’m new to deep learning and learning it from your tutorials, which previously helped me understand Machine Learning very well.
    In the following code, I want to know why the number of neurons differ from input_dim in first layer of Nueral Net.
    # create model
    model = Sequential()
    model.add(Dense(12, input_dim=8, init=’uniform’, activation=’relu’))
    model.add(Dense(8, init=’uniform’, activation=’relu’))
    model.add(Dense(1, init=’uniform’, activation=’sigmoid’))

    • Jason Brownlee March 28, 2017 at 8:22 am #

      You can specify the number of inputs via “input_dim”, you can specify the number of neurons in the first hidden layer as the first parameter to Dense().

      • Saurabh Bhagvatula March 28, 2017 at 4:15 pm #

        Thanx a lot.

  91. Nalini March 29, 2017 at 2:52 am #

    Hi Jason

    while running this code for k fold cross validation it is not working.please give the code for k fold cross validation in binary class

  92. trangtruong March 29, 2017 at 7:04 pm #

    Hi Jason, why i use function evaluate to get accuracy score my model with test dataset, it return result >1, i can’t understand.

  93. enixon April 3, 2017 at 3:08 am #

    Hey Jason, thanks for this great article! I get the following error when running the code above:

    TypeError: Received unknown keyword arguments: {‘epochs’: 150}

    Any ideas on why that might be? I can’t get ‘epochs’, nb_epochs, etc to work…

    • Jason Brownlee April 4, 2017 at 9:07 am #

      You need to update to Keras version 2.0 or higher.

  94. Ananya Mohapatra April 5, 2017 at 9:30 pm #

    def baseline_model():
    # create model
    model = Sequential()
    model.add(Dense(10, input_dim=25, init=’normal’, activation=’softplus’))
    model.add(Dense(3, init=’normal’, activation=’softmax’))
    # Compile model
    model.compile(loss=’mean_squared_error’, optimizer=’adam’, metrics=[‘accuracy’])
    return model
    sir here mean_square_error has been used for loss calculation. Is it the same as LMS algorithm. If not, can we use LMS , NLMS or RLS to calculate the loss?

  95. Ahmad Hijazi April 5, 2017 at 10:19 pm #

    Hello Jason, thank you a lot for this example.

    My question is, after I trained the model and an accuracy of 79.2% for example is obtained successfully, how can I test this model on new data?

    for example if a new patient with new records appear, I want to guess the result (0 or 1) for him, how can I do that in the code?

    • Jason Brownlee April 9, 2017 at 2:36 pm #

      You can fit your model on all available training data then make predictions on new data as follows:

  96. Perick Flaus April 6, 2017 at 12:16 am #

    Thanks Jason, how can we test if new patient will be diabetic or no (0 or 1) ?

    • Jason Brownlee April 9, 2017 at 2:36 pm #

      Fit the model on all training data and call:

  97. Gangadhar April 12, 2017 at 1:28 am #

    Dr Jason,

    In compiling the model i got below error

    TypeError: compile() got an unexpected keyword argument ‘metrics’

    unable to resolve the below error

    • Jason Brownlee April 12, 2017 at 7:53 am #

      Ensure you have the latest version of Keras, v2.0 or higher.

  98. Omogbehin Azeez April 13, 2017 at 1:48 am #

    Hello sir,
    Thank you for the post. A quick question, my dataset has 24 input and 1 binary output( 170 instances, 100 epoch , hidden layer=6 and 10 batch, kernel_initializer=’normal’) . I adapted your code using Tensor flow and keras. I am having an accuracy of 98 to 100 percent. I am scared of over-fitting in my model. I need your candid advice. Kind regards sir

    • Jason Brownlee April 13, 2017 at 10:07 am #

      Yes, evaluate your model using k-fold cross-validation to ensure you are not tricking yourself.

      • Omogbehin Azeez April 14, 2017 at 1:08 am #

        Thank you sir

  99. Sethu Baktha April 13, 2017 at 5:19 am #

    Hi Jason,
    If I want to use the diabetes dataset (NOT Pima) to predict Blood Glucose which tutorials and e-books of yours would I need to start with…. Also, the data in its current format with time, code and value is it usable as is or do I need to convert the data in another format to be able to use it.

    Thanks for your help

    • Jason Brownlee April 13, 2017 at 10:13 am #

      This process will help you frame and work through your dataset:

      I hope that helps as a start.

      • Sethu Baktha April 13, 2017 at 10:25 am #

        Dr. Jason,
        The data is time series(time based data) with categorical(20) with two numbers one for insulin level and another for blood sugar level… Each time series data does not have every categorical data… For example one category is blood sugar before breakfast, another category is blood sugar after breakfast, before lunch and after lunch… Some times some of these category data is missing… I read through the above link, but does not talk about time series, categorical data with some category of data missing what to do in those cases…. Please let me know if any of your books will help clarify these points?

  100. Omogbehin Azeez April 14, 2017 at 9:49 am #

    Hello sir,

    Is it compulsory to normalize the data before using ANN model. I read it somewhere I which the author insisted that each attribute be comparable on the scale of [0,1] for a meaningful model. What is your take on that sir. Kind regards.

    • Jason Brownlee April 15, 2017 at 9:29 am #

      Yes. You must scale your data to the bounds of the activation used.

  101. shiva April 14, 2017 at 10:38 am #

    Hi Jason, You are simply awesome. I’m one of the many who got benefited from your book “machine learning mastery with python”. I’m working with a medical image classification problem. I have two classes of medical images (each class having 1000 images of 32*32) to be worked upon by the convolutional neural networks. Could you guide me how to load this data to the keras dataset? Or how to use my data while following your simple steps? kindly help.

    • Jason Brownlee April 15, 2017 at 9:30 am #

      Load the data as numpy arrays and then you can use it with Keras.

  102. Omogbehin Azeez April 18, 2017 at 12:09 am #

    Hello sir,

    I adapted your code with the cross validation pipelined with ANN (Keras) for my model. It gave me 100% still. I got the data from UCI ( Chronic Kidney Disease). It was 400 instances, 24 input attributes and 1 binary attribute. When I removed the rows with missing data I was left with 170 instances. Is my dataset too small for (24 input layer, 24 hidden layer and 1 output layer ANN, using adam and kernel initializer as uniform )?

    • Jason Brownlee April 18, 2017 at 8:32 am #

      It is not too small.

      Generally, the size of the training dataset really depends on how you intend to use the model.

      • Omogbehin Azeez April 18, 2017 at 11:10 pm #

        Thank you sir for the response, I guess I have to contend with the over-fitting of my model.

  103. Padmanabhan Krishnamurthy April 19, 2017 at 6:26 pm #

    Hi Jason,

    Great tutorial. Love the site 🙂
    Just a quick query : why have you used adam as an optimizer over sgd? Moreover, when do we use sgd optimization, and what exactly does it involve?


    • Jason Brownlee April 20, 2017 at 9:23 am #

      ADAM seems to consistently work well with little or no customization.

      SGD requires configuration of at least the learning rate and momentum.

      Try a few methods and use the one that works best for your problem.

      • Padmanabhan Krishnamurthy April 20, 2017 at 4:32 pm #

        Thanks 🙂

  104. Omogbehin Azeez April 25, 2017 at 8:13 am #

    Hello sir,

    Good day sir, how can I get all the weights and biases of the keras ANN. Kind regards.

  105. Shiva April 27, 2017 at 5:43 am #

    Hi Jason,
    I am currently working with the IMDB sentiment analysis problem as mentioned in your book. Am using Anaconda 3 with Python 3.5.2. In an attempt to summarize the review length as you have mentioned in your book, When i try to execute the command:

    result = map(len, X)
    print(“Mean %.2f words (%f)” % (numpy.mean(result), numpy.std(result)))

    it returns the error: unsupported operand type(s) for /: ‘map’ and ‘int’

    kindly help with the modified syntax. looking forward…

    • Jason Brownlee April 27, 2017 at 8:47 am #

      I’m sorry to hear that. Perhaps comment out that line?
      Or change it to remove the formatting and just print the raw mean and stdev values for you to review?

  106. Elikplim May 1, 2017 at 1:58 am #

    Hello, quite new to Python, Numpy and Keras(background in PHP, MYSQL etc). If there are 8 input variables and 1 output varable(9 total), and the Array indexing starts from zero(from what I’ve gathered it’s a Numpy Array, which is built on Python lists) and the order is [rows, columns], then shouldn’t our input variable(X) be X = dataset[:,0:7] (where we select from the 1st to 8th columns, ie. 0th to 7th indices) and output variable(Y) be Y = dataset[:,8] (where we the 9th column, ie. 8th index)?

  107. Jackie Lee May 1, 2017 at 12:47 pm #

    I’m having troubles with the predictions part. It saves ValueError: Error when checking model input: expected dense_1_input to have shape (None, 502) but got array with shape (170464, 502)

    testset = numpy.loadtxt(“right_stim_FD1.csv”, delimiter=”,”)
    A = testset[:,0:502]
    B = testset[:,502]
    probabilities = model.predict(A, batch_size=10, verbose=1)
    predictions = float(round(a) for a in probabilities)
    accuracy = numpy.mean(predictions == B)
    #round predictions
    #rounded = [round(x[0]) for x in predictions]
    print(“Prediction Accuracy: %.2f%%” % (accuracy*100))

    • Jason Brownlee May 2, 2017 at 5:55 am #

      It looks like you might be giving the entire dataset as the output (y) rather than just the output variable.

  108. Anastasios Selalmazidis May 2, 2017 at 12:27 am #

    Hi there,

    I have a question regarding deep learning. In this tutorial we build a MLP with Keras. Is this Deep Learning or is it just a MLP Backpropagation ?

  109. Eric T May 2, 2017 at 8:59 pm #

    Would you mind if I use this code as an example of a simple network in a school project of mine?
    Need to ask before using it, since I cannot find anywhere in this tutorial that you are OK with anyone using the code, and the ethics moment of my course requires me to ask (and of course give credit where credit is due).
    Kind regards
    Eric T

  110. BinhLN May 7, 2017 at 3:11 am #

    Hi Jason
    I have a problem
    My Dataset have 500 record. But My teacher want my dataset have 100.000 record. I must have a new algorithm for data generation. Please help me

  111. Dp May 11, 2017 at 2:26 am #

    Can you give a deep cnn code which includes 25 layers , in the first conv layer the filter sizs should be 39×39 woth a total lf 64 filters , in the 2nd conv layer , 21 ×21 with 32 filters , in the 3rd conv layer 11×11 with 64 filters , 4th Conv layer 7×7 with 32 layers . For a input size of image 256×256. Im Competely new in this Deep learning Thing but if you can code that for me it would be a great help. Thanks

  112. Maple May 13, 2017 at 12:58 pm #

    I have to follow with the facebook metrics. But the result is very low. Help me.
    I changed the input but did not improve

  113. Alessandro May 14, 2017 at 1:01 am #

    Hi Jason,

    Great Tutorial and thanks for your effort.

    I have a question, since I am beginner with keras and tensorflow.
    I have installed both of them, keras and tensorflow, the latest version and I have run your example but I get always the same error:

    Traceback (most recent call last):
    File “”, line 18, in
    model.compile(loss=’binary_crossentropy’, optimizer=’adam’, metrics=[‘accuracy’])
    File “/Users/MacBookPro1/.virtualenvs/keras_tf/lib/python2.7/site-packages/keras/”, line 777, in compile
    File “/Users/MacBookPro1/.virtualenvs/keras_tf/lib/python2.7/site-packages/keras/engine/”, line 910, in compile
    sample_weight, mask)
    File “/Users/MacBookPro1/.virtualenvs/keras_tf/lib/python2.7/site-packages/keras/engine/”, line 436, in weighted
    score_array = fn(y_true, y_pred)
    File “/Users/MacBookPro1/.virtualenvs/keras_tf/lib/python2.7/site-packages/keras/”, line 51, in binary_crossentropy
    return K.mean(K.binary_crossentropy(y_pred, y_true), axis=-1)
    File “/Users/MacBookPro1/.virtualenvs/keras_tf/lib/python2.7/site-packages/keras/backend/”, line 2771, in binary_crossentropy
    TypeError: sigmoid_cross_entropy_with_logits() got an unexpected keyword argument ‘labels’

    Could you help? Thanks


    • Jason Brownlee May 14, 2017 at 7:30 am #

      Ouch, I have not seen this error before.

      Some ideas:
      – Consider trying the theano backend and see if that makes a difference.
      – Try searching/posting on the keras user group and slack channel.
      – Try searching/posting on stackoverflow or cross validated.

      Let me know how you go.

      • Alessandro May 14, 2017 at 9:44 am #

        Hi Jason,

        I found the issue. The tensorflow installation was outdated; so I have updated it and everything
        is working nicely.

        Good night,

  114. Sheikh Rafiul Islam May 25, 2017 at 3:36 pm #

    Thank you Mr. Brownlee for your wonderful easy to understand explanation

  115. WAZED May 29, 2017 at 12:31 am #

    Hi Jason,
    Thank you very much for your wonderful tutorial. I have a question regarding the metrices.Is there default way to declare metrices “Precision” and “Recall” in addtion with the “Accurace”.


  116. chiranjib konwar May 29, 2017 at 4:30 am #

    Hi Jason,

    please send me a small note containing resources from where i can learn deep learning from scratch. thanks for the wonderful read you had prepared.

    Thanks in advance

    yes, my email id is [email protected]

  117. Jeff June 1, 2017 at 11:48 am #

    Why the NN have mistakes many times?

  118. kevin June 2, 2017 at 5:53 pm #

    Hi Jason,

    I seem to be getting an error when applying the fit method:

    ValueError: Error when checking input: expected dense_1_input to have shape (None, 12) but got array with shape (767, 8)

    I looked this up and the most prominent suggestion seemed to be upgrade keras and theno, which I did, but that didn’t resolve the problem.

    • Jason Brownlee June 3, 2017 at 7:24 am #

      Ensure you have copied the code exactly from the post.

  119. Hemanth Kumar K June 3, 2017 at 2:15 pm #

    hi Jason,
    I am stuck with an error
    TypeError: sigmoid_cross_entropy_with_logits() got an unexpected keyword argument ‘labels’
    my tensor flow and keras virsions are
    keras: 2.0.4
    Tensorflow: 0.12

    • Jason Brownlee June 4, 2017 at 7:46 am #

      I’m sorry to hear that, I have not seen that error before. Perhaps you could post a question to stackoverflow or the keras user group?

  120. xena June 4, 2017 at 6:36 pm #

    can anyone tell me which neural network is being used here? Is it MLP??

    • Jason Brownlee June 5, 2017 at 7:40 am #

      Yes, it is a multilayer perceptron (MLP) feedforward neural network.

  121. Nirmesh Shah June 9, 2017 at 11:00 pm #

    Hi Jason,

    I have run this code successfully on PC with CPU.

    If I have to run the same code n another PC which contains GPU, What line should I add to make it sure that it runs on the GPU

    • Jason Brownlee June 10, 2017 at 8:24 am #

      The code would stay the same, your configuration of the Keras backend would change.

      Please refer to TensorFlow or Theano documentation.

  122. Prachi June 12, 2017 at 7:30 pm #

    What if I want to train my neural which should detect whether the luggage is abandoned or not ? How do i proceed for it ?

  123. Ebtesam June 14, 2017 at 11:15 pm #

    I was build neural machine translation model but the score i was get is 0 i am not sure why

  124. Sarvottam Patel June 20, 2017 at 7:31 pm #

    HHey Jason , first of all thank you very much from the core of my heart to make me understand this perfectly, I have an error after completing 150 iteration.

    File “”, line 53, in
    print(“\n%s: %.2f” %(model.metrics_names[1]*100))
    TypeError: not enough arguments for format string

    • Sarvottam Patel June 20, 2017 at 8:05 pm #

      Sorry Sir my bad , actually I wrote it wrongly

    • Jason Brownlee June 21, 2017 at 8:12 am #

      Confirm that you have copied the line exactly:

  125. Joydeep June 30, 2017 at 4:15 pm #

    Hi Dr Jason,

    Thanks for the tutorial to get started using Keras.

    I used the below snippet to directly load the dataset from the URL rather than downloading and saving as this makes the code more streamlined without having to navigate elsewhere.

    # load pima indians dataset
    datasource = numpy.DataSource().open(“”)
    dataset = numpy.loadtxt(datasource, delimiter=”,”)

  126. Yvette July 7, 2017 at 9:01 pm #

    Thanks for this helpful resource!

  127. Andeep July 10, 2017 at 1:14 am #

    Hi Dr Brownlee,

    thank you very much for this great tutorial!
    I would be grateful, if you could answer some questions:

    1. What does the 7 in “numpy.random.seed(7)” means?

    2. In my case I have 3 input neurons and 2 output neurons. Is the correct notation:
    X = dataset[:,0:3]
    Y = dataset[:,3:4] ?

    3. The batch size means how many training data are used in one epoch, am I right?
    I have thought we have to use the whole training data set for the training. In this case I would determine the batch size as the number of training data pairs I have achieved through experiments etc.. In your example, does the batch (sized 10) means that the computer always uses the same 10 training data in every epoch or are the 10 training data randomly chosen among all training data before every epoch?

    4. When evaluating the model what does the loss means (e.g. in loss: 0.5105 – acc: 0.7396)?
    Is it the sum of values of the error function (e.g. mean_squared_error) of the output neurons?

  128. Patrick Zawadzki July 11, 2017 at 5:35 am #

    Is there anyway to see the relationship between these inputs? Essentially understand which inputs affect the output the most, or perhaps which pairs of inputs affect the output the most?

    Maybe pairing this with unsupervised deep learning? I want to have less of a “black box” for the developed network if at all possible. Thank you for your great content!

  129. Bernt July 13, 2017 at 10:12 pm #

    Hi Jason,
    Thank you for sharing your skills and competence.

    I want to study the change in weights and predictions between each epoch run.
    Have tried to use the model.train_on_batch method and the method with epoch=1 and batch_size equal all the samples.

    But it seems like the model doesn’t save the new updated weights.
    I print predictions before and after I dont see a change in the evaluation scores.

    Parts of the code is printed below.

    Any idea?

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

    # evaluate the model
    scores = model.evaluate(X, Y)
    print(“\n%s: %.2f%%” % (model.metrics_names[1], scores[1]*100))

    # Run one update of the model trained run with X and compared with Y
    model.train_on_batch(X, Y)

    # Fit the model, Y, epochs=1, batch_size=768)

    scores = model.evaluate(X, Y)
    print(“\n%s: %.2f%%” % (model.metrics_names[1], scores[1]*100))

    • Jason Brownlee July 14, 2017 at 8:29 am #

      Sorry, I have not explored evaluating a Keras model this way.

      Perhaps it is a fault, I would recommend preparing the smallest possible example that demonstrates the issue and post to the Keras GitHub issues.

  130. iman July 18, 2017 at 11:18 pm #

    Hi, I tried to apply this to the titanic data set, however the predictions were all 0.4. What do you suggest for:
    # create model
    model = Sequential()
    model.add(Dense(12, input_dim=4, activation=’relu’))
    model.add(Dense(4, activation=’relu’))
    model.add(Dense(1, activation=’sigmoid’))

    model.compile(loss=’binary_crossentropy’, optimizer=’adam’, metrics=[‘accuracy’]) #’sgd’, Y, epochs=15, batch_size=10)

  131. Camus July 19, 2017 at 2:14 am #

    Hi Dr Jason,
    This is probably a stupid question but I cannot find out how to do it … and I am beginner on Neural Network.
    I have relatively same number of inputs (7) and one output. This output can take numbers between -3000 and +3000.
    I want to build a neural network model in python but I don’t know how to do it.
    Do you have an example with outputs different from 0-1.
    Tanks in advance


    • Jason Brownlee July 19, 2017 at 8:28 am #

      Ensure you scale your data then use the above tutorial to get started.

  132. Khalid Hussain July 21, 2017 at 11:28 pm #

    Hi Jason Brownlee

    I am using the same data “pima-indians-diabetes.csv” but all predicted values are less then 1 and are in fraction which could not distinguish any class.

    If I round off then all become 0.

    I am using model.predict(x) function

    You are requested to kindly guide me what I am doing wrong are how can I achieve correct predicted value.

    Thank you

    • Jason Brownlee July 22, 2017 at 8:36 am #

      Consider you have copied all of the code exactly from the tutorial.

  133. Ludo July 25, 2017 at 6:59 pm #

    Hello Jason,

    Thanks you for your great example. I have some comments.

    – Why you have choice “12” inputs hidden layers ? and not 24 / 32 .. it’s arbitary ?
    – Same question about epochs and batch_size ?

    This value are very sensible !! i have try with 32 inputs first layer , epchos=500 and batch_size=1000 and the result is very differents… i’am at 65% accurancy.

    Thx for you help.

    • Jason Brownlee July 26, 2017 at 7:50 am #

      Yes, it is arbitrary. Tune the parameters of the model to your problem.

  134. Almoutasem Bellah Rajab July 25, 2017 at 7:32 pm #

    Wow, you’re still replying to comments more than a year later!!!… you’re great,, thanks..

  135. Jane July 26, 2017 at 1:23 am #

    Thanks for your tutorial, I found it very useful to get me started with Keras. I’ve previously tried TensorFlow, but found it very difficult to work with. I do have a question for you though. I have both Theano and TensorFlow installed, how do I know which back-end Keras is using? Thanks again

    • Jason Brownlee July 26, 2017 at 8:02 am #

      Keras will print which backend it uses every time you run your code.

      You can change the backend in the Keras configuration file (~/.keras/keras.json) which looks like:

  136. Masood Imran July 28, 2017 at 12:00 am #

    Hello Jason,

    My understanding of Machine Learning or evaluating deep learning models is almost 0. But, this article gives me lot of information. It is explained in a simple and easy to understand language.

    Thank you very much for this article. Would you suggest any good read to further explore Machine Learning or deep learning models please?

  137. Peggy August 3, 2017 at 7:14 pm #

    If I have trained prediction models or neural network function scripts. How can I use them to make predictions in an application that will be used by end users? I want to use python but it seems I will have to redo the training in Python again. Is there a way I can rewrite the scripts in Python without retraining and just call the function of predicting?

  138. Shane August 8, 2017 at 2:38 pm #

    Jason, I used your tutorial to install everything needed to run this tutorial. I followed your tutorial and ran the resulting program successfully. Can you please describe what the output means? I would like to thank you for your very informative tutorials.

    • Shane August 8, 2017 at 2:39 pm #

      768/768 [==============================] – 0s – loss: 0.4807 – acc: 0.7826
      Epoch 148/150
      768/768 [==============================] – 0s – loss: 0.4686 – acc: 0.7812
      Epoch 149/150
      768/768 [==============================] – 0s – loss: 0.4718 – acc: 0.7617
      Epoch 150/150
      768/768 [==============================] – 0s – loss: 0.4772 – acc: 0.7812
      32/768 [>………………………..] – ETA: 0s
      acc: 77.99%

      • Jason Brownlee August 8, 2017 at 5:12 pm #

        It is summarizing the training of the model.

        The final line evaluates the accuracy of the model’s predictions – really just to demonstrate how to make predictions.

    • Jason Brownlee August 8, 2017 at 5:11 pm #

      Well done Shane.

      Which output?

  139. Bene August 9, 2017 at 1:02 am #

    Hello Jason, i really liked your Work and it helped me a lot with my first steps.

    But i am not really familiar with the numpy stuff:

    So here is my Question:

    dataset = numpy.loadtxt(“pima-indians-diabetes.csv”, delimiter=”,”)
    # split into input (X) and output (Y) variables
    X = dataset[:,0:8]
    Y = dataset[:,8]

    I get that the numpy.loadtxt is extracting the information from the cvs File

    but what does the stuff in the Brackets mean like X = dataset[:,0:8]

    why the “:” and why , 0:8

    its probably pretty dumb but i can’t find a good explanation online 😀

    thanks really much!

  140. Chen August 12, 2017 at 5:43 pm #

    Can I translate it to Chinese and put it to Internet in order to let other Chinese people can read your article?

  141. Deep Learning August 12, 2017 at 7:36 pm #

    It seems that using this line:


    …is redundant i.e. the Keras output in a loop running the same model with the same configuration will yield a similar variety of results regardless if it’s set at all, or which number it is set to. Or am I missing something?

    • Jason Brownlee August 13, 2017 at 9:52 am #

      Deep learning algorithms are stochastic (random within a range). That means that they will make different predictions/learn different things when the same model is trained on the same data. This is a feature:

      You can fix the random seed to ensure you get the same result, and it is a good idea for tutorials to help beginners out:

      When evaluating the skill of a model, I would recommend repeating the experiment n times and taking skill as the average of the runs. See here for the procedure:

      Does that help?

      • Deep Learning August 14, 2017 at 3:08 am #

        Thanks Jason 🙂

        I totally get what it should do, but as I had pointed out, it does not do it. If you run the codes you have provided above in a loop for say 10 times. First 10 with random seed set and the other 10 times without that line of code all together. Then compare the result. At least the result I’m getting, is suggesting the effect is not there i.e. both sets of 10 times will have similar variation in the result.

        • Jason Brownlee August 14, 2017 at 6:26 am #

          It may suggest that the model is overprescribed and easily addresses the training data.

  142. Deep Learning August 14, 2017 at 3:12 am #

    Nice post by the way >

    Thanks for sharing it. Been lately thinking about the aspect of accuracy a lot, it seems that at the moment it’s a “hot mess” in terms of the way common tools do it out of the box. I think a lot of non PhD / non expert crowd (most people) will at least initially be easily confused and make the kinds of mistakes you point out in your post.

    Thanks for all the amazing contributions you are making in this field!

    • Jason Brownlee August 14, 2017 at 6:26 am #

      I’m glad it helped.

      • Haneesh December 7, 2019 at 10:36 pm #

        Hi Jason,

        i’m actually trying to find “spam filter for quora questions” where i have a dataset with label-0’s and 1’s and questions columns. please let me know the approach and path to build a model for this.


  143. RATNA NITIN PATIL August 14, 2017 at 8:16 pm #

    Hello Jason, Thanks for a wonderful tutorial.
    Can I use Genetic Algorithm for feature selection??
    If yes, Could you please provide the link for it???
    Thanks in advance.

    • Jason Brownlee August 15, 2017 at 6:34 am #

      Sure. Sorry, I don’t have any examples.

      Generally, computers are so fast it might be easier to test all combinations in an exhaustive search.

  144. sunny1304 August 15, 2017 at 3:44 pm #

    Hi Json,
    Thank you for your awesome tutorial.
    I have a question for you.

    Is there any guideline on how to decide on neuron number for our network.
    for example you used 12 for thr 1st layer and 8 for the second layer.
    how do you decide on that ?


    • Jason Brownlee August 15, 2017 at 4:58 pm #

      No, there is no way to analytically determine the configuration of the network.

      I use trial and error. You can grid search, random search, or copy configurations from tutorials or papers.

  145. yihadad August 16, 2017 at 6:53 pm #

    Hi Json,
    Thanks for a wonderful tutorial.

    Run a model generated by a CNN it takes how much ram, cpu ?


    • Jason Brownlee August 17, 2017 at 6:39 am #

      It depends on the data you are using to fit the model and the size of the model.

      Very large models could be 500MB of RAM or more.

  146. Ankur September 1, 2017 at 3:15 am #

    Hi ,
    Please let me know , how can i visualise the complete neural network in Keras……………….

    I am looking for the complete architecture – like number of neurons in the Input Layer, hidden layer , output layer with weights.

    Please have a look at the link present below, here someone has created a beutiful visualisation/architecture using neuralnet package in R.
    Please let me know, can we create such type of model in KERAS

    • Jason Brownlee September 1, 2017 at 6:50 am #

      Use the Keras visualization API:

    • ASAD October 17, 2017 at 3:23 am #

      Hello ANKUR,,,, how are you?

      you have try visualization in keras which is suggested by Jason Brownlee?
      if you have tried then please send me code i am also trying but didnot work..

      please guide me

  147. Adam September 3, 2017 at 1:45 am #

    Thank you Dr. Brownlee for the great tutorial,

    I have a question about your code:
    is the argument metrics=[‘accuracy’] necessary in the code and does it change the results of the neural network or is it just for showing me the accuracy during compiling?

    thank you!!

  148. PottOfGold September 5, 2017 at 12:14 am #

    Hi Jason,

    your work here is really great. It helped me a lot.
    I recently stumbled upon one thing I cannot understand:

    For the pimas dataset you state:
    When I look at the table of the pimas dataset, the examples are in rows and the features in columns, so your input dimension is the number of columns. As far as I can see, you don’t change the table.

    For neural networks, isn’t the input normally: examples = columns, features=rows?
    Is this different for Keras? Or can I use both shapes? An if yes, what’s the difference in the construction of the net?

    Thank you!!

    • Jason Brownlee September 7, 2017 at 12:36 pm #

      No, features are columns, rows are instances or examples.

      • PottOfGold September 7, 2017 at 3:35 pm #

        Thanks! 🙂
        I had a lot of discussions because of that.
        In Andrew Ng new Coursera course it’s explained as examples = columns, features=rows, but he doesn’t use Keras of course, but programms the neural networks from scratch.

        • Jason Brownlee September 9, 2017 at 11:38 am #

          I doubt that, I think you may have mixed it up. Columns are never examples.

          • PottOfGold October 6, 2017 at 6:26 pm #

            Thats what I thought, but I looked it up in the notation for the new coursera course ( and there it says: m is the numer of examples in the dataset and n is the input size, where X superscript n x m is the input matrix …
            But either way, you helped me! Thank you. 🙂

  149. Lin Li September 16, 2017 at 1:50 am #

    Hi Jason, thank you so much for your tutorial, it helps me a lot. I need your help for the question below:
    I copy the code and run it. Although I got the classification results, there were some warning messages in the process. As follows:

    Warning (from warnings module):
    File “C:\Users\llfor\AppData\Local\Programs\Python\Python35\lib\site-packages\keras\”, line 120
    % delta_t_median)
    UserWarning: Method on_batch_end() is slow compared to the batch update (0.386946). Check your callbacks.

    I don’t know why, and cannot find any answer to this question. I’m looking forward to your reply. Thanks again!

    • Jason Brownlee September 16, 2017 at 8:43 am #

      Sorry, I have not seen this message before. It looks like a warning, you might be able to ignore it.

      • Lin Li September 16, 2017 at 12:24 pm #

        Thanks for your reply. I’m a start-learner on deep learning.I’d like to put it aside temporarily.

  150. Sagar September 22, 2017 at 2:51 pm #

    Hi Jason,
    Great article, thumbs up for that. I am getting this error when I try to run the file on the command prompt. Any suggestions. Thanks for you response.

    Using TensorFlow backend.
    2017-09-22 10:11:11.189829: W C:\tf_jenkins\home\workspace\rel-win\M\windows\PY\
    36\tensorflow\core\platform\] The TensorFlow library wasn
    ‘t compiled to use AVX instructions, but these are available on your machine and
    could speed up CPU computations.
    2017-09-22 10:11:11.190829: W C:\tf_jenkins\home\workspace\rel-win\M\windows\PY\
    36\tensorflow\core\platform\] The TensorFlow library wasn
    ‘t compiled to use AVX2 instructions, but these are available on your machine an
    d could speed up CPU computations.
    32/768 [>………………………..] – ETA: 0s
    acc: 78.52%

    • Jason Brownlee September 23, 2017 at 5:35 am #

      Looks like warning messages that you can ignore.

      • Sagar September 24, 2017 at 3:52 am #

        Thanks I got to know what the problem was. According to section 6 I had set verbose argument to 0 while calling “”. Now all the epochs are getting printed.

  151. Valentin September 26, 2017 at 6:35 pm #

    Hi Jason,

    Thanks for the amazing article . Clear and straightforward.
    I had some problems installing Keras but was advised to prefix
    with tf.contrib.keras
    so I have code like


    Now I try to train Keras on some small datafile to see how things work out:

    The first 4 columns are inputs and the 5-th column is output.
    I use the same code for training (adjust number of inputs) as in your article,
    but the network only gets to 12.5% accuracy.
    Any advise?


  152. Priya October 3, 2017 at 2:28 pm #

    Hi Jason,

    I tried replacing the pima data with random data as follows:

    X_train = np.random.rand(18,61250)
    X_test = np.random.rand(18,61250)
    Y_train = np.array([0.0, 1.0, 1.0, 0.0, 1.0, 1.0, 1.0, 0.0, 1.0,
    0.0, 1.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 1.0,])
    Y_test = np.array([1.0, 0.0, 0.0, 1.0, 1.0, 0.0, 1.0, 1.0, 1.0,
    1.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 1.0, 0.0,])

    _, input_size = X_train.shape #put this in input_dim in the first dense layer

    I took the round() off of the predictions so I could see the full value and then inserted my random test data in

    predictions = model.predict(X_test)
    preds = [x[0] for x in predictions]
    print(preds), Y_train, epochs=100, batch_size=10, verbose=2, validation_data=(X_test,Y_test))

    I found something slightly odd; I expected the predicted values to be around 0.50, plus or minus some, but instead, I got this:

    [0.49525392, 0.49652839, 0.49729034, 0.49670222, 0.49342978, 0.49490061, 0.49570397, 0.4962129, 0.49774086, 0.49475089, 0.4958384, 0.49506786, 0.49696651, 0.49869373, 0.49537542, 0.49613148, 0.49636957, 0.49723724]

    which is near 0.50 but always less than 0.50. I ran this a few times with different random seeds, so it’s not coincidental. Would you have any explanation for why it does this?


    • Jason Brownlee October 3, 2017 at 3:46 pm #

      Perhaps calculate the mean of your training data and compare it to the predicted value. It might be simple sampling error.

    • Priya October 4, 2017 at 1:02 am #

      I found out I was doing predictions before fitting the model. (I suppose that would mean the network hadn’t adjusted to the data’s distribution yet.)

  153. Saurabh October 7, 2017 at 5:59 am #

    Hello Jason,

    I tried to train this model on my laptop, it is working fine. But I tried to train this model on google-cloud with the same instructions as in your example-5. But it is failing.
    Can you just let me know, which changes are to required for the model, so that I can train this on cloud.

  154. tobegit3hub October 12, 2017 at 6:40 pm #

    Great post. Thanks for sharing.

  155. Manoj October 12, 2017 at 11:43 pm #

    Hi Jason,
    Is there a way to store the model, once it is created so that I can use it for different input data sets as and when needed.

  156. Cam October 23, 2017 at 6:11 pm #

    I get a syntax error for the line in this example. Is it due to library conflicts with theano and tensorflow if i have both installed?

  157. Diego Quintana October 25, 2017 at 7:37 am #

    Hi Jason, thanks for the example.

    How would you predict a single element from X? X[0] raises a ValueError

    ValueError: Error when checking : expected dense_1_input to have shape (None, 8) but got array with shape (8, 1)


  158. Shahbaz Wasti October 28, 2017 at 1:30 pm #

    Dear Sir ,
    I have installed and configured the environment according to your directions but while running the program i have following error

    “from keras.utils import np_utils”

  159. Zhengping October 30, 2017 at 12:12 am #

    Hi Jason, thanks for the great tutorials. I just learnt and repeated the program in your “Your First Machine Learning Project in Python Step-By-Step” without problem. Now trying this one, getting stuck at the line “model = Sequential()” when the Interactive window throws: NameError: name ‘Sequential’ is not defined. tried to google, can’t find a solution. I did import Sequential from keras.models as in ur example code. copy pasted as it is. Thanks in advance for your help.

    • Zhengping October 30, 2017 at 12:14 am #

      I’m running ur examples in Anaconda 4.4.0 environment in visual studio community version. relevant packages have been installed as in ur earlier tutorials instructed.

      • Zhengping October 30, 2017 at 12:18 am #

        >> # create model
        … model = Sequential()

        Traceback (most recent call last):
        File “”, line 2, in
        NameError: name ‘Sequential’ is not defined
        >>> model.add(Dense(12, input_dim=8, init=’uniform’, activation=’relu’))
        Traceback (most recent call last):
        File “”, line 1, in
        AttributeError: ‘SVC’ object has no attribute ‘add’

    • Jason Brownlee October 30, 2017 at 5:38 am #

      Looks like you need to install Keras. I have a tutorial here on how to do that:

  160. Akhil October 30, 2017 at 5:04 pm #

    Ho Jason,

    Thanks a lot for this wonderful tutorial.

    I have a question:

    I want to use your code to predict the classification (1 or 0) of unknown samples. Should I create one common csv file having the train (known) as well as the test (unknown) data. Whereas the ‘classification’ column for the known data will have a known value, 1 or 0, for the unknown data, should I leave the column empty (and let the code decide the outcome)?

    Thanks a lot

  161. Guilherme November 3, 2017 at 1:26 am #

    Hi Jason,

    This is really cool! I am blown away! Thanks so much for making it so simple for a beginner to have some hands on. I have a couple questions:

    1) where are the weights, can I save and/or retrieve them?

    2) if I want to train images with dogs and cats and later ask the neural network whether a new image has a cat or a dog, how do I get my input image to pass as an array and my output result to be “cat” or “dog”?

    Thanks again and great job!

  162. Michael November 5, 2017 at 8:33 am #

    Hi Jason,

    Are you familiar with a python tool/package that can build neural network as in the tutorial, but suitable for data stream mining?


  163. bea November 8, 2017 at 1:58 am #

    Hi, there. Could you please clarify why exactly you’ve built your network with 12 neurons in the first layer?

    “The first layer has 12 neurons and expects 8 input variables. The second hidden layer has 8 neurons and finally, the output layer has 1 neuron to predict the class (onset of diabetes or not)…”

    Should’nt it have 8 neurons at the start?


    • Jason Brownlee November 8, 2017 at 9:28 am #

      The input layer has 8, the first hidden layer has 12. I chose 12 through a little trial and error.

  164. Guilherme November 9, 2017 at 12:54 am #

    Hi Jason,

    Do you have or else could you recommend a beginner’s level image segmentation approach that uses deep learning? For example, I want to train some neural net to automatically “find” a particular feature out of an image.


    • Jason Brownlee November 9, 2017 at 10:00 am #

      Sorry, I don’t have image segmentation examples, perhaps in the future.

  165. Andy November 12, 2017 at 6:56 pm #

    Hi Jason,

    I just started my DL training a few weeks ago. According to what I learned in course, in order to train the parameters for the NN, we need to run the Forward and Backward propagation; however, looking at your Keras example, i don’t find any of these propagation processes. Does it mean that Keras has its own mechanism to find the parameters instead of using Forward and Backward propagation?


    • Jason Brownlee November 13, 2017 at 10:13 am #

      It is performing those operations under the covers for you.

  166. Badr November 13, 2017 at 11:42 am #

    Hi Jason,

    Can you explain why I got the following output:

    ValueError Traceback (most recent call last)
    in ()
    —-> 1 model.compile(loss=’binary_crossentropy’, optimizer=’adam’, metrics=[‘accuracy’])
    2, Y, epochs=150, batch_size=10)
    3 scores = model.evaluate(X, Y)
    4 print(“\n%s: %.2f%%” % (model.metrics_names[1], scores[1]*100))

    /Users/badrshomrani/anaconda/lib/python3.5/site-packages/keras/ in compile(self, optimizer, loss, metrics, sample_weight_mode, **kwargs)
    545 metrics=metrics,
    546 sample_weight_mode=sample_weight_mode,
    –> 547 **kwargs)
    548 self.optimizer = self.model.optimizer
    549 self.loss = self.model.loss

    /Users/badrshomrani/anaconda/lib/python3.5/site-packages/keras/engine/ in compile(self, optimizer, loss, metrics, loss_weights, sample_weight_mode, **kwargs)
    620 loss_weight = loss_weights_list[i]
    621 output_loss = weighted_loss(y_true, y_pred,
    –> 622 sample_weight, mask)
    623 if len(self.outputs) > 1:
    624 self.metrics_tensors.append(output_loss)

    /Users/badrshomrani/anaconda/lib/python3.5/site-packages/keras/engine/ in weighted(y_true, y_pred, weights, mask)
    322 def weighted(y_true, y_pred, weights, mask=None):
    323 # score_array has ndim >= 2
    –> 324 score_array = fn(y_true, y_pred)
    325 if mask is not None:
    326 # Cast the mask to floatX to avoid float64 upcasting in theano

    /Users/badrshomrani/anaconda/lib/python3.5/site-packages/keras/ in binary_crossentropy(y_true, y_pred)
    47 def binary_crossentropy(y_true, y_pred):
    —> 48 return K.mean(K.binary_crossentropy(y_pred, y_true), axis=-1)

    /Users/badrshomrani/anaconda/lib/python3.5/site-packages/keras/backend/ in binary_crossentropy(output, target, from_logits)
    1418 output = tf.clip_by_value(output, epsilon, 1 – epsilon)
    1419 output = tf.log(output / (1 – output))
    -> 1420 return tf.nn.sigmoid_cross_entropy_with_logits(output, target)

    /Users/badrshomrani/anaconda/lib/python3.5/site-packages/tensorflow/python/ops/ in sigmoid_cross_entropy_with_logits(_sentinel, labels, logits, name)
    147 # pylint: disable=protected-access
    148 nn_ops._ensure_xent_args(“sigmoid_cross_entropy_with_logits”, _sentinel,
    –> 149 labels, logits)
    150 # pylint: enable=protected-access

    /Users/badrshomrani/anaconda/lib/python3.5/site-packages/tensorflow/python/ops/ in _ensure_xent_args(name, sentinel, labels, logits)
    1696 if sentinel is not None:
    1697 raise ValueError(“Only call %s with ”
    -> 1698 “named arguments (labels=…, logits=…, …)” % name)
    1699 if labels is None or logits is None:
    1700 raise ValueError(“Both labels and logits must be provided.”)

    ValueError: Only call sigmoid_cross_entropy_with_logits with named arguments (labels=…, logits=…, …)

    • Jason Brownlee November 14, 2017 at 10:05 am #

      Perhaps double check you have the latest versions of the keras and tensorflow libraries installed?!

  167. Badr November 14, 2017 at 10:50 am #

    keras was outdated

  168. Mikael November 22, 2017 at 8:20 am #

    Hi Jason, thanks for your short tutorial, helps a lot to actually get your hands dirty with a simple example.
    I have tried 5 different parameters and got some interesting results to see what would happen. Unfortunately, I didnt record running time.

    Test 1 Test 2 Test 3 Test 4 Test 5 Test 6 Test 7
    number of layers 3 3 3 3 3 3 4
    Train set 768 768 768 768 768 768 768
    Iterations 150 100 1000 1000 1000 150 150
    Rate of update 10 10 10 5 1 1 5
    Errors 173 182 175 139 161 169 177
    Values 768 768 768 768 768 768 768
    % Error 23,0000% 23,6979% 22,7865% 18,0990% 20,9635% 22,0052% 23,0469%

    I can’t seem to see a trend here.. That could put me on the right track to adjust my hyperparameters.

    Do you have any advice on that?

  169. Nikolaos November 28, 2017 at 10:58 am #

    Hi, I try to implement the above example with fer2013.csv but I receive an error, it is possible to help me to implement this correctly?

    • Jason Brownlee November 29, 2017 at 8:10 am #

      Sorry, I cannot debug your code.

      What is the problem exactly?

  170. Tanya December 2, 2017 at 12:06 am #

    i have a a bit general question.
    I have to do a forecasting for restaurant sales (meaning that I have to predict 4 meals based on a historical daily sales data), weather condition (such as temperature, rain, etc), official holiday and in-off-season. I have to perform that forecasting using neuronal networks.
    I am unfortunately not a very skilled in python. On my computer I have Python 2.7 and I have install anaconda. I am trying to learn exercising with your codes, Mr. Brownlee. But somehow I can not run the code at all (in Spyder). Can you tell me what kind of version of python and anaconda I have to install on my computer and in which environment (jupiterlab,notebook,qtconsole, spyder, etc) I can run the code, so to work and not to give error from the very beginning?
    I will be very thankful for your response

  171. Eliah December 3, 2017 at 10:53 am #

    Hi Dr. Brownlee.

    I looked over the tutorial and I had a question regarding reading the data from a binary file? For instance I working on solving the sliding tiled n-puzzle using neural networks, but I seem to have trouble to getting my data which is in a binary file and it generates the number of move required for the n-puzzle to be solve in. Am not sure if you have dealt with this before, but any help would be appreciated.

    • Jason Brownlee December 4, 2017 at 7:43 am #

      Sorry, I don’t know about your binary file.

      Perhaps after you load your data, you can convert it to a numpy array so that you can provide it to a neural net?

      • Eliah December 4, 2017 at 9:28 am #

        Thanks for the tip, I’ll try it.

  172. Wafaa December 7, 2017 at 4:59 pm #

    Thank you very very much for all your great tutorials.

    If I wanted to add batch layer after the input layer, how should I do it?

    Cuz I applied this tutorial on a different dataset and features and I think I need normalization or standardization and I want to do it the easiest way.

    Thank you,

    • Jason Brownlee December 8, 2017 at 5:35 am #

      I recommend preparing the data prior to fitting the model.

  173. zaheer December 9, 2017 at 3:03 am #

    thanks for sharing such nice tutorials, it helped me alot. i want to print the confusion matrix from the above example. and one more question.
    if i have
    20-input variable
    1- class label (binary)
    and 400 instances
    how i would know , setting up the dense layer parameter in the first layer and hidden layer and output layer. like above example you have placed. 12,8,1

    • Jason Brownlee December 9, 2017 at 5:44 am #

      I recommend trial and error to configure the number of neurons in the hidden layer to see what works best for your specific problem.

  174. zaheer December 9, 2017 at 3:29 am #

    C:\Users\zaheer\AppData\Local\Programs\Python\Python36\python.exe C:/Users/zaheer/PycharmProjects/PythonBegin/
    Using TensorFlow backend.
    Traceback (most recent call last):
    File “C:/Users/zaheer/PycharmProjects/PythonBegin/”, line 28, in , y_train , epochs=100, batch_size=100)
    File “C:\Users\zaheer\AppData\Local\Programs\Python\Python36\lib\site-packages\keras\”, line 960, in fit
    File “C:\Users\zaheer\AppData\Local\Programs\Python\Python36\lib\site-packages\keras\engine\”, line 1574, in fit
    File “C:\Users\zaheer\AppData\Local\Programs\Python\Python36\lib\site-packages\keras\engine\”, line 1407, in _standardize_user_data
    File “C:\Users\zaheer\AppData\Local\Programs\Python\Python36\lib\site-packages\keras\engine\”, line 153, in _standardize_input_data
    ValueError: Error when checking input: expected dense_1_input to have shape (None, 20) but got array with shape (362, 1)

  175. Anam Zahra December 10, 2017 at 7:40 pm #

    Dear Jason! Great job a very simple guide.
    I am trying to run the exact code but there is an eror

    ValueError: Error when checking target: expected dense_3 to have shape (None, 1) but got array with shape (768, 8)

    How can I resolve.

    I have windows 10 and spyder.

    • Jason Brownlee December 11, 2017 at 5:24 am #

      Sorry to hear that, perhaps confirm that you have the latest version of Numpy and Keras installed?

  176. nazek hassouneh December 11, 2017 at 7:33 am #

    after run this code , i will calculate the accuracy , how i did , i
    i want to split the data set into test data , training data
    and evaluate the model and calculate the accuracy
    thank dr.

  177. Suchith December 21, 2017 at 2:35 pm #

    In the model how many hidden layers are there ?

    • Jason Brownlee December 21, 2017 at 3:35 pm #

      There are 2 hidden layers, 1 input layer and 1 output layer.

  178. Amare Mahtesenu December 22, 2017 at 9:55 am #

    hi there. this blog is very awesome like the Adrian’s pyimagesearch blog. I have one question and that is do you have or will you have a tutorial on keras frame work with SSD or Yolo architechtures?

    • Jason Brownlee December 22, 2017 at 4:16 pm #

      Thanks for the suggestion, I hope to cover them in the future.

  179. Kyujin Chae January 8, 2018 at 2:22 pm #

    Thanks for your awesome article.
    I am really enjoying
    ‘Machine Learning Mastery’!!

  180. Luis Galdo January 9, 2018 at 8:41 am #

    Hello Jason!

    This is an awesome article!
    I am writing a report for a subject in university and I have used your code during my implementation, would it be possible to cite this post in bibtex?

    Thank you!

  181. Nikhil Gupta January 25, 2018 at 8:05 pm #

    My question is regarding predict. I used to get decimals in the prediction array. Suddenly, I started seeing only Integers (0 or 1) in the run. Any idea what could be causing the change?

    predictions = model.predict(X2)

    array([[ 0.],
    [ 0.],
    [ 0.],
    [ 0.],
    [ 0.],
    [ 0.]], dtype=float32)

    • Jason Brownlee January 26, 2018 at 5:39 am #

      Perhaps check the activation function on the output layer?

      • Nikhil Gupta January 28, 2018 at 3:30 am #

        # create model. Fully connected layers are defined using the Dense class
        model = Sequential()
        model.add(Dense(12, input_dim=len(x_columns), activation=’relu’)) #12 neurons, 8 inputs
        model.add(Dense(8, activation=’relu’)) #Hidden layer with 8 neurons
        model.add(Dense(1, activation=’sigmoid’)) #1 output layer. Sigmoid give 0/1

  182. joe January 27, 2018 at 1:25 am #

    ================== RESTART: /Users/apple/Documents/ ==================
    Using TensorFlow backend.

    Traceback (most recent call last):
    File “/Users/apple/Documents/”, line 20, in
    model.compile(loss=’binary_crossentropy’, optimizer=’adam’, metrics=[‘accuracy’])
    File “/Library/Frameworks/Python.framework/Versions/2.7/lib/python2.7/site-packages/keras/”, line 826, in compile
    File “/Library/Frameworks/Python.framework/Versions/2.7/lib/python2.7/site-packages/keras/engine/”, line 827, in compile
    sample_weight, mask)
    File “/Library/Frameworks/Python.framework/Versions/2.7/lib/python2.7/site-packages/keras/engine/”, line 426, in weighted
    score_array = fn(y_true, y_pred)
    File “/Library/Frameworks/Python.framework/Versions/2.7/lib/python2.7/site-packages/keras/”, line 77, in binary_crossentropy
    return K.mean(K.binary_crossentropy(y_true, y_pred), axis=-1)
    File “/Library/Frameworks/Python.framework/Versions/2.7/lib/python2.7/site-packages/keras/backend/”, line 3069, in binary_crossentropy
    TypeError: sigmoid_cross_entropy_with_logits() got an unexpected keyword argument ‘labels’

    • Jason Brownlee January 27, 2018 at 5:58 am #

      I have not seem this error, sorry. Perhaps try posting to stack overflow?

  183. Atefeh January 27, 2018 at 4:04 pm #

    Hello Mr.Janson
    After installing Anaconda and deep learning libraries, I read your Free mini-course and I tried to write the code about the handwritten digit recognition.
    I wrote the codes in jupyter notebook, am I right?
    if not where should I write the codes ?
    and if I want to use another dataset (my own data set) how can I use in the code?
    and how can I see the result, for example the accuracy percentage?
    I am really sorry for my simple questions! I have written a lot of code in “Matlab” but I am really a beginner in Python and Anaconda, my teacher force me to use Python and keras for my project.

    thank you very much for your help

    • Jason Brownlee January 28, 2018 at 8:22 am #

      A notebook is fine.

      You can write code in a Python script and then run the script directly.

  184. Atefeh January 28, 2018 at 12:01 am #

    Hello Mr.Janson again
    I wrote the code below from your Free mini course for hand written digit recognition, but after running I faced the syntaxerror:

    from keras.datasets import mnist

    (X_train, y_train), (X_test, y_test) = mnist.load_data()

    X_train = X_train.reshape(X_train.shape[0], 1, 28, 28)
    X_test = X_test.reshape(X_test.shape[0], 1, 28, 28)

    from keras.utils import np_utils

    y_train = np_utils.to_categorical(y_train)
    y_test = np_utils.to_categorical(y_test)

    model = Sequential()
    model.add(Conv2D(32, (3, 3), padding=’valid’, input_shape=(1, 28, 28),
    model.add(MaxPooling2D(pool_size=(2, 2)))
    model.add(Dense(128, activation=’relu’))
    model.add(Dense(num_classes, activation=’softmax’))
    model.compile(loss=’categorical_crossentropy’, optimizer=’adam’, metrics=[‘accuracy’])

    File “”, line 2
    2 model.add(Conv2D(32, (3, 3), padding=’valid’, input_shape=(1, 28, 28),
    SyntaxError: invalid syntax

    would you please help me?!

    thanks a lot

    • Jason Brownlee January 28, 2018 at 8:25 am #


      should be:

  185. Lila January 29, 2018 at 8:04 am #

    Thank you for the awsome blog and explanations. I have just a question: How can we get predicted values by the model. . Many thanks

    • Jason Brownlee January 29, 2018 at 8:21 am #

      As follows:

      • Lila January 30, 2018 at 1:22 am #

        Thank you for your prompt answer. I am trying to learn how keras models work and I used. I trained the model like this:

        model.compile(loss=’mean_squared_error’, optimizer=’sgd’, metrics=[‘MSE’])

        As output I have those lines

        Epoch 10000/10000

        10/200 [>………………………..] – ETA: 0s – loss: 0.2489 – mean_squared_error: 0.2489
        200/200 [==============================] – 0s 56us/step – loss: 0.2652 – mean_squared_error: 0.2652

        and my question what the difference between the two lines (MSE values)

        • Jason Brownlee January 30, 2018 at 9:53 am #

          They should be the same thing. One may be calculated at the end of each batch, and one at the end of each epoch.

  186. Atefeh January 30, 2018 at 4:28 am #


    after running again it show an error:

    NameError Traceback (most recent call last)
    in ()
    —-> 1 model = Sequential()
    2 model.add(Conv2D(32, (3, 3), padding=’valid’, input_shape=(1, 28, 28), activation=’relu’))
    3 model.add(MaxPooling2D(pool_size=(2, 2)))
    4 model.add(Flatten())
    5 model.add(Dense(128, activation=’relu’))

    NameError: name ‘Sequential’ is not defined

    • Jason Brownlee January 30, 2018 at 9:55 am #

      You are missing the imports. Ensure you copy all code from the complete example at the end.

  187. Atefeh January 31, 2018 at 1:02 am #

    from keras.datasets import mnist

    (X_train, y_train), (X_test, y_test) = mnist.load_data()
    X_train = X_train.reshape(X_train.shape[0], 1, 28, 28)
    X_test = X_test.reshape(X_test.shape[0], 1, 28, 28)
    from keras.utils import np_utils

    y_train = np_utils.to_categorical(y_train)
    y_test = np_utils.to_categorical(y_test)

    model = Sequential()
    2 model.add(Conv2D(32, (3, 3), padding=’valid’, input_shape=(1, 28, 28), activation=’relu’))
    3 model.add(MaxPooling2D(pool_size=(2, 2)))
    4 model.add(Flatten())
    5 model.add(Dense(128, activation=’relu’))
    6 model.add(Dense(num_classes, activation=’softmax’))
    7 model.compile(loss=’categorical_crossentropy’, optimizer=’adam’, metrics=[‘accuracy’])

  188. Atefeh February 2, 2018 at 5:01 am #

    please tell me how can I find out that tensorflow and keras are correctly installed on my system.
    maybe the problem is that, because no code runs in my jupyter. and no “import” acts well(for example import pandas)
    thank you

  189. Dan February 3, 2018 at 12:29 am #

    Hi. I’m totally new to machine learning and I’m trying to wrap my head around it.
    I have a problem I can’t quite solve yet. And don’t know where to start actually.
    I have a dictionary with a few key:value pairs. The key is a random 4 digit number from 0000 to 9999. And the value for each key is set as follows: if a digit in a number is either 0, 6 or 9 then its weight is 1, if a digit is 8 then it’s weight is 2, any other digit has a weight of 0. All the weights are summarised then and here you have the value for the key. (example: { ‘0000’: 4, ‘1234’: 0, ‘1692’: 2, ‘8800’: 6} – and so on).

    Now I’m trying to build a model that will predict the correct value of a given key. (i.e if I give it 2222 the answer is 0, if I give it 9011 – it’s 2). What I did first is created a CSV file with 5 columns, first four is a split (by a single digit) key from my dictionary, and the fifth column is the value for each key. Next I created a dataset and defined a model (like this tutorial but with input_dim=4). Now when I train the model the accuracy won’t go higher then ~30%. Also your model is based on binary output, whereas mine should have an integer from 0 to 8. Where do I go from here?

    Thank you for all your effort in advance! 🙂

  190. Alex February 5, 2018 at 5:22 am #

    There is one thing I just dont get.

    An example of row data is 6,148,72,35,0,33.6,0.627,50,1

    I guess the number at the end is if the person has diabetes (1) or does not (0) , but what I dont understand is how I know the ‘prediction’is about that 0 or 1, tehere are a lot of other variables in the data, and I dont see ‘diabetes’ being a label for any of that.

    So, how do I know or how do I set wich variable (number) I want to predict?

    • Jason Brownlee February 5, 2018 at 7:49 am #

      You interpret the prediction in your application or usage.

      The model does not care what the inputs and outputs are, it does the best it can. It does not intrinsically care about diabetes.

  191. blaisexen February 6, 2018 at 9:14 am #

    @Jason Brownlee, Master of Keras Python.

    I’m developing a face recognition testing, I successfully used Rprop, it was good for static images or face pictures, I also have test svm results.

    What do you think in your experienced that Keras is better or powerful than Rprop?

    because I was also thinking to used Keras(1:1) for final result of Rprop(1:many).

    or which do you think is better system?

    thanks in advance for the advices.

    I also heard one of the leader of commercial face recognizers uses PNN(uses libopenblas), so I really doubt which one to choose for my final thesis and application.

    • Jason Brownlee February 6, 2018 at 9:29 am #

      What do you mean by rprop? I believe it is just an optimization algorithm, whereas Keras is a deep learning library.

      • blaisexen February 17, 2018 at 10:46 am #

        Ok, I think I understand you.

        I used Accord.Net
        Rprop testing was good
        MLR testing was good
        SVM testing was good
        RBM testing was good

        I used classification for face images
        They are only good for static face pictures 100×100

        but if I used another picture from them,
        these 4 testing I have failed.

        Do you think if I used Keras in image face recognition will have a good result or good prediction?

        because if Keras will have a good result then I’ll have to used cesarsouza keras c#

        thanks for the reply.

  192. CHIRANJEEVI February 8, 2018 at 8:52 pm #

    What is the difference between the accuracy we get when we fit the model and the accuracy_score() of sklearn.metrics , what they mean exactly ?

    • Jason Brownlee February 9, 2018 at 9:05 am #

      Accuracy is a summary of the number of predictions that were made correctly out of all predictions that were made.

      It is used as an estimate of model skill on new out of sample data.

  193. Shinan February 8, 2018 at 9:09 pm #

    is weather forecasting can done using RNN?

    • Jason Brownlee February 9, 2018 at 9:06 am #

      No. Weather forecasting is done with ensembles of physics simulations on very large computers.

  194. CHIRANJEEVI February 9, 2018 at 3:56 pm #

    we haven’t predicting anyting during the fit (its just a training , like mapping F(x)=Y)
    but still getting acc , what is this acc?

    Epoch 1/150
    768/768 [==============================] – 1s 1ms/step – loss: 0.6771 – acc: 0.6510

    Thank you in advance

    • Jason Brownlee February 10, 2018 at 8:50 am #

      Predictions are made as part of back propagating error.

  195. lcy1031 February 12, 2018 at 1:00 pm #

    Hi Jason,

    Many thanks to you for a great tutorial. I have couple questions to you as followings.
    1). How can I get the score of Prediction?
    2). How can I output the result of predict run to a file in which the output is listed by vertical?

    I see you everywhere to answer questions and help people. Your time and patience were greatly appreciated!


    • Jason Brownlee February 12, 2018 at 2:50 pm #

      You can make predictions with a model as follows:

      yhat = model.predict(X)

      You can then save the numpy array result to file.