# How to Visualize a Deep Learning Neural Network Model in Keras

Last Updated on September 11, 2019

The Keras Python deep learning library provides tools to visualize and better understand your neural network models.

In this tutorial, you will discover exactly how to summarize and visualize your deep learning models in Keras.

After completing this tutorial, you will know:

• How to create a textual summary of your deep learning model.
• How to create a graph plot of your deep learning model.
• Best practice tips when developing deep learning models in 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.

How to Visualize a Deep Learning Neural Network Model in Keras
Photo by Ed Dunens, some rights reserved.

## Tutorial Overview

This tutorial is divided into 4 parts; they are:

1. Example Model
2. Summarize Model
3. Visualize Model
4. Best Practice Tips

## Example Model

We can start off by defining a simple multilayer Perceptron model in Keras that we can use as the subject for summarization and visualization.

The model we will define has one input variable, a hidden layer with two neurons, and an output layer with one binary output.

For example:

If you are new to Keras or deep learning, see this step-by-step Keras tutorial.

The code listing for this network is provided below.

## Summarize Model

Keras provides a way to summarize a model.

The summary is textual and includes information about:

• The layers and their order in the model.
• The output shape of each layer.
• The number of parameters (weights) in each layer.
• The total number of parameters (weights) in the model.

The summary can be created by calling the summary() function on the model that returns a string that in turn can be printed.

Below is the updated example that prints a summary of the created model.

Running this example prints the following table.

We can clearly see the output shape and number of weights in each layer.

## Visualize Model

The summary is useful for simple models, but can be confusing for models that have multiple inputs or outputs.

Keras also provides a function to create a plot of the network neural network graph that can make more complex models easier to understand.

The plot_model() function in Keras will create a plot of your network. This function takes a few useful arguments:

• model: (required) The model that you wish to plot.
• to_file: (required) The name of the file to which to save the plot.
• show_shapes: (optional, defaults to False) Whether or not to show the output shapes of each layer.
• show_layer_names: (optional, defaults to True) Whether or not to show the name for each layer.

Below is the updated example that plots the created model.

Note, the example assumes that you have the graphviz graph library and the Python interface installed.

Running the example creates the file model_plot.png with a plot of the created model.

Plot of Neural Network Model Graph

## Best Practice Tips

I generally recommend to always create a summary and a plot of your neural network model in Keras.

I recommend this for a few reasons:

• Confirm layer order. It is easy to add layers in the wrong order with the sequential API or to connect them together incorrectly with the functional API. The graph plot can help you confirm that the model is connected the way you intended.
• Confirm the output shape of each layer. It is common to have problems when defining the shape of input data for complex networks like convolutional and recurrent neural networks. The summary and plot can help you confirm the input shape to the network is as you intended.
• Confirm parameters. Some network configurations can use far fewer parameters, such as the use of a TimeDistributed wrapped Dense layer in an Encoder-Decoder recurrent neural network. Reviewing the summary can help spot cases of using far more parameters than expected.

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

## Summary

In this tutorial, you discovered how to summarize and visualize your deep learning models in Keras.

Specifically, you learned:

• How to create a textual summary of your deep learning model.
• How to create a graph plot of your deep learning model.
• Best practice tips when developing deep learning models in Keras.

Do you have any questions?

## Develop Deep Learning Projects with Python!

#### What If You Could Develop A Network in Minutes

...with just a few lines of Python

Discover how in my new Ebook:
Deep Learning With Python

It covers end-to-end projects on topics like:
Multilayer PerceptronsConvolutional Nets and Recurrent Neural Nets, and more...

### 89 Responses to How to Visualize a Deep Learning Neural Network Model in Keras

1. Mishti December 13, 2017 at 6:05 pm #

Running plot_model leaves an error.
It cant import pydot.
I did install it and try.
No improvement at all.

• Jason Brownlee December 14, 2017 at 5:34 am #

What is the error?

• Mishti December 14, 2017 at 9:54 pm #

I get this.
attribute error: ‘module’ object has no attribute ‘find_graphviz’

• Jason Brownlee December 15, 2017 at 5:33 am #

You may need to install two packages: pydot and pygraphviz.

• Shantanu Oak January 2, 2018 at 9:53 pm #

if you are using conda environment then may be this will work.

!conda install –yes graphviz
!conda install –yes pydotplus

• Mitchell January 4, 2018 at 5:01 pm #

I have the same problem.
It shows “ImportError: Failed to import pydot. You must install pydot and graphviz for pydotprint to work.” even I had installed graphviz and pydot or pydotplus.

• Jason Brownlee January 5, 2018 at 5:19 am #

You must install pydot and pygraphviz.

• Channing October 13, 2018 at 12:02 am #

You can also try add graphviz folder to system path. It works for me.

• pajeet December 18, 2019 at 3:58 am #

If you’re having an error, it’s 99% likely that it’s because you didn’t put the /graphviz/bin/ filepath to your PATH. pip install graphviz is not sufficient. You need to add its files to PATH

2. Murtaza December 17, 2017 at 11:00 am #

Hi Jason , you could also include a tutorial for tensorboard in which each time a model is run we can log it using callback function and display all runs on tensorboard

3. vinci December 24, 2017 at 3:02 pm #

thank you Jason Brownlee for your help……god bless you

4. Stefano December 27, 2017 at 1:42 am #

All the prints do not contain the activation function, I think important in defining a layer!…

5. Daniel January 9, 2018 at 7:05 pm #

check out this project for a better visual result:
https://github.com/yu4u/convnet-drawer

6. mistermoper February 5, 2018 at 3:32 am #

What does a dense layer respect other kind of layers?
Another question, un all the neurons of the same layer It Will be used the same activation función?

• Jason Brownlee February 5, 2018 at 7:47 am #

All of the neurons in a layer do use the same activation function.

7. Junho Lee February 20, 2018 at 2:34 am #

HI, I got this error : “raise ImportError(‘Failed to import pydot. You must install pydot’
ImportError: Failed to import pydot. You must install pydot and graphviz for pydotprint to work.”

I have installed both pydot and graphviz via “sudo pip install **”, but still not solving the problem. Can I have a comment?

Thank you

• Jason Brownlee February 21, 2018 at 6:29 am #

I recommend installing pygraphviz instead of graphviz.

8. Jacob J February 23, 2018 at 3:57 pm #

I’m currently trying to get comfortable with Keras. I think I need to gather examples of importing things like image data. Also, I only really understand basic neural networks so far, I’ll have to study further on how to make convolutional networks with Keras. Finally, I’m not sure I understand how a network can work with batches of images at the same time.

I plan to use Keras with this repo: github.com/diginessforever/machinelearning -> it already has a nifty webscraper and some of my noobish neural network code.

• Jason Brownlee February 24, 2018 at 9:08 am #

Cool, thanks for sharing. Good luck!

9. Pawan March 9, 2018 at 8:06 pm #

Can we visualize the model in tensorboard

• Jason Brownlee March 10, 2018 at 6:24 am #

I believe so. I do not have a tutorial, sorry.

10. Gautam March 25, 2018 at 5:20 am #

Jason, really didn’t get it working. I have installed pydot, pygraphviz and gpraphviz using both conda and python – m pip command but still same message. what is the trick. If you can specify command in sequence may help. Thanks in advace.

• Jason Brownlee March 25, 2018 at 6:34 am #

There was no special trick, I just pip-installed those two libraries.

Perhaps try posting/searching for your error on stackoverflow.com?

11. Alexandru March 25, 2018 at 9:48 pm #

Hi Jason, cool tutorial! Could you please expand it to a Model with multiple inputs? Nothing complicated, maybe a Dense layer on input a, a Dense layer on input b, merge them together, map to some classes, done.Thanks in advance.

12. Amit March 26, 2018 at 3:43 am #

For training purpose in Keras, Do we need to explicitly configure to use GPU ? Or it will automatically use the GPU?

• Jason Brownlee March 26, 2018 at 10:03 am #

You must configure the backend (e.g. tensorflow, etc.) to use the GPU. If done, Keras will then use the GPU.

13. Akhtar Munir April 1, 2018 at 8:46 pm #

ImportError: Failed to import pydot. You must install pydot and graphviz for pydotprint to work.

I have already installed pydot and graphviz using conda install pydot and conda install graphivz,

but still showing me this error?

• Jason Brownlee April 2, 2018 at 5:22 am #

Sorry to hear that. Perhaps post the error to stackoverflow?

14. Tomas Mendoza April 3, 2018 at 8:13 pm #

Hello Jason.

Digging in stackoverflow i found out that the following error is actually a bug in Keras and many people have it, it does not get fixed even if you install both libraries:

ImportError: Failed to import pydot. You must install pydot and graphviz for pydotprint to work.

Apparently there is no way to solve this to this day. Is there any other trick to visualize the model? Thanks.

• Jason Brownlee April 4, 2018 at 6:11 am #

No trick. I installed the libraries and wrote/ran the code.

15. Yu-Chia Hsu June 24, 2018 at 8:55 pm #

Hello Jason,

I use your code to picture the same model. And I got a problem that the picture had no input

layers. Instead, it pictures numbers such as 140223915681232. Do you have the same problem

before?

• Jason Brownlee June 25, 2018 at 6:21 am #

I have not seen that, perhaps it is a bug? You could try posting to issues section of the Keras github project?

• Walid October 17, 2018 at 5:21 am #

Me too had the same problem

• Jason Brownlee October 17, 2018 at 6:56 am #

You must update your version of Keras.

16. Vivek June 27, 2018 at 5:07 pm #

Hi, Jason I have used conv3d layers in network as my dataset is 3d. now i want to plot the features of intermidiate layers. what is the process to do that and how can i visualize feature maps in intermidiate layers in 3d dataset. w hich library should i use?

• Jason Brownlee June 28, 2018 at 6:11 am #

Sorry, I don’t have examples of visualizing the feature maps within a CNN.

17. Nathan Glenn July 28, 2018 at 2:51 am #

What do the “None” values mean in the printout?

• Jason Brownlee July 28, 2018 at 6:38 am #

Great question!

It means “unspecified” and that it can vary with the size of the input.

18. Zenon Uchida July 31, 2018 at 8:09 pm #

Thank you for this and all the other deep learning how tos. You are the best!

• Jason Brownlee August 1, 2018 at 7:42 am #

You’re welcome, happy it helped.

19. Franziska August 7, 2018 at 11:06 pm #

My model looks like this :

model = Sequential()

If I run the plot_model function I get a picture like you show above but my very first input is a box with a hudge number of 732677288.
What does this mean?

Is there any possibility to print a dot plot of the network so I can better understand the LSTM layer.

• Jason Brownlee August 8, 2018 at 6:20 am #

I see this too, it might be a fault in the latest version of Keras.

• Lukas August 27, 2018 at 7:39 pm #

Same problem here. Is this a confirmed bug? I couldn’t find any information on th web.

20. MARCOS AUGUSTO GRZECA August 19, 2018 at 11:49 am #

There is a similar function in R Keras?

Thank you

21. Ajit September 14, 2018 at 10:19 am #

Hi Jason,
Thanks for a great post.
One question: my plot_model command did generate a file but it did not display it in Jupyter notebook.
I know there is command (%matplotlib inline) to have matplotlib images displayed in Jupyter. Is there similar command that is needed to display output of plot_model?

Thanks
Ajit

• Jason Brownlee September 14, 2018 at 2:35 pm #

I recommend running code from the command line, it always works!

22. Christopher September 24, 2018 at 12:20 pm #

Hi Jason,
Yet another cool post!
The following will not pring the activation layer.
print (model.summary())

However, the activation layer gets printed if you use add it piecewise like the following:
print (model.summary())

Would you know why?

Regards

• Jason Brownlee September 24, 2018 at 2:11 pm #

Yes, in one case it is part of a layer and in the second it is implemented as a layer.

The models are different, it is a design decision, and in this case they are functionally equivalent. This will not always be the case.

23. Sober October October 31, 2018 at 11:42 pm #

Maybe about time to write a tutorial on Tensorboard, Jason? Keep up the god work!

24. JG December 1, 2018 at 11:58 pm #

nice post Jason !. Thks.

I appreciate this particualr tip regarding summary and model_plot functions l :

“It is common to have problems when defining the shape of input data for complex networks like convolutional and recurrent neural networks. The summary and plot can help you confirm the input shape to the network is as you intended”.

because it is very usual get wrong and confused when matching input-output dimension and using reshaping for training data shape, input shape argument, convolutional specific shape treatments, etc …of the different layers and models

• Jason Brownlee December 2, 2018 at 6:21 am #

Yes, it might be the most common question about neural net configuration (issues with size mismatches)!

25. amin February 9, 2019 at 3:10 pm #

you rock!!!!
have been reading blogs here and there for the last 20 years, but your blog is without a doubt best.

26. gordon June 13, 2019 at 7:50 pm #

Most excellent tutorial!

Is there any way to do this ‘in reverse’? For example, with a classifying CNN with 10 output classes, you can ‘stimulate’ one of the outputs and see how the filters produce it?

• Jason Brownlee June 14, 2019 at 6:41 am #

Not sure I understand, can you elaborate?

• gordon June 14, 2019 at 7:29 pm #

Sorry, I’ll try and be a bit clearer. The visualisations here show the layer activations when you present a CNN with, say, a bird, giving a final classification of ‘robin’. We see what parts of the image various layers are interested in on the way from ‘this bird image’ to ‘robin classification’.

My question was related to the inverse operation. If we’ve got that same CNN, is there any way we can use it (or a modified form of it) to see what filters are generally more or less important for the the ‘robin’ classification versus other ones.

• Jason Brownlee June 15, 2019 at 6:29 am #

This sounds like a significantly harder problem. Off hand, I don’t see a clean way.

Perhaps some clever engineer has devised a way, I’m not across it sorry.

27. Hemanth Kumar September 18, 2019 at 10:44 pm #

ValueError: Error when checking input: expected conv2d_3_input to have shape (966, 1296, 3) but got array with shape (966, 1296, 1)

sir i am facing this error while model.fit is executing

and i have and other doubt is that we can train our model using only one image data with label

• Jason Brownlee September 19, 2019 at 5:58 am #

The error suggests a mismatch between the shape of your data and your model.

You can reshape your data or change the expectations of your model.

28. JW B October 22, 2019 at 1:30 am #

Hi Jason,
Yet another cool post!
Can you make a study on how to visualize the changes of LSTM cell state?

Thanks!

• Jason Brownlee October 22, 2019 at 5:55 am #

Thanks.

Interesting suggestion.

Why are you interested in changes in LSTM state?

29. Sreedevi October 30, 2019 at 6:21 pm #

Hi Jason,
One more great tutorial!
Is it possible to save/capture the optimizer & loss function of the model used?

• Jason Brownlee October 31, 2019 at 5:28 am #

Thanks!

What do you mean exactly?

• Sreedevi October 31, 2019 at 3:00 pm #

As the model architecture (layers) can be saved as an image. But this does not contain the information regarding the optimizer and loss function (used in the compile function). Is there a way to retrieve this information from the model?

30. sage June 19, 2020 at 8:40 pm #

I was looking for a way to visualize models so that I can explain to a student properly.
I remember doing this in torch7 which is by far the easiest way to teach this stuff.
Unfortunately developement on lua stalled

31. yaser June 20, 2020 at 8:39 am #

Great tutorial!

After building the DNN or CNN model, and choosing the Dropout rate, # of Epochs, and Batch size, Number of Hidden Neurons in Each Hidden Layer, How to connect the findings with scientific justification?

In other wording: How can I interpret and discuss the results?

• Jason Brownlee June 21, 2020 at 5:59 am #

Results can only provide empirical evidence, and even then typically relative to other models you have explicitly tested on the same test harness, i.e. relative improvement.

Often this is the best we can do in applied machine learning.

32. MariosGavaletakis September 1, 2020 at 6:20 am #

I have installed both required libraries . I suppose that the code will pop out a window showing the visualized model , right ? If yes , i don’t get that window . Should i add some “import” lines?

By the way , i open manually the png file.

• Jason Brownlee September 1, 2020 at 6:40 am #

I believe the image is saved to file, not popped up.

33. Shashank September 10, 2020 at 5:36 pm #

OSError: pydot failed to call GraphViz.Please install GraphViz (https://www.graphviz.org/) and ensure that its executables are in the \$PATH.

i have already installed GraphViz and pydot

• Jason Brownlee September 11, 2020 at 5:51 am #

Also install pygraphviz and pydot.

34. Rahma October 30, 2020 at 8:50 am #

I don’t found “none” it’s a “?” on her place

35. Amandin April 24, 2021 at 9:33 pm #

AttributeError Traceback (most recent call last)
in
41
42 Model.summary()
—> 43 plot_model(model, to_file=’model_plot.png’,show_shapes=True, show_layer_names=True)

~/anaconda3/lib/python3.7/site-packages/keras/utils/vis_utils.py in plot_model(model, to_file, show_shapes, show_layer_names, rankdir, expand_nested, dpi)
238 “””
239 dot = model_to_dot(model, show_shapes, show_layer_names, rankdir,
–> 240 expand_nested, dpi)
241 _, extension = os.path.splitext(to_file)
242 if not extension:

~/anaconda3/lib/python3.7/site-packages/keras/utils/vis_utils.py in model_to_dot(model, show_shapes, show_layer_names, rankdir, expand_nested, dpi, subgraph)
97 if not model.built:
98 model.build()
—> 99 layers = model._layers
100
101 # Create graph nodes.

AttributeError: ‘XGBRegressor’ object has no attribute ‘_layers’

• Jason Brownlee April 25, 2021 at 5:17 am #

Looks like you are trying to apply the code to an xgboost model, which would be invalid.

36. Bharath May 25, 2021 at 2:33 am #

I faced the same problems as mentioned by others above in using pydot and graphviz.
I could get the above issue to work out using google colab, couldnt exactly get the solution but this works for me.
hope this can be of help for someone facing the same issue.

• Jason Brownlee May 25, 2021 at 6:10 am #

Perhaps just comment out the call to plot_model.