Visualize effects of neural net architectural choices on 2D data
Switch branches/tags
Nothing to show
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Failed to load latest commit information.
pyAIUtils @ d933ea0

Easily visualize the effects of neural network architectural choices on 2D data

The code allows you to easily visualize the prediction surface of any neural network for any 2D data. This was written partially for generating results for one of my blog posts. But in addition I have tried to make it easy to try out various tweaks, either your own or one of the many published on arXiv, for improving training and data efficiency of neural networks.


The following result compares the effect of commonly used activations on learning 4 different decision functions. effect_of_activations



Install all the above dependencies except pyAIUtils. Clone the repository and run the following commands inside the created directory to setup pyAIUtils which is used as submodule.

git submodule init
git submodule update

After running these commands, make sure the pyAIUtils directory is created and has aiutils directory inside it.

Code Overview

See nn_pred_surf/experiments directory for example scripts. Each of these scripts uses the Constants class to store network specifications. An object of this class is passed on to which samples data, creates a tensorflow computation graph, trains the network and produces a decision surface.


Feel free to use the code and any material on the blog as long as you attribute it correctly. For publications you may use the following BibTex citation

  author = {Gupta, Tanmay},
  title = {Visualizing the effects of neural network architectural choices},
  year = {2016},
  howpublished = {}

Have fun visualizing!