A collection of tools to analyze and visualize deep neural networks.

The original goal of the toolbox was visualizing networks that solve image classifications tasks. It was motivated by the paper:

• Jason Yosinski, Jeff Clune, Anh Nguyen, Thomas Fuchs, and Hod Lipson. Understanding neural networks through deep visualization Presented at the Deep Learning Workshop, International Conference on Machine Learning (ICML), 2015.

The main design goals are (not fully achieved yet):

• modularity: new tools can be added, building on common core functionality
• framework agnosticity: the toolbox should be able to use different neural network frameworks like TensorFlow, Torch, Caffe, ...
• clear API: the toolbox should define an interface allowing to use tools from within other programs
• comprehensive CLI: allowing to run all tools from the command line or scripts
• easy to use GUI: providing intuitive access to the tools

This is still work in progress. Below we document some aspects already realized.

Example of a network visualization that classifies different shapes

The main window consists of activation maps for selected layer, the input and the structure of a network where the layers of interest can be chosen

Installation process is just installing packages listed in the requirements.txt

• python virtualenv:

  1. Install your virtualenv following The Hitchhiker’s Guide to Python, check that python version >=3.8 is set as primary interpreter

  2. After activating the enviroment, run the command to install necessary libraries:

     $pip install -r requirements.txt

Install PyTorch.

$ conda install pytorch torchvision cuda80 -c soumith

Then run the PyTorch MNIST example, tpye

$ python main.py --framework=torch

Run $ pytest to execute all tests.

In order to build the Sphinx documentation, cd into the doc directory and run make html. Then open build/html/index.html. Note: Sphinx executes all found modules, which can exhibit side-effects. For instance, loading TensorFlow takes a lot of time, and so building the documentation does as well.

When adding a new python file, run sphinx-apidoc <folder> -o doc/source to generate an .rst file indexing the modules. When the apidoc tool finds modules it has already created indexes for, it will complain that those already exist. It may be neccessary to delete them and regenerate them if you add a submodule, I'm not currently sure.