sanjeev309/cnn_heatmap

Visualising hidden layers of a Convolutional Neural Network trained on MNIST dataset

Visualising Hidden Layers in a convolutional neural network

Visualisation of how an input gets filtered for inference in a trained CNN network

Getting Started

Clone the repository into the system

git clone https://github.com/sanjeev309/cnn_heatmap.git

Launch Jupyter Notebook and run MNIST HeatMap.ipynb

Prerequisites

Install :

• Python 3.6
• tensorflow
• matplotlib
• numpy

pip3 install tensorflow==1.8.0 matplotlib numpy

Description

A CNN network is trained on the classic MNIST dataset.

If you do not already have the dataset, the script will download it for you.

Post training, a single input image is fed again into a trained CNN network. The weights of the network post "activation" by the convoluted image is fetched from the graph and visualised in a binary image format. is done by the following function:

getActivations(layer,stimuli,Heatmap=False)

where

• layer : name of the layer to visualise
• stimuli: input sample image from the mnist dataset
• Heatmap: boolean to show heatmap

Results

When input of 3 is given, the first and second hidden layer's output are:

Further Reading

visit Website : AlphaDataOne for more details.