VizConvnets

Visualizing Channels of 2D Convolutional Layers through Tensorflow.js

Personal Project

Check out the Live Demo: https://mishig25.github.io/vizconvnets/

Description:

After AlexNet won ImageNet 2012, popularity and usage of convnets have increased exponentially. Visualizing channels/filters of conv layers turned out to be a powerful tool for analyzing how Convolutional Neural Networks work. Zeiler and Fergus were one of the first people to visualize convnets throughly and went on to win ImageNet 2013. Afterwards, there was a plethora of papers and demos about visualizing convnets, including the popular one by Yosinski.

This project is a continuation of the convnet visualizing trend. By using Tensorflow.js and MobileNet, an efficient CNN architecture, the project visualizes sample channels/filters from MobileNet and does so through web browser only.

Contents of the repository:

• Model Activation model is created through Keras Functionall API in Jupyter Notebooks.
• Frontend Using Tensorflowjs and HTML5 Canvas to create a convnet visualizations in web-browser environemnt.

Development:

git clone https://github.com/mishig25/vizconvnets.git
cd ./vizconvnets
cd frontend
yarn
yarn watch

License

MIT