Skip to content

Small app for visualizing changes in filters and representation produced by a Convolutional Neural Network during training.

Notifications You must be signed in to change notification settings

vb690/cnn_representation_visualizer

Repository files navigation

Anatomy of a Convolutional Neural Network

License: MIT Open in Streamlit

Motivation

This App aims to visualize changes in filters and representation produced by a Convolutional Artificial Neural Network (LeNet-5) while learning how to classify different clothing categories present in the Fashion-MNIST dataset.

For maximizing the observed differences and reducing memory overhead, the observed training period is represented by the sequence of random batches observed during a single training epoch.

Features

  • Script for trainining-on-batch of LeNet-5 on Fashion-MNIST.
    • Extracting learned filters over training.
    • Extracting learned embeddings oser training.
    • Running AlignedUMAP ovser sequence of extracted embeddings.
  • Streamlit App
    • Visulize changes in learned filters over training.
    • Visulize changes in UMAP reduction of learned embedding over training.

How to use

Open the app clicking on the streamlit badge on top of this README.

License

The MIT License

About

Small app for visualizing changes in filters and representation produced by a Convolutional Neural Network during training.

Resources

Stars

Watchers

Forks

Languages