erogol/deepviz

DeepViz is an interactive visualization tool that helps you understand how deep convolutional neural networks learn to recognize images. Instead of treating a trained model as a black box, it lets you see what the network is doing at each stage of learning, watching how it evolves from random initialization through training checkpoints, and inspecting what features it learns to detect in images. The core idea is that neural networks trained for image classification go through many iterations, and at each step they get incrementally better at the task. DeepViz captures snapshots of the model at different training stages and computes statistics about how the model performs on a dataset of images. You then open a web interface (running locally on your computer) where you can browse these snapshots, visualize what the network "sees," and understand how its understanding improves over time. To use it, you need three things: a set of training images (the tool comes with instructions for using CIFAR-10, a standard dataset of small photographs), a series of saved model checkpoints (snapshots of the network at different points in training), and pre-computed statistics about the model's behavior. The README provides sample datasets you can download to try it out immediately. You run a Python script that starts a web server, then open your browser to explore the visualizations interactively. This tool would be useful for researchers studying how neural networks learn, for students trying to understand what happens inside these models, or for anyone troubleshooting a model's performance, you can literally watch where it starts making mistakes or struggling with certain types of images. The visualization aspect is particularly valuable because neural networks operate on millions of numbers that are nearly impossible to understand as raw data, seeing them visualized as learned features and predictions makes the learning process tangible.