Model Inspector
A tool for watching what's happening inside a neural network as it trains.
Loss curves don’t tell you much about what a model is actually doing, especially with many reinforcement learning algorithms. Model Inspector logs a PyTorch model’s internals during training and replays them in a web UI.
- Reads the network’s architecture automatically using PyTorch’s ONNX export
- Draws the network as a graph
- Heatmaps of parameters, and of how far they’ve moved from their starting values, animated over time
- Histograms and mean/variance of parameters and gradients over the course of training
The UI runs entirely from static files. Try the MNIST demo.