Grada is an interactive tool for observing real-time changes while training a multilayer perceptron, built entirely from scratch without libraries like TensorFlow, PyTorch, or NumPy. An extension of Karpathy's micrograd, Grada originally used a scalar-based engine, but has now been reengineered with a custom tensor-based engine. The scalar version is available on the scalar-value branch.
With a simple drag-and-drop interface, you can easily construct neural networks and watch how training affects parameters and outputs in real time. Grada also features a component for handwritten digit recognition, enabling you to test your model interactively by drawing digits and visualizing predictions.
The website welcomes you with a quick manual that guides you through using the app.
Available here
https://www.youtube.com/watch?v=VMj-3S1tku0
https://github.com/karpathy/micrograd
https://github.com/ixartz/handwritten-digit-recognition-tensorflowjs
https://cs.stanford.edu/people/karpathy/convnetjs/demo/mnist.html
https://people.ece.ubc.ca/bradq/ELEC502Slides/ELEC502-Part5VectorizedBackpropagation.pdf (***)
Contributions, issues and feature requests are welcome.