Cascade: A visual machine learning framework (mockup)

• Visually arrange a computational graph in the web-based frontend and perform computations on the backend (similar to Jupyter)
• A demo version lets you try it out without the backend - using pyodide to emulate python in javascript
• Arbitrary python code can be executed
• Many specific features are possible: Dynamic resource allocation, a history of training runs, smart hyperparameter management and more...

Purpose

In the current state this demo is not capable of making workflows more productive. The aim is to showcase a different way of setting up machine learning pipelines. In an ideal world, tasks such as managing devices, tracking hyperparameter choices and model backups could be handled by a framework instead of being done by every data scientist on their own. In my opinion, a visual format would help a lot in making machine learning more accessible, easier to communicate and potentially avoid many mistakes.

Demo Tutorial

There is a solely web-based demo that showcases how to train a simple GAN to imitate the standard normal distribution in Cascade. Find it here and follow these steps:

More Details on how it works

• Ìnitialize executes all code in the boxes and saves the results to variables.
• Executing an output or plot box will trigger code to be executed along the computational graph
• Code Boxes without an input will return their value directly
• Code Boxes with inputs are treated as functions and will process the output from connected code boxes
• Plot boxes expect a list of plots, which are two-element lists of lists of data points (this might change in the future)
• Normal output boxes always return the result as string
• In the demo version, t is bintorch and np is numpy, other imports are currently not available but can be added in server.py in the offline version
• Open the console to see (python) errors
• Execute blocks in the routine pane must be linked to output boxes by clicking first on the routine execute and then on the output box in the computational graph. Both should be highlighted in red. Click elsewhere to finish the link.

Technical Details

• the frontend uses jsplumb for the drag-and-drop connectors
• SortableJS is used for the drag-and-drop routines
• Pyodide is used for Python emulation in the web-demo
• The backend and frontend manage a shared state in json format. state.json therefore acts as a save file (for now)