Brief Reminder: We are building Vizstack: a suite of programming tools that visualize code structure & execution — for better understanding & simpler debugging in existing languages. Consider the last time you encountered an unfamilar codebase or had to debug a nasty algorithm. We believe that this code complexity can be tamed with better developer tools. We believe the time is right to reinvent the programming experience, so that developers can build bigger, faster, safer, and more intricate software.
To learn more, visit the home page.
A crucial part of our visualization library is being able to visualize graphs. For various reasons outlined here, we've decided to develop our own graph layout library called nodal.
We're continuing to work on our graph component for a neural network interpretability tool. This a great use case for many components of the Vizstack ecosystem: core, vz-logger, and nodal.
Finally, we've implemented edge routing so that graphs actually look reasonably uncluttered now!
This week, we focused on:
- Added edge routing
- Builing more complex pytorch models
- Improved python API
- Finalized compound force model
- Fixing ports bug
- Fixing overlap bug
- Show all Tensor operations in compgraph
repo: https://github.com/vizstack/vz-pytorch
Edge routing refers to the process of transforming edges in a graph after the layout process has completed so that they avoid obstacles and generally make the graph look cleaner. In particular, we chose to implement orthogonal edge routing, since having edges made up of only horizontal or vertical segments makes graphs feel more inviting to explore, which is especially critical in large graphs.
Before:
After:
As this Georgia Tech prof can attest (https://twitter.com/mark_riedl/status/1067625823333687296), the available computation graphs in things like Tensorboard are terrible! We're currently focused on making computation graphs better — more like what you'd find in academic papers — so that you can code neural networks while visualizing it.
This week, we focused on further improving the force model so that graphs lay out in a more natural way.
Before:
After:
