tinygraph: PyTorch Geometric for tinygrad
PyG (PyTorch Geometric) is a library built upon PyTorch to easily write and train Graph Neural Networks (GNNs) for a wide range of applications related to structured data.
GitHub Repository: github.com/squarra/tinygraph
In order to learn about GNNs and tinygrad I decided t built the equivalent of what PyG is for pytorch for tinygrad. This was a lot of fun. Unfortunately I lost a lot of progress which was sitting on my server, when I accidently formatted the partition. Here are the main things I implemented:
- Compatibility layer for working with PyG datasets
DataandHeteroDatalike you would find them in PyG, in a lot less lines of code- The basic
GCNConvandSAGEConvlayers. On the work I lost, I had a working (but very slowHGTConv) - Some basic examples using the beauty of tinygrad
My biggest takeaway is how much better it felt working with tinygrad compared to pytorch. The API is feels much more correct and I hope it finds some more adoption with increasing speed.