This is my equivariant graph neural network library. I have one goal only:
- to make the implementation so simple that I can come back to it in a few months and understand how it works
This repo was heavily inspired by code in e3nn, e3nn-jax, e3nn.c, and e3x
Equivariant Graph Neural Network libraries are pretty complex and not well-explained. I'm doing this so I can learn the math and the minute details.
pip install -e .
- Read this! https://docs.google.com/presentation/d/1ZNM52MDDc183y5j4AIX27NjePoJP1qLnAhYsyKaBzqI/edit?usp=sharing
- Read this! https://docs.google.com/presentation/d/1tuhAtmkWthONETgRxBx1pVVXcoYxj4ooa6HvpOHFsVw/edit?usp=sharing
- make sure you're using cartesian order in all places (when retrieving the spherical harmonics coefficients, the clebsch-gordan coefficients, and setting the coefficients that the tensor product outputs)
- When getting the clebsch gordan coefficients, check the shape of the matrix you're reading it from. Make sure you're only reading the coefficients for degrees l1,l2,l3 NOT all the degrees up to l1+l2+l3 (which is a larger matrix).
- make sure you normalize the vectors before you calculate the spherical harmonics coefficients to get the irreps
- not normalizing the resulting tensor by sqrt(1/num_paths) when we aggregate irreps of the same id.
- See the e3nn paper when they talk about noramlization in the tensor product.
- I only do this operation in the linear layer since OUR tensor product is a REAL tensor product (I output all of the irreps, even if it's of higher l than the input).
- our linear layer does the actual logic of consolidating weights for each irrep of the same id (so we need to normalize there)
- only performing equivariance tests for l=0 and l=1. The tensor product passes for that, but for higher l, it fails.
- I made custom message passing functions since we are not taking advantage of nice tensors (with consistent shapes during message passing)
- It also showcases the simplicity of message passing and how we can maintain equivariance at the same time
- There is no batch dimension into the model. We pass one graph at a time into the model to train.
- LinearLayer tests
- Add an equivariance test for 3D outputs
- support adding scalar features as features
- simplify files. put o3 utils in an o3 folder
- a "debugger" to determine where we're losing precision. are we losing it cause we're throwing away higher order irreps (larger ls)?