Open JiahaoYao opened 1 year ago
One of the shortcomings for this approach is for the different splits, the lengths are not equally divided, for example Boron example, the electrons are (3, 2). In this way, it is quite hard to share the self attention layers, because the shape is different. The ideal case should go back to different individual self attention layers.
I wonder whether it would be improving if we change the transformer to be individual for different split?
The code is here: https://github.com/jeffminlin/vmcnet/blob/master/vmcnet/models/equivariance.py#L170