Is your feature request related to a problem? Please describe.
The models SR-GNN, GC-SAN contain NumPy operations in their forward function which requires repeated data transfers between CPU and GPU during training and inference. This is very inefficient.
Describe the solution you'd like
Replace the Numpy operations with PyTorch code.
@bkersbergen Thank you for your suggestions!
The numpy operations in forward are mainly from function "_get_slice". We will rewrite the function and update the model.
Is your feature request related to a problem? Please describe. The models SR-GNN, GC-SAN contain NumPy operations in their
forward
function which requires repeated data transfers between CPU and GPU during training and inference. This is very inefficient.Describe the solution you'd like Replace the Numpy operations with PyTorch code.
Describe alternatives you've considered
Additional context