Unofficial PyTorch implementation of "Involution: Inverting the Inherence of Convolution for Visual Recognition" by Li et al. presented at CVPR 2021.
[abs, pdf, Yannic's Video]
You can install involution_pytorch via pip:
pip install involution_pytorchYou can use the Inv2d layer as you would with any PyTorch layer:
import torch
from involution_pytorch import Inv2d
inv = Inv2d(
channels=16,
kernel_size=3,
stride=1
)
x = torch.rand(1, 16, 32, 32)
y = inv(x) # [1, 16, 32, 32]The paper talks about using Self-Attention for the dynamic kernel generation function. I'll try implementing it later if time permits.
If I've made any errors anywhere in the implementation, please do let me know by raising an issue. If there's any cool addition you want to introduce, all PRs appreciated!