This repository has a Python implementation of Two-Level Kronecker Product Factorization Bilinear Pooling for PyTorch by Yu et al, 2022 [paper].
This implementation uses tensorly for tensor folding and tensor modal product calculations.
pip install -U tensorly
class TKPFBilinearPooling(input_dim, a=64, b=64, r=16, q=2, cuda=True)
Basic usage:
from TKPFBilinearPooling import TKPFBilinearPooling
w, h, channels = 7, 7, 2048
bp = TKPFBilinearPooling(channels, cuda=False)
x = torch.rand(4,channels, w, h)
z = bp(x)
- Yu Tan et al. "Efficient Bilinear Pooling via Kronecker Product", 2022