Awesome

LegoNet

This code is the implementation of ICML2019 paper LegoNet: Efficient Convolutional Neural Networks with Lego Filters

Run

python train.py

You could achieve an VGG16 with 93.88% accuracy on CIFAR10 dataset, the lego filters occupy ~3.8M parameters.

LegoConv2d

self.lego = nn.Parameter(nn.init.kaiming_normal_(torch.rand(self.n_lego, self.basic_channels, self.kernel_size, self.kernel_size)))
self.aux_coefficients = nn.Parameter(init.kaiming_normal_(torch.rand(self.n_split, self.out_channels, self.n_lego, 1, 1)))
self.aux_combination = nn.Parameter(init.kaiming_normal_(torch.rand(self.n_split, self.out_channels, self.n_lego, 1, 1)))

lego: Lego Filters

aux_coefficients: combination coefficients used during combination

aux_combination: combination index

Note

The aux_coefficients and aux_combination should be saved as sparse matrix for saving memory. This code does not include this part.

Citation

@inproceedings{legonet,
	title={LegoNet: Efficient Convolutional Neural Networks with Lego Filters},
	author={Yang, Zhaohui and Wang, Yunhe and Liu, Chuanjian and Chen, Hanting and Xu, Chunjing and Shi, Boxin and Xu, Chao and Xu, Chang},
	booktitle={International Conference on Machine Learning},
	pages={7005--7014},
	year={2019}
}