RegNet Implementation with TorchVision Style

PyTorch implementation of Designing Network Design Spaces by Ilija Radosavovic, Raj Prateek Kosaraju, Ross Girshick, Kaiming He, and Piotr Dollár.

Compared to the official codebase, this repository follows the torchvision's ResNeXt style, which is expected to be more easily interpreted and utilized by pre-existing downstream applications.

We train the following models on 8x TITAN XP GPUs with 12G VRAM. During the first five epochs, we linearly ramp up the learning rate from 0.1.

Pre-trained Models

Model	Params (M)	GFLOPs	Batch size	Top-1 acc (%) (our impl.)	Top-1 acc (%) (official)
RegNetX-200M	2.685	0.199	1024	68.210	68.9
RegNetX-400M	5.158	0.398	1024	72.278	72.7
RegNetX-600M	6.196	0.601	1024	73.862	74.1
RegNetX-800M	7.260	0.800	1024	74.940	75.2
RegNetX-1.6G	9.190	1.603	1024	76.706	77.0
RegNetX-3.2G	15.296	3.177	512	78.188	78.3
RegNetX-4.0G	22.118	3.965	512	78.690	78.6
RegNetX-6.4G	26.209	6.460	512	79.152	79.2
RgeNetX-8.0G	39.573	7.995	512	79.380	79.3
RegNetX-12G	46.106	12.087	256	79.998	79.7
RegNetX-16G	54.279	15.941	256	80.118	80.0
RegNetX-32G	107.812	31.736	256	80.516	80.5

Citation

@InProceedings{Radosavovic_2020_CVPR,
author = {Radosavovic, Ilija and Kosaraju, Raj Prateek and Girshick, Ross and He, Kaiming and Doll{\'a}r, Piotr},
title = {Designing Network Design Spaces},
booktitle = {The IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
month = {June},
year = {2020}
}