A pytorch implementation of MicroNet. If you use this code in your research please consider citing
@article{li2021micronet, title={MicroNet: Improving Image Recognition with Extremely Low FLOPs}, author={Li, Yunsheng and Chen, Yinpeng and Dai, Xiyang and Chen, Dongdong and Liu, Mengchen and Yuan, Lu and Liu, Zicheng and Zhang, Lei and Vasconcelos, Nuno}, journal={arXiv preprint arXiv:2108.05894}, year={2021} }
Requirements
| Model | #Param | MAdds | Top-1 | download |
|---|---|---|---|---|
| MicroNet-M3 | 2.6M | 21M | 62.5 | model |
| MicroNet-M2 | 2.4M | 12M | 59.4 | model |
| MicroNet-M1 | 1.8M | 6M | 51.4 | model |
| MicroNet-M0 | 1.0M | 4M | 46.6 | model |
Download the pretrained MicroNet M0-M3 with the link above. The scripts used for evaluation can be found here. For example, if you want to test MicroNet-M3, you can use the following command.
sh scripts/eval_micronet_m3.sh /path/to/imagenet /path/to/output /path/to/pretrained_modelThe scripts used for training MicroNet M0-M3 can be found here and can be implemented as follows (You can choose to use different scripts for 2 gpu or 4 gpu training based on the resources you can access).
sh scripts/train_micronet_m3_4gpu.sh /path/to/imagenet /path/to/output