Implementation of the paper ''Attention Inspiring Receptive-fields Network'' (under review), which contains the evaluation code and trained models. By:
Lu Yang, Qing Song, Yingqi Wu and Mengjie Hu
- Install PyTorch>=0.3.0
- Install torchvision>=0.2.0
- Clone
git clone https://github.com/soeaver/AirNet-PyTorch
- Download the trained models, and move them to the
ckptsfolder. - Run the
eval.py:python eval.py --gpu_id 0 --arch airnet50_1x64d --model_weights ./ckpts/air50_1x64d.pth - The results will be consistent with the paper.
Single-crop (224x224) validation error rate is reported.
| Network | Flops (G) | Params (M) | Top-1 Error (%) | Top-5 Error (%) | Download |
|---|---|---|---|---|---|
| AirNet50-1x64d (r=16) | 4.36 | 25.7 | 22.11 | 6.18 | GoogleDrive |
| AirNet50-1x64d (r=2) | 4.72 | 27.4 | 21.83 | 5.89 | GoogleDrive |
| AirNeXt50-32x4d | 5.29 | 25.5 | 20.87 | 5.52 | GoogleDrive |
Other Resources (from DPNs)
ImageNet-1k Trainig/Validation List:
- Download link: GoogleDrive
ImageNet-1k category name mapping table:
- Download link: GoogleDrive