Gluon implementation of anti-aliased CNNs: https://arxiv.org/abs/1904.11486
Based on original PyTorch implementation: https://github.com/adobe/antialiased-cnns
Example of training resnet20_v1 with anti-aliasing and random crop augmentation:
python3 train_cifar10.py --mode hybrid --num-gpus 1 -j 8 --batch-size 128 --num-epochs 186 --lr 0.003 --lr-decay 0.1 --lr-decay-epoch 81,122 --wd 0.0001 --optimizer adam --model cifar_resnet20_v1 --antialiasing --random-crop
| Model | random crop | anti-aliasing | Train accuracy | Test accuracy |
|---|---|---|---|---|
| cifar_resnet20_v1 | ✘ | ✘ | 1.0000 | 0.8879 |
| ✘ | ✔ | 1.0000 | 0.9026 | |
| ✔ | ✘ | 0.9918 | 0.9165 | |
| ✔ | ✔ | 0.9960 | 0.9184 | |
| cifar_resnet20_v2 | ✘ | ✘ | 1.0000 | 0.8850 |
| ✘ | ✔ | 0.9999 | 0.9051 | |
| ✔ | ✘ | 0.9891 | 0.9114 | |
| ✔ | ✔ | 0.9953 | 0.9084 |