octconv-chainer

Implementation of octave convolution in Chainer (https://arxiv.org/abs/1904.05049)

ResNet50 ImageNet training for 90 epochs:

alpha	validation accuracy	theoretical flop cost
(origin)	0.762	100%
0.25	0.762	67%
0.50	0.757	44%
0.75	0.749	30%

Pretrained models are in the pretrained_models/ directory

To run the pretrained model for validation:

$ python eval_imagenet.py <ImageNet validation directory> --model=octresnet50 --alpha=0.25 --pretrained-model=pretrained_models/octresnet50_alpha25.npz --gpu=0
$ python eval_imagenet.py <ImageNet validation directory> --model=octresnet50 --alpha=0.50 --pretrained-model=pretrained_models/octresnet50_alpha50.npz --gpu=0
$ python eval_imagenet.py <ImageNet validation directory> --model=octresnet50 --alpha=0.75 --pretrained-model=pretrained_models/octresnet50_alpha75.npz --gpu=0

The imageNet dataset preparation follows examples in chainercv. https://github.com/chainer/chainercv/tree/master/examples/classification#how-to-prepare-imagenet-dataset

---

See chainer-cifar10/ for training with cifar10 dataset. (The resnet model with octave convolution is chainer-cifar10/models/oct_resnet.py.)

---

reference: https://github.com/d-li14/octconv.pytorch