This is an experimental Pytorch implementation of the Ghost-Resnet56. I finish the adaption follow iamhankai’s and akamaster’s work.
- Weights Just as the paper arxiv describes, the number of the parameters of the adapted resnet56 decreased from 0.85m to 0.44m.
- Training Now we can train the Ghostnet and the Ghost Resnet56 on the Cifar-10 dataset, but I cannot obtain the same performance on both models. I have to follow this paper ([12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770–778, 2016.) to achieve the same results.
- Somethng interesting This implements including Ghostnet can be trained on the ipad. You probably have to download an app called "Python AI", which pre-installed the pytorch arm version in it and makes compilation possible. The trained weights of ghostnet take about 20mb.
Simply run the "train_Ghost_ResNet56.py".
- If you want to try different optimization strategy, modify the milestone in Ln88 or uncommen the Ln89 to switch to the adam optimizer.
- Uncommen Ln142, it will beep during each epochs.