• 设为首页
  • 点击收藏
  • 手机版
    手机扫一扫访问
    迪恩网络手机版
  • 关注官方公众号
    微信扫一扫关注
    公众号

TropComplique/knowledge-distillation-keras: A machine learning experiment

原作者: [db:作者] 来自: 网络 收藏 邀请

开源软件名称(OpenSource Name):

TropComplique/knowledge-distillation-keras

开源软件地址(OpenSource Url):

https://github.com/TropComplique/knowledge-distillation-keras

开源编程语言(OpenSource Language):

Jupyter Notebook 79.7%

开源软件介绍(OpenSource Introduction):

Knowledge distillation with Keras

Keras implementation of Hinton's knowledge distillation (KD), a way of transferring knowledge from a large model into a smaller model.

Summary

  • I use Caltech-256 dataset for a demonstration of the technique.
  • I transfer knowledge from Xception to MobileNet-0.25 and SqueezeNet v1.1.
  • Results:
model accuracy, % top 5 accuracy, % logloss
Xception 82.3 94.7 0.705
MobileNet-0.25 64.6 85.9 1.455
MobileNet-0.25 with KD 66.2 86.7 1.464
SqueezeNet v1.1 67.2 86.5 1.555
SqueezeNet v1.1 with KD 68.9 87.4 1.297

Implementation details

  • I use pretrained on ImageNet models.
  • For validation I use 20 images from each category.
  • For training I use 100 images from each category.
  • I use random crops and color augmentation to balance the dataset.
  • I resize all images to 299x299.
  • In all models I train the last two layers.

Notes on flow_from_directory

I use three slightly different versions of Keras' ImageDataGenerator.flow_from_directory:

  • original version for initial training of Xception and MobileNet.
  • ver1 for getting logits from Xception. Now DirectoryIterator.next also outputs image names.
  • ver2 for knowledge transfer. Here DirectoryIterator.next packs logits with hard true targets. All three versions only differ in DirectoryIterator.next function.

Requirements

  • Python 3.5
  • Keras 2.0.6
  • torchvision, Pillow
  • numpy, pandas, tqdm

References

[1] Geoffrey Hinton, Oriol Vinyals, Jeff Dean, Distilling the Knowledge in a Neural Network




鲜花

握手

雷人

路过

鸡蛋
该文章已有0人参与评论

请发表评论

全部评论

专题导读
热门推荐
阅读排行榜

扫描微信二维码

查看手机版网站

随时了解更新最新资讯

139-2527-9053

在线客服(服务时间 9:00~18:00)

在线QQ客服
地址:深圳市南山区西丽大学城创智工业园
电邮:jeky_zhao#qq.com
移动电话:139-2527-9053

Powered by 互联科技 X3.4© 2001-2213 极客世界.|Sitemap