All CSDN users to vote have participated in lottery activities
Announced the award, presented within the group there are more benefits
Author & editor | There are three words
Source | Three AI (ID: yanyousan_ai )
REVIEW distilled knowledge and transfer learning not only belongs to one of the important model optimization, but also enhance the important cross-cutting technology model generalization, then what is currently available knowledge distillation and transfer learning open source tools?
1 PaddleSlim
PaddleSlim model is proposed Optimizer Baidu contained in PaddlePaddle framework supports several algorithms knowledge distillation may be added in combination loss teacher and student networks of any network layer, comprising FSP loss, L2 loss, softmax with cross-entropy loss and the like.
https://github.com/PaddlePaddle/models/tree/develop/PaddleSlim
2 Distiller
Distiller is Intel optimization tools Pytorch open source model, support Knowledge distillation Hinton, who proposed algorithm.
https://github.com/NervanaSystems/distiller
3 MxNet
MxNet is integrated Bayesian Dark Knowledge reproduction method, have a role in the recommendation system.
https://github.com/apache/incubator-mxnet/tree/master/example/bayesian-methods
4 unofficial Pytorch project
Knowledge-Distillation-Zoo is a GitHub user AberHu distilled knowledge distillation projects, such as support for multiple models fitnet.
https://github.com/AberHu/Knowledge-Distillation-Zoo
deep-transfer-learning是GitHub用户easezyc整理的迁移学习项目。
https://github.com/easezyc/deep-transfer-learning
5 一些经典模型实现
Hinton等人在2015年提出了经典的知识蒸馏框架Knowledge distillation,下面是各大框架的复现。
[1] Keras
https://github.com/TropComplique/knowledge-distillation-keras
[2] Tensorflow
https://github.com/DushyantaDhyani/kdtf
[3] Caffe
https://github.com/wentianli/knowledge_distillation_caffe
更多的一些经典算法如下,留待读者自己学习,咱们就不沉迷于收藏了。
[1] Relational Knowledge Distillation算法
https://github.com/lenscloth/RKD
[2] Teacher-Assistant-Knowledge-Distillation算法
https://github.com/imirzadeh/Teacher-Assistant-Knowledge-Distillation
[3] Contrastive Representation Distillation算法
https://github.com/HobbitLong/RepDistiller
[4] Zero-shot_Knowledge_Distillation算法
https://github.com/sseung0703/Zero-shot_Knowledge_Distillation
[5] net2net算法
https://github.com/soumith/net2net.torch
[6] fitnet算法
https://github.com/adri-romsor/FitNets
6 开源综述资料
接下来再给大家介绍两个综述性质的开源资料。
第一个是知识蒸馏综述项目,包括相关的论文合集,也包括一些开源算法的实现。
https://github.com/dkozlov/awesome-knowledge-distillation
第二个是迁移学习综述项目,包括相关的论文合集,开源代码。
https://github.com/jindongwang/transferlearning
其他类似的还有https://github.com/artix41/awesome-transfer-learning,https://github.com/AI-ON/Multitask-and-Transfer-Learning,资料收集爱好者不妨关注。
7 更多理论学习
如果想要系统性学习模型优化相关的理论,可以移步有三AI知识星球 -> 网络结构1000变 -> 模型压缩板块 -> 模型剪枝,量化与蒸馏板块,知识蒸馏与迁移学习的一些解读案例如下:
总结
知识蒸馏与迁移学习相比于大量数据进行监督学习的方式,更符合人类学习的特点,在工业界有着实际落地需求,学术界有很重要的研究意义。
(*本文为AI科技大本营转载文章,转载请联系原作者)
◆
精彩推荐
◆
点击阅读原文,或扫描文首贴片二维码
所有CSDN 用户都可参与投票活动
加入福利群,每周还有精选学习资料、技术图书等福利发送
点击投票页面「讲师头像」,60+公开课免费学习
推荐阅读
Python 三十大实践、建议和技巧
2020年AI 2000最具影响力学者榜单发布,何恺明排名超过李飞飞
只十分钟,唾手可得的工作机会就被我搞砸了!
2020年趋势一览:AutoML、联邦学习、云寡头时代的终结
千万不要和程序员一起合租!
十大新兴前端框架大盘点
联盟链走向何方
用比特币的中间层知识完善区块链认知
你点的每个“在看”,我都认真当成了AI