Deep Compression/Acceleration（模型压缩加速总结）

淩亂°似流年 2022-04-06 04:22 457阅读 0赞

### 模型压缩论文目录 ###

*  结构\`structure\`
 *   *   *  \[CVPR2019\] Searching for MobileNetV3
         *  \[BMVC2018\] IGCV3: Interleaved Low-Rank Group Convolutions for Efficient Deep Neural Networks
         *  \[CVPR2018\] IGCV2: Interleaved Structured Sparse Convolutional Neural Networks
         *  \[CVPR2018\] MobileNetV2: Inverted Residuals and Linear Bottlenecks
         *  \[ECCV2018\] ShuffleNet V2: Practical Guidelines for Efficient CNN Architecture Design
 *  量化\`quantization\`
 *   *   *  Binarized Neural Networks: Training Deep Neural Networks with Weights and Activations Constrained to +1 or -1
         *  \[ACM2017\] FINN: A Framework for Fast, Scalable Binarized Neural Network Inference
         *  \[CVPR2016\] DoReFa-Net: Training Low Bitwidth Convolutional Neural Networks with Low Bitwidth Gradients
         *  \[CVPR2016\] XNOR-Net: ImageNet Classification Using Binary Convolutional Neural Networks
         *  \[CVPR2016\] Ternary Weight Networks
         *  Quantization and Training of Neural Networks for Efficient Integer-Arithmetic-Only Inference
         *  \[ACM2017\] Quantized Neural Networks: Training Neural Networks with Low Precision Weights and Activations
         *  Two-Step Quantization for Low-bit Neural Networks
 *  剪枝\`pruning\`
 *   *  通道裁剪\`channel pruning\`
     *   *  \[NIPS2018\] Discrimination-aware Channel Pruning for Deep Neural Networks
         *  \[ICCV2017\] Channel Pruning for Accelerating Very Deep Neural Networks
         *  \[ECCV2018\] AMC: AutoML for Model Compression and Acceleration on Mobile Devices
         *  \[ICCV2017\] Learning Efficient Convolutional Networks through Network Slimming
         *  \[ICLR2018\] Rethinking the Smaller-Norm-Less-Informative Assumption in Channel Pruning of Convolution Layers
         *  \[CVPR2017\] NISP: Pruning Networks using Neuron Importance Score Propagation
         *  \[ICCV2017\] ThiNet: A Filter Level Pruning Method for Deep Neural Network Compression
     *  稀疏\`sparsity\`
     *   *  SBNet: Sparse Blocks Network for Fast Inference
         *  To Prune, or Not to Prune: Exploring the Efficacy of Pruning for Model Compression
         *  Submanifold Sparse Convolutional Networks
         *  MorphNet: Fast & Simple Resource-Constrained Learning of Deep Network Structure
 *  融合\`fusion\`
 *  蒸馏\`distillation\`
 *   *   *  \[NIPS2014\] Distilling the Knowledge in a Neural Network
 *  综合\`comprehensive\`
 *   *   *  \[ICLR2016\] Deep Compression: Compressing Deep Neural Network with Pruning, Trained Quantization and Huffman Coding
         *  Model Distillation with Knowledge Transfer from Face Classification to Alignment and Verification

根据个人理解将模型压缩方面研究分为以下七个方向：

# 结构`structure` #

### \[CVPR2019\] Searching for MobileNetV3 ###

*  intro: 神经网络结构搜索NAS 、强化学习
 *  arxiv: [https://arxiv.org/abs/1905.02244][https_arxiv.org_abs_1905.02244]
 *  github: [https://github.com/xiaolai-sqlai/mobilenetv3][https_github.com_xiaolai-sqlai_mobilenetv3]
 *  github: [https://github.com/leaderj1001/MobileNetV3-Pytorch][https_github.com_leaderj1001_MobileNetV3-Pytorch]

### \[BMVC2018\] IGCV3: Interleaved Low-Rank Group Convolutions for Efficient Deep Neural Networks ###

*  intro:
 *  arxiv:[https://arxiv.org/abs/1806.00178][https_arxiv.org_abs_1806.00178]
 *  github:[https://github.com/homles11/IGCV3][https_github.com_homles11_IGCV3]

### \[CVPR2018\] IGCV2: Interleaved Structured Sparse Convolutional Neural Networks ###

*  intro:
 *  arxiv:[https://arxiv.org/abs/1804.06202][https_arxiv.org_abs_1804.06202]
 *  同上

### \[CVPR2018\] MobileNetV2: Inverted Residuals and Linear Bottlenecks ###

*  intro:
 *  arxiv:[https://arxiv.org/abs/1801.04381][https_arxiv.org_abs_1801.04381]
 *  github:[https://github.com/tensorflow/models/tree/master/research/slim/nets/mobilenet][https_github.com_tensorflow_models_tree_master_research_slim_nets_mobilenet]

### \[ECCV2018\] ShuffleNet V2: Practical Guidelines for Efficient CNN Architecture Design ###

*  intro:
 *  arxiv:[https://arxiv.org/abs/1807.11164][https_arxiv.org_abs_1807.11164]
 *  github:

# 量化`quantization` #

### Binarized Neural Networks: Training Deep Neural Networks with Weights and Activations Constrained to +1 or -1 ###

*  intro:二值网络
 *  arxiv:[https://arxiv.org/abs/1602.02830][https_arxiv.org_abs_1602.02830]
 *  github: [https://github.com/MatthieuCourbariaux/BinaryNet][https_github.com_MatthieuCourbariaux_BinaryNet]  
    [https://github.com/itayhubara/BinaryNet][https_github.com_itayhubara_BinaryNet]

### \[ACM2017\] FINN: A Framework for Fast, Scalable Binarized Neural Network Inference ###

*  intro:二值网络
 *  pdf:[http://www.idi.ntnu.no/~yamanu/2017-fpga-finn-preprint.pdf][http_www.idi.ntnu.no_yamanu_2017-fpga-finn-preprint.pdf]
 *  github:[https://github.com/Xilinx/FINN][https_github.com_Xilinx_FINN]

### \[CVPR2016\] DoReFa-Net: Training Low Bitwidth Convolutional Neural Networks with Low Bitwidth Gradients ###

*  intro:低bit位
 *  arxiv:[https://arxiv.org/abs/1606.06160][https_arxiv.org_abs_1606.06160]
 *  github:[https://github.com/tensorpack/tensorpack/tree/master/examples/DoReFa-Net][https_github.com_tensorpack_tensorpack_tree_master_examples_DoReFa-Net]

### \[CVPR2016\] XNOR-Net: ImageNet Classification Using Binary Convolutional Neural Networks ###

*  intro:darknet团队出品
 *  arxiv:[https://arxiv.org/abs/1603.05279][https_arxiv.org_abs_1603.05279]
 *  github:[https://github.com/allenai/XNOR-Net][https_github.com_allenai_XNOR-Net]

### \[CVPR2016\] Ternary Weight Networks ###

*  intro:
 *  arxiv:[https://arxiv.org/abs/1605.04711][https_arxiv.org_abs_1605.04711]
 *  github:[https://github.com/fengfu-chris/caffe-twns][https_github.com_fengfu-chris_caffe-twns]

### Quantization and Training of Neural Networks for Efficient Integer-Arithmetic-Only Inference ###

*  Google出品
 *  arxiv:[https://arxiv.org/abs/1712.05877][https_arxiv.org_abs_1712.05877]
 *  github:[https://github.com/tensorflow/tensorflow/tree/master/tensorflow/contrib/quantize][https_github.com_tensorflow_tensorflow_tree_master_tensorflow_contrib_quantize]

### \[ACM2017\] Quantized Neural Networks: Training Neural Networks with Low Precision Weights and Activations ###

*  intro:QNNs
 *  arxiv:[https://arxiv.org/abs/1609.07061][https_arxiv.org_abs_1609.07061]
 *  github:[https://github.com/peisuke/qnn][https_github.com_peisuke_qnn]

### Two-Step Quantization for Low-bit Neural Networks ###

*  intro:
 *  paper:[http://openaccess.thecvf.com/content\_cvpr\_2018/papers/Wang\_Two-Step\_Quantization\_for\_CVPR\_2018\_paper.pdf][http_openaccess.thecvf.com_content_cvpr_2018_papers_Wang_Two-Step_Quantization_for_CVPR_2018_paper.pdf]
 *  github:

# 剪枝`pruning` #

## 通道裁剪`channel pruning` ##

### \[NIPS2018\] Discrimination-aware Channel Pruning for Deep Neural Networks ###

*  intro:
 *  arxiv:[https://arxiv.org/abs/1810.11809][https_arxiv.org_abs_1810.11809]
 *  github:[https://github.com/Tencent/PocketFlow][https_github.com_Tencent_PocketFlow]支持DisChnPrunedLearner

### \[ICCV2017\] Channel Pruning for Accelerating Very Deep Neural Networks ###

*  intro:Lasso回归
 *  arxiv:[https://arxiv.org/abs/1707.06168][https_arxiv.org_abs_1707.06168]
 *  github:[https://github.com/yihui-he/channel-pruning][https_github.com_yihui-he_channel-pruning]

### \[ECCV2018\] AMC: AutoML for Model Compression and Acceleration on Mobile Devices ###

*  intro:自动学习优化
 *  arxiv:[https://arxiv.org/abs/1802.03494][https_arxiv.org_abs_1802.03494]

1.  [https://www.jiqizhixin.com/articles/AutoML-for-Model-Compression-and-Acceleration-on-Mobile-Devices][https_www.jiqizhixin.com_articles_AutoML-for-Model-Compression-and-Acceleration-on-Mobile-Devices]论文翻译

*  github:[https://github.com/Tencent/PocketFlow][https_github.com_Tencent_PocketFlow]

### \[ICCV2017\] Learning Efficient Convolutional Networks through Network Slimming ###

*  intro:Zhuang Liu
 *  arxiv:[https://arxiv.org/abs/1708.06519][https_arxiv.org_abs_1708.06519]
 *  github:[https://github.com/Eric-mingjie/network-slimming][https_github.com_Eric-mingjie_network-slimming]  
    [https://github.com/foolwood/pytorch-slimming][https_github.com_foolwood_pytorch-slimming]

### \[ICLR2018\] Rethinking the Smaller-Norm-Less-Informative Assumption in Channel Pruning of Convolution Layers ###

*  intro:
 *  arxiv:[https://arxiv.org/abs/1802.00124][https_arxiv.org_abs_1802.00124]
 *  github:[\[PyTorch\]https://github.com/jack-willturner/batchnorm-pruning][PyTorch_https_github.com_jack-willturner_batchnorm-pruning]  
    [\[TensorFlow\]https://github.com/bobye/batchnorm\_prune][TensorFlow_https_github.com_bobye_batchnorm_prune]

### \[CVPR2017\] NISP: Pruning Networks using Neuron Importance Score Propagation ###

*  intro:
 *  arxiv:[https://arxiv.org/abs/1711.05908][https_arxiv.org_abs_1711.05908]
 *

### \[ICCV2017\] ThiNet: A Filter Level Pruning Method for Deep Neural Network Compression ###

*  intro:
 *  web:[http://lamda.nju.edu.cn/luojh/project/ThiNet\_ICCV17/ThiNet\_ICCV17\_CN.html][http_lamda.nju.edu.cn_luojh_project_ThiNet_ICCV17_ThiNet_ICCV17_CN.html]
 *  github:[https://github.com/Roll920/ThiNet][https_github.com_Roll920_ThiNet]  
    [https://github.com/Roll920/ThiNet\_Code][https_github.com_Roll920_ThiNet_Code]

## 稀疏`sparsity` ##

### SBNet: Sparse Blocks Network for Fast Inference ###

*  intro: Uber
 *  arxiv:[https://arxiv.org/abs/1801.02108][https_arxiv.org_abs_1801.02108]
 *  github:[https://github.com/uber/sbnet][https_github.com_uber_sbnet]

### To Prune, or Not to Prune: Exploring the Efficacy of Pruning for Model Compression ###

*  intro:稀疏
 *  arxiv:[https://arxiv.org/abs/1710.01878][https_arxiv.org_abs_1710.01878]
 *  github:[https://github.com/tensorflow/tensorflow/tree/master/tensorflow/contrib/model\_pruning][https_github.com_tensorflow_tensorflow_tree_master_tensorflow_contrib_model_pruning]

### Submanifold Sparse Convolutional Networks ###

*  intro:Facebook
 *  arxiv:[https://arxiv.org/abs/1706.01307][https_arxiv.org_abs_1706.01307]
 *  github:[https://github.com/facebookresearch/SparseConvNet][https_github.com_facebookresearch_SparseConvNet]

### MorphNet: Fast & Simple Resource-Constrained Learning of Deep Network Structure ###

*  intro: Google，正则化
 *  arxiv: [https://arxiv.org/abs/1711.06798][https_arxiv.org_abs_1711.06798]
 *  github: [https://github.com/google-research/morph-net][https_github.com_google-research_morph-net]

# 融合`fusion` #

# 蒸馏`distillation` #

### \[NIPS2014\] Distilling the Knowledge in a Neural Network ###

*  intro:Hinton出品
 *  arxiv:[https://arxiv.org/abs/1503.02531][https_arxiv.org_abs_1503.02531]
 *  github:[https://github.com/peterliht/knowledge-distillation-pytorch][https_github.com_peterliht_knowledge-distillation-pytorch]

# 综合`comprehensive` #

### \[ICLR2016\] Deep Compression: Compressing Deep Neural Network with Pruning, Trained Quantization and Huffman Coding ###

*  intro:开创先河
 *  arxiv:[https://arxiv.org/abs/1510.00149][https_arxiv.org_abs_1510.00149]
 *  github:[https://github.com/songhan][https_github.com_songhan]

### Model Distillation with Knowledge Transfer from Face Classification to Alignment and Verification ###

*  intro:实验比较多，适合工程化
 *  arxiv:[https://arxiv.org/abs/1709.02929][https_arxiv.org_abs_1709.02929]

[https_arxiv.org_abs_1905.02244]: https://arxiv.org/abs/1905.02244
[https_github.com_xiaolai-sqlai_mobilenetv3]: https://github.com/xiaolai-sqlai/mobilenetv3
[https_github.com_leaderj1001_MobileNetV3-Pytorch]: https://github.com/leaderj1001/MobileNetV3-Pytorch
[https_arxiv.org_abs_1806.00178]: https://arxiv.org/abs/1806.00178
[https_github.com_homles11_IGCV3]: https://github.com/homles11/IGCV3
[https_arxiv.org_abs_1804.06202]: https://arxiv.org/abs/1804.06202
[https_arxiv.org_abs_1801.04381]: https://arxiv.org/abs/1801.04381
[https_github.com_tensorflow_models_tree_master_research_slim_nets_mobilenet]: https://github.com/tensorflow/models/tree/master/research/slim/nets/mobilenet
[https_arxiv.org_abs_1807.11164]: https://arxiv.org/abs/1807.11164
[https_arxiv.org_abs_1602.02830]: https://arxiv.org/abs/1602.02830
[https_github.com_MatthieuCourbariaux_BinaryNet]: https://github.com/MatthieuCourbariaux/BinaryNet
[https_github.com_itayhubara_BinaryNet]: https://github.com/itayhubara/BinaryNet
[http_www.idi.ntnu.no_yamanu_2017-fpga-finn-preprint.pdf]: http://www.idi.ntnu.no/~yamanu/2017-fpga-finn-preprint.pdf
[https_github.com_Xilinx_FINN]: https://github.com/Xilinx/FINN
[https_arxiv.org_abs_1606.06160]: https://arxiv.org/abs/1606.06160
[https_github.com_tensorpack_tensorpack_tree_master_examples_DoReFa-Net]: https://github.com/tensorpack/tensorpack/tree/master/examples/DoReFa-Net
[https_arxiv.org_abs_1603.05279]: https://arxiv.org/abs/1603.05279
[https_github.com_allenai_XNOR-Net]: https://github.com/allenai/XNOR-Net
[https_arxiv.org_abs_1605.04711]: https://arxiv.org/abs/1605.04711
[https_github.com_fengfu-chris_caffe-twns]: https://github.com/fengfu-chris/caffe-twns
[https_arxiv.org_abs_1712.05877]: https://arxiv.org/abs/1712.05877
[https_github.com_tensorflow_tensorflow_tree_master_tensorflow_contrib_quantize]: https://github.com/tensorflow/tensorflow/tree/master/tensorflow/contrib/quantize
[https_arxiv.org_abs_1609.07061]: https://arxiv.org/abs/1609.07061
[https_github.com_peisuke_qnn]: https://github.com/peisuke/qnn
[http_openaccess.thecvf.com_content_cvpr_2018_papers_Wang_Two-Step_Quantization_for_CVPR_2018_paper.pdf]: http://openaccess.thecvf.com/content_cvpr_2018/papers/Wang_Two-Step_Quantization_for_CVPR_2018_paper.pdf
[https_arxiv.org_abs_1810.11809]: https://arxiv.org/abs/1810.11809
[https_github.com_Tencent_PocketFlow]: https://github.com/Tencent/PocketFlow
[https_arxiv.org_abs_1707.06168]: https://arxiv.org/abs/1707.06168
[https_github.com_yihui-he_channel-pruning]: https://github.com/yihui-he/channel-pruning
[https_arxiv.org_abs_1802.03494]: https://arxiv.org/abs/1802.03494
[https_www.jiqizhixin.com_articles_AutoML-for-Model-Compression-and-Acceleration-on-Mobile-Devices]: https://www.jiqizhixin.com/articles/AutoML-for-Model-Compression-and-Acceleration-on-Mobile-Devices
[https_arxiv.org_abs_1708.06519]: https://arxiv.org/abs/1708.06519
[https_github.com_Eric-mingjie_network-slimming]: https://github.com/Eric-mingjie/network-slimming
[https_github.com_foolwood_pytorch-slimming]: https://github.com/foolwood/pytorch-slimming
[https_arxiv.org_abs_1802.00124]: https://arxiv.org/abs/1802.00124
[PyTorch_https_github.com_jack-willturner_batchnorm-pruning]: https://github.com/jack-willturner/batchnorm-pruning
[TensorFlow_https_github.com_bobye_batchnorm_prune]: https://github.com/bobye/batchnorm_prune
[https_arxiv.org_abs_1711.05908]: https://arxiv.org/abs/1711.05908
[http_lamda.nju.edu.cn_luojh_project_ThiNet_ICCV17_ThiNet_ICCV17_CN.html]: http://lamda.nju.edu.cn/luojh/project/ThiNet_ICCV17/ThiNet_ICCV17_CN.html
[https_github.com_Roll920_ThiNet]: https://github.com/Roll920/ThiNet
[https_github.com_Roll920_ThiNet_Code]: https://github.com/Roll920/ThiNet_Code
[https_arxiv.org_abs_1801.02108]: https://arxiv.org/abs/1801.02108
[https_github.com_uber_sbnet]: https://github.com/uber/sbnet
[https_arxiv.org_abs_1710.01878]: https://arxiv.org/abs/1710.01878
[https_github.com_tensorflow_tensorflow_tree_master_tensorflow_contrib_model_pruning]: https://github.com/tensorflow/tensorflow/tree/master/tensorflow/contrib/model_pruning
[https_arxiv.org_abs_1706.01307]: https://arxiv.org/abs/1706.01307
[https_github.com_facebookresearch_SparseConvNet]: https://github.com/facebookresearch/SparseConvNet
[https_arxiv.org_abs_1711.06798]: https://arxiv.org/abs/1711.06798
[https_github.com_google-research_morph-net]: https://github.com/google-research/morph-net
[https_arxiv.org_abs_1503.02531]: https://arxiv.org/abs/1503.02531
[https_github.com_peterliht_knowledge-distillation-pytorch]: https://github.com/peterliht/knowledge-distillation-pytorch
[https_arxiv.org_abs_1510.00149]: https://arxiv.org/abs/1510.00149
[https_github.com_songhan]: https://github.com/songhan
[https_arxiv.org_abs_1709.02929]: https://arxiv.org/abs/1709.02929