Skip to content

JackYang825/network-prunning

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

5 Commits
 
 
 
 

Repository files navigation

前提: DNN的过参数化

prunning 优点:

  1. 占的内存少
  2. 运行时计算量少

核心问题是成如何有效地裁剪模型且最小化精度的损失.

基于度量标准的剪枝

通常是提出一个判断神经元是否重要的标准,依据这个标准将不重要的神经元剪掉。主要分为三大块 weight/activation/gradient

Weight:

1.Filter Pruning via Geometric Median for Deep Convolutional Neural Networks Acceleration

去除相似的kernel Filter Pruning via Geometric Median

reference: Robust statistics on riemannian manifolds via the geometric median

Actviation

1.Network Trimming: A Data-Driven Neuron Pruning Approach towards Efficient Deep Architectures 用activations中0的比例 (Average Percentage of Zeros, APoZ)作为度量标准

reference: Learning both weights and connections for efficient neural networks

2.An Entropy-based Pruning Method for CNN Compression 用全局平均池化来转化一个layer的output,得到一个向量。用验证集的一些图片进入网络,就可以得到一个矩阵,这个filter的熵就可以用这个矩阵计算出来。然后排序删除

reference: Binarized neural networks: Training deep neural networks with weights and activations constrained to +1 or -1

Gradient

Pruning Convolutional Neural Networks For Resource Efficient Inference 将prune filter看作一个优化问题。反向传播代价函数以泰勒展开作为准则决定prune 哪些filters。

##基于重建误差的剪枝 这类方法通过最小化特征输出的重建误差来确定哪些filters要进行剪裁,即找到当前层对后面的网络层输出没啥影响的信息。

1.NISP: Pruning Networks using Neuron Importance Score Propagation通过最小化网络倒数第二层的重建误差,并将反向传播的误差累积考虑在内,来决定前面哪些filters需要裁剪。

##基于稀疏训练的剪枝

1.Synaptic Strength For Convolutional Neural Network 基于连接强度修剪。【smaller-norm-less-important】

001.png

reference: Pruning filters for efficient convnets

NAS

1Approximated Oracle Filter Pruning for Destructive CNN Width Optimization行操作网络的所有层,用二分搜索的方式确定每层的剪枝数。用下一层的输出featuremap的变化程度,而非最终的输出来评估删掉一个filter的影响。

2.DSA: More Efficient Budgeted Pruning via Differentiable Sparsity Allocation 在结构化剪枝情况下 对budge训练使得acc可微。

GAN

Towards Optimal Structured CNN Pruning via Generative Adversarial Learning 006.png

  1. 第一个阶段固定生成器和掩膜,通过对抗训练更新判别器D,损失函数包含对抗损失与对抗正则项;
  2. 第二阶段固定判决器,更新生成器与掩膜,损失函数包含对抗损失中的生成器与baseline特征输出的MSE损失以及生成器和掩膜的正则项。

##评估 1. EagleEye: Fast Sub-net Evaluation for Efficient Neural Network Pruning [Bailin Li1, Bowen Wu2, Jiang Su1, Guangrun Wang2, and Liang Lin1,2]

采用自适应批归一化技术进行修剪, 将相关性分析引入修剪算法领域的。

截屏2020-11-26 上午11.02.54.png

Reference:

Revisiting batch normalization for practical domain adaptation.

About

No description, website, or topics provided.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published