考虑加入极市专业CV交流群,与 10000+来自港科大、北大、清华、中科院、CMU、腾讯、百度 等名校名企视觉开发者互动交流!
同时提供每月大咖直播分享、真实项目需求对接、干货资讯汇总,行业技术交流。关注 极市平台 公众号 ,回复 加群,立刻申请入群~
文章结构: 一、前言 二、论文阅读 三、代码实验(实验1. 训练模型/实验2. 计算Weights和FLOPs/实验3. 可视化特征图) 四、Ghost Module下一步工作探讨(方案1:借助baseline的卷积核参数/方案2:真正意义的即插即用/方案3:定制符合Ghost module的baseline) 五、结语
"For exam- ple, Figure 1 presents some feature maps of an input image generated by ResNet-50, and there exist many similar pairs of feature maps, like a ghost of each another".
第三段”the output feature maps of convolutional layers often contain much redundancy, and some of them could be similar with each other“
“Note that the linear operations Φ operate on each channel whose computational cost is much less than the ordinary convolution. In practice, there could be several different linear operations in a Ghost module, e.g. 3 × 3 and 5 × 5 linear kernels, which will be analyzed in the experiment part.”
“Besides convolutions used in the above experiments, we can also explore some other low-cost linear operations to construct the Ghost module such as affine transformation and wavelet transformation. However, convolution is an efficient operation already well support by current hardware and it can cover a number of widely used linear operations such as smoothing, blurring, motion, etc. Moreover, although we can also learn the size of each filter w.r.t. the linear operation Φ, the irregular module will reduce the efficiency of computing units (e.g. CPU and GPU).”
nohup python 02_main.py -gpu 1 0 -arc resnet56 > resnet56.log 2>&1 &
nohup python bin/02_main.py -max_epoch 190 -lr 0.1 -gpu 1 0 -arc resnet56 -replace_conv > ghost-resnet56.log 2>&1 &
nohup python bin/02_main.py -max_epoch 190 -lr 0.1 -gpu 1 0 -arc vgg16 > vgg16.log 2>&1 &
nohup python bin/02_main.py -max_epoch 190 -lr 0.1 -gpu 1 0 -arc vgg16 -replace_conv > ghost-vgg16.log 2>&1 &
联系客服