论文笔记:Training Region-based Object Detectors with Online Hard Example Mining
2017-10-27 16:35
671 查看
论文地址:https://arxiv.org/abs/1604.03540v1
应该就是机器学习里常用的Boosting算法吧,有名就有AdaBoosting,就是每次训练完成后,把训练错误的样本的权重增加,多次训练得到多个分类器,最后多个分类器联合做决策。但是在现在神经网络中不好用,因为咱训练的时间太长了,不能等到训练完一次再调。
我们就不训完再换阶段了。我们可以在每次的mini-batch SGD中寻找难样本。
loss排序之后保留loss大的有个错误,因为难样本也是多个proposal对应的,可能前几中都是一个区域。所以先NMS(非极大值抑制)
如果按照原来的方法,我们可以把非难样本的loss设为0,就跟往常训练一样了。但是这样0的样本也要bp,浪费计算。所以想了个法子。网络一式两份,一个readonly的做前向传播所有样本,筛选出难样本;一个只负责前向传播和bp训练难样本。网络结构如下:
博主就一直想:为啥要两个?一个不行吗?
原文也说了:
a limitation of current deep learning toolboxes
问题:
正负样本不均衡,总是训练好训练的样本。已有解决办法:
Bootstrapping应该就是机器学习里常用的Boosting算法吧,有名就有AdaBoosting,就是每次训练完成后,把训练错误的样本的权重增加,多次训练得到多个分类器,最后多个分类器联合做决策。但是在现在神经网络中不好用,因为咱训练的时间太长了,不能等到训练完一次再调。
主要解决办法:
Our main observation is that these alternating steps can be combined with how FRCN is trained using online SGD. The key is that although each SGD iteration samples only a small number of images, each image contains thousands of example RoIs from which we can select the hard examples rather than a heuristically sampled subset. This strategy fits the alternation template to SGD by “freezing” the model for only one mini-batch. Thus the model is updated exactly as frequently as with the baseline SGD approach and therefore learning is not delayed.我们就不训完再换阶段了。我们可以在每次的mini-batch SGD中寻找难样本。
实现细节
每次训练的难样本,就是正向计算中loss大的。loss排序之后保留loss大的有个错误,因为难样本也是多个proposal对应的,可能前几中都是一个区域。所以先NMS(非极大值抑制)
如果按照原来的方法,我们可以把非难样本的loss设为0,就跟往常训练一样了。但是这样0的样本也要bp,浪费计算。所以想了个法子。网络一式两份,一个readonly的做前向传播所有样本,筛选出难样本;一个只负责前向传播和bp训练难样本。网络结构如下:
博主就一直想:为啥要两个?一个不行吗?
原文也说了:
a limitation of current deep learning toolboxes
相关文章推荐
- 论文笔记 OHEM: Training Region-based Object Detectors with Online Hard Example Mining
- 论文笔记 | Training Region-based Object Detectors with Online Hard Example Mining
- Training Region-based Object Detectors with Online Hard Example Mining - cvpr 2016 oral
- 【论文笔记】(CVPR2016 Oral) Training Region-based Object Detectors with Online Hard Example Mining
- 论文提要“Training Region-based Object Detectors with Online Hard Example Mining”
- 论文阅读-《Training Region-based Object Detectors with Online Hard Example Mining》
- Training Region-based Object Detectors with Online Hard Example Mining - cvpr 2016 oral
- Training Region-based Object Detectors with Online Hard Example Mining - cvpr 2016 oral
- Training Region-based Object Detectors with Online Hard Example Mining
- 目标检测--Training Region-based Object Detectors with Online Hard Example Mining
- Training Region-based Object Detectors with Online Hard Example Mining
- OHEM-Training Region-based Object Detectors with Online Hard Example Mining - cvpr 2016 oral
- OHEM安装运行(Training Region-based Object Detectors with Online Hard Example Mining)
- Training Region-based Object Detectors with Online Hard Example Mining(CVPR2016 Oral)
- [文献阅读]Training Region-based Object Detectors with Online Hard Example Mining
- 论文笔记--FaceNet & Online Hard Example Mining
- 论文笔记 R-FCN: Object Detection via Region-based Fully Convolutional Networks
- 论文笔记:Faster R-CNN:Towards Real-Time Object Detection with Region Proposal Networks
- 【论文笔记】Region-based Convolutional Networks for Accurate Object Detection and Segmentation
- 论文笔记:Faster R-CNN:Towards Real-Time Object Detection with Region Proposal Networks