Train a linear SVM on caffe

[Alexandre Dalyac]

Hello,

I'm interested in reproducing the work done in 'CNN features off the shelf' by Razvian et al, where a linear SVM is trained on the feature space from AlexNet layer 7.

What architecture do I need (on top of alexnet) to achieve that? 1 inner product layer with num_outputs equal to number of classes, then one hinge loss layer on top?

Thanks in advance!

[Alexandre Dalyac]

edit: 1 inner product layer with num_outputs equal to number of classes, then 1 eltwise product layer, then one hinge loss layer on top?

[lixin7...@gmail.com]

Hi , have you implemented linear svm in this way?

在 2014年8月27日星期三UTC+8上午10时19分43秒，Alexandre Dalyac写道：

[evancompu...@gmail.com]

Has anyone tried using a random forest classifier instead of SVM for this scenario? I'd love to learn pros/cons.