load individually trained models to parallel/ensemble network

[Siavash Sakhavi]

Hi

If we have N parallel trained models  (Which are not necessarily the same), each of them having their own .prototxt and .caffe file, 

How should a prototxt file and solver be written in order to initialize a parallel network with each of these individual networks and concatenate their output blobs?

As an example, you can look at this paper : http://arxiv.org/abs/1505.01749 , page 3

[张乐]

Hi, I am thinking about the same question.

So far have you got any idea about that?

在 2015年5月18日星期一 UTC+8上午11:58:09，Siavash Sakhavi写道：

[Evan Shelhamer]

Call `caffe train` with a comma separated list of `.caffemodels` to draw the weights from as in https://github.com/BVLC/caffe/pull/1456 thanks to @jyegerlehner. Be careful that the models have layers with distinct names.

In more complicated cases you may have to do net surgery: http://nbviewer.ipython.org/github/BVLC/caffe/blob/master/examples/net_surgery.ipynb. In this way you can assign whatever weights you like by loading multiple nets and transferring parameters through Python ndarrays.

Evan Shelhamer

--
You received this message because you are subscribed to the Google Groups "Caffe Users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to caffe-users...@googlegroups.com.
To post to this group, send email to caffe...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/caffe-users/af4f2b80-c173-4255-834a-270b6e4c1bd3%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.