How to Set up the 3D CNN config to do Regression

40 views
Skip to first unread message

steven...@gmail.com

unread,
Jul 1, 2019, 2:30:08 PM7/1/19
to mdCNN: Multidimensional CNN library in Matlab
Hi 

    It seems no description in the WORD document  file of "mdCNN – Multi dimensional CNN" for doing regression using
the 3D cnn in the package.  A straightforward question is:  can we simply replace 'softmax'  with 'regression' in the following
code in the mnist3d.conf file ? 

net.layers{end+1}.properties = struct('type','softmax');


but it had error when I changed this.  



Steven

Hagay Garty

unread,
Jul 2, 2019, 10:06:36 AM7/2/19
to mdCNN: Multidimensional CNN library in Matlab
First, you should remove the softmax layer and change the output layer to use the MSE cost function.
The last layer before the output layer will usually be a 'reshape' layer to match the output size you are expecting.
See autoEnc.conf file for regression example. This conf file is a part of the AutoEnc demo (autoencoder), try to follow the network structure and training.

In the doc file there are some explanation on how to run regression, section 3
Reply all
Reply to author
Forward
0 new messages