Preparing config to train

[cmog...@gmail.com]

Hi,
I've been trying to make symbol detector. I prepared 32x32 bit-images (but from jpg) and put it to 2 folders: "symbols" and "background".
Then I run:
dscompile imfolder -outdir outdir -dims 32x32x1 -kernelsz 4x4
and
dssplit . ds ds_train ds_test -maxperclass 60000

I decided to start from existing conf-file and edited mnist.conf from demos folder, changed strings:
root = .
train = ${root}/ds_data.mat # training data
train_labels = ${root}/ds_labels.mat # training labels
val = ${root}/ds_data.mat # validation data
val_labels = ${root}/ds_labels.mat # validation labels

But train my.conf didn't work.
Error: 2x2x2 and 2x1x1 should have the same number of elements, in idx_sqrdist at /home/cmogilko/eblearn/core/libidx/include/idxops.hpp:1795

I thought that difference is only in number of classes and edited mnist config with other dataset will work. What's wrong?

[the_minion]

Hey,

that means that the last layer of your putput is not giving 1x1 but is giving 2x2. You can't do that as the classifier output should be nclasses x 1 x 1. Check your kernel sizes and make sure that the output adheres to the expected size.