Weights for examples about matconvnet

[lenck]

Hi, I think that this may be little bit complicated because the derivatives are collected over a batch inside the e.g. vl_nnconv.cu... So you would probably need to change the mex files and have adjustable alpha parameter in the sgemm operation for the backrpop (e.g. on line 820 in vl_nnconv.cu but on many more places too...).

But if you want to just have a proof of concept, and you don't mind that the computation would take longer, you can pass batches of size 1 and then integrate the gradients in Matlab... Simplest would be to have your own version of vl_simplenn which would iterate over all images in a batch summing up the derivatives...

[lenck]

Hmm, just realised that I might have been wrong about this - you just need to change the softmaxloss function implementation to weight the loss per learning example... [feeling stupid]

[dbbert]

How would I go about changing the softmaxloss function to weigh the examples?

I would adapt lines 59-60 as follows, where the weights vector is based on the label vector c:

t = Xmax + log(sum(ex,3)) - reshape(X(c_), [sz(1:2) 1 sz(4)]) ;
t = t .* weights;
Y = sum(t(:)) / n ;

But what about the backward step?

[shenyu2013]

I also have this problem. After changing the softmaxloss function to weight the examples, how to change the backpropagation steps?
I have tried it if nothing is done, that would cause NaNs and no regression.