FCN32 training

[Robert Krutsch]

Hi,

I am facing some issue in understanding the fcn32 loss that I am getting ( took the staff from shelhamer , https://github.com/shelhamer/fcn.berkeleyvision.org ). I looked at the label and the output from the network (the network is the one provided in url). Interestingly, from visual inspection the label and output match.. I've attached the images, they correspond to VOC image 2009_001885.

The loss that I am getting is 36968.641 , as computed by the score.py script. In the prototext I do have ignore_label: 255 .. The interesting thing is , without it I get some blass error anyway. It looks to me like it the error should not be that big..

When training, the loss also oscillates and is generally big (8-30K) for a learning rate 1e-10..

Did someone face the same ?

Thank you,
 Rob

[Evan Shelhamer]

The loss oscillates in FCN training but goes down as learning progresses. The loss values will generally be high since they are the sum of the loss over all pixels of the image. The loss is not normalized so that every pixel in the data contributes to the loss equally instead of having a weight dependent on the size of the image it is in.

In short: you will learn to stop worrying and love the unnormalized loss.

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/eacc972d-d837-40ce-a191-50941b4014f1%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.