FCN-8 loss value is constant throughout training

[jas]

Hello,

I'm training the fcn-8 network, but throughout the training loss function is constant. I'm training the network using my dataset, which has around 176 samples and three classes including background. Should I change the solver parameters, if yes which are parameters which I should I change.

[Evan Shelhamer]

Constant loss usually suggests the initial parameters are not initialized properly. See https://github.com/shelhamer/fcn.berkeleyvision.org/blob/master/voc-fcn32s/solve.py for an example of how to solve.

Evan Shelhamer

On Mon, May 16, 2016 at 6:40 AM, jas <m.s...@gmail.com> wrote:

Hello,

I'm training the fcn-8 network, but throughout the training loss function is constant. I'm training the network using my dataset, which has around 176 samples and three classes including background. Should I change the solver parameters, if yes which are parameters which I should I change.

--
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/05b55e6f-60f2-479e-97d6-e4e48bd009ec%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

[jas]

Thank you Evan. The link which you have provided, I understood that it is used to train fcn32 using vgg16fc.caffemodel. However, I'm using fcn-8 for training, so for that too can I use vgg16fc.caffemodel as base weights.

[Evan Shelhamer]

See the all-at-once version of FCN-8s for how to train it in one step from VGG-16: https://github.com/shelhamer/fcn.berkeleyvision.org/tree/master/voc-fcn8s-atonce

The FCN-32s, FCN-16s, and FCN-8s are otherwise intended to be trained in stages.

Evan Shelhamer

To view this discussion on the web visit https://groups.google.com/d/msgid/caffe-users/b9ffe193-1e11-49bf-92f5-95861f0ea03e%40googlegroups.com.