class balancing

[Klaymen]

Hello,

I'm trying to add weights to the classes during training right before the final softmax layer, the output shape of the layer is (batch_size, length, nb_class) with one-hot-label of the same size, and I'm using categorical_crossentropy loss here.

I've tried to use class_weight in fit and send class weight as dictionary: {0:w1, 1:w2, ...}, but got the error: Exception: class_weight not supported for 3+ dimensional targets.

Is there any way to do this? Any suggestions will be greatly appreciated!

[Sander Stepanov]

good question

[Isaac Gerg]

I'm using class_weight for a 2 class problem with no issues.

Can you post some code?

On Tuesday, February 2, 2016 at 4:06:01 AM UTC-5, Klaymen wrote:

[Alex]

I am getting the same error with just 2 class weights. See details here: https://github.com/fchollet/keras/issues/5437  Any suggestions?

[al...@goodman.io]

You should implement a custom loss objective, e.g.

https://github.com/fchollet/keras/issues/2115

[Alex]

Thanks for the reply! The thread you pointed out should help. I was expecting that Keras would have integrated the class balancing in `class_weight`.