Incubator-mxnet: weighting gradient feedback of each instance wrt class distribution in case of class imbalance

[erogol]

Given a imbalanced data I would like to set the contribution of each instance to the loss value and the gradient wrt the portion of data with the same category. For instance, if this is 2 classes problem then

Class A with K instances
ClassB with K/2 instances

which means Class B contributes to gradient feedback with coefficient 1 and Class A with coefficient 1/2.

What would be the be4st way to implement it in mxnet? Or do you have any other trick to cope with class imbalance?

[piiswrong]

You can do it by adding an extra aux variable to the loss layer that holds the scales.
Would you like to propose a PR?

[erogol]

What would be the use of this idea for a new PR? Is it an additional argument to loss layers or any other method? Caffe solves this by using infogain layer which uses a class-wise relation matrix and you can define diagonals of the matrix with these frequency scalers.

[jsusskin]

I'm interested in this as well. It can be done by writing custom loss functions of course.

[fedorzh]

How did you resolve it finally?

[benwu232]

is there any example for this?

[yajiedesign]

This issue is closed due to lack of activity in the last 90 days. Feel free to reopen if this is still an active issue. Thanks!