Unbalanced training data

[Douglas Almonfrey]

Hi everyone,

I would like to fine-tune a CNN in Caffe with unbalanced training data. I would like to specify 30% of positive examples and 70% of negative examples in a batch of 256 images. I know that this is possible with the Windows Data Layer, but this layer is about to be deprecated, I think. Besides that, this Windows Data Layer has a difficult to understand image entrance protocol.

So,

Does anybody know how to fine-tune a CNN in Caffe with unbalanced data?