InfogainLossLayer doesn't work

[Henry]

Hi all

I had very imbalanced training data and I plan to use the InfogainLoss layer to solve it.  The code is shown below:

n.softmax_score = L.Softmax(n.score)

n.loss = L.InfogainLoss(n.softmax_score, n.label,

            loss_param=dict(normalization=True, ignore_label=0),

            infogain_loss_param=dict(source="infogainH.binaryproto"))

The values in the diagonal line of tinfogainH.binaryproto matrix are 1, 1, 1, 2, 2, 4, 1, 1 ............. and the values in other position are 0. But during training, the loss doesn't decrease. When I use SoftmaxWithLoss layer, the loss decrease very quickly.

I changed the infogainH.binaryproto matrix to an identity matrix. But the loss still doesn't decrease. The official document of InfogainLoss layer shows that "If H = I, this layer is equivalent to the MultinomialLogisticLossLayer." 

Any idea where is my mistakes? Many thanks for your help!

Best

Henry

[Henry]

bump!