How can I assign class weights in gluon?


How can I assign class weights in gluon? I see the “weight” parameter in the loss function SoftmaxCrossEntropy, but the documentation is very poor.

Best regards

weight is just a global scalar value. You need to use sample_weight that allows you to weight each sample in the batch.

If by “class weights” you mean class parameters like weights and bias, then here is an example:

class NormalizationHybridLayer(gluon.HybridBlock):
    def __init__(self, hidden_units, scales):
        super(NormalizationHybridLayer, self).__init__()

        with self.name_scope():
            # here you are creating your weights
            self.weights = self.params.get('weights', # you can set name whatever you want
                                           shape=(hidden_units, 0),
            # here you are creating another parameter
            self.scales = self.params.get('scales',
                                      init=mx.init.Constant(scales.asnumpy().tolist()), # Convert to regular list to make this object serializable
    def hybrid_forward(self, F, x, weights, scales):
        normalized_data = F.broadcast_div(F.broadcast_sub(x, F.min(x)), (F.broadcast_sub(F.max(x), F.min(x))))
        weighted_data = F.FullyConnected(normalized_data, weights, num_hidden=self.weights.shape[0], no_bias=True)
        scaled_data = F.broadcast_mul(scales, weighted_data)
        return scaled_data

More info available here


The weight argument that you saw within SoftmayCrossEntropyLoss actually is just a value that would be multiplied by the final loss computed, for instance if your SoftmaxCrossEntropyLoss is L, then it will return weight * L.

Hope this helps.

Hi, could you provide some sample codes? or some links more specific for applying class weights