Why choose weights?

[NEMOlv]

Let's review the theory of Class Balanced Loss:
First, the effective number of samples are calculated for all classes as:

Then the class balanced loss function is defined as:

where n_y is the number of samples in the ground-truth class y. In other words, when the number of class is N, the number of E_n is N. This means that each class has a weight. However, the weight is selected in your code.
This is your code：

weights = weights.unsqueeze(0)
weights = weights.repeat(batch_size, 1) * labels_one_hot
weights = weights.sum(1)
weights = weights.unsqueeze(1)
weights = weights.repeat(1, num_classes)

Suppose we have the following data and execute according to your code：

labels_one_hot = [[1,0,0,0],
                  [0,1,0,0],
                  [0,0,1,0],
                  [0,0,0,1]]

weights = [weight_1, weight_2, weight_3, weight_4]

weights = weights.repeat(batch_size, 1) * labels_one_hot

At this time, you can see the output weight：

>>> [[weight_1,0,0,0],
     [0,weight_2,0,0],
     [0,0,weight_3,0],
     [0,0,0,weight_4]]

The weight after executing your subsequent code is：

>>> [[weight_1,weight_1,weight_1,weight_1],
     [weight_2,weight_2,weight_2,weight_2],
     [weight_3,weight_3,weight_3,weight_3],
     [weight_4,weight_4,weight_4,weight_4]]

This means that when the label is [1,0,0,0], loss of all categories should be multiplied by weight_1. Why not multiply by weight_1, weight_2, weight_3 and weight_4. That is, the loss of each category is multiplied by the corresponding weight.