Loss function is not LMCL

[petrLorenc]

How is this loss function:

def large_margin_cosine_loss(y_true, y_pred, scale=30, margin=0.35):
    y_pred = y_true * (y_pred - margin) + (1 - y_true) * y_pred
    y_pred *= scale
    return K.categorical_crossentropy(y_true, y_pred, from_logits=True)

related to 2.2 Large Margin Cosine Loss (LMCL) in the reference paper

It seems totally different, right?

[petrLorenc]

According to the original idea of the paper, the softmax layer should be applied after the margin/scale. In this case, it was applied beforehand so it seems not the same.