loss_func

def func_loss(tar, logits):
    loss_object = tf.keras.losses.CategoricalCrossentropy(from_logits=True, reduction='none')
    mask = tf.math.logical_not(tf.math.equal(tar, 0))
    one_hot_labels = tf.one_hot(tar, depth=tok_bpe_obj.voc_size_punct)
    class_weights = tf.constant(tf.range(1.0, tok_bpe_obj.voc_size_punct+1))
    weights = tf.reduce_sum(class_weights * one_hot_labels, axis=-1)
    loss = loss_object(one_hot_labels, logits)
    mask = tf.cast(mask, dtype=loss.dtype)
    loss = loss * weights * mask
    loss = tf.reduce_mean(loss)
    return loss