def get_ghm_weight(predict, target, valid_mask, bins=10, alpha=0.75,
               dtype=tf.float32, name='GHM_weight'):
    """ Get gradient Harmonized Weights.
    This is an implementation of the GHM ghm_weights described
    in https://arxiv.org/abs/1811.05181.
    Args:
        predict:
            The prediction of categories branch, [0, 1].
            -shape [batch_num, category_num].
        target:
            The target of categories branch, {0, 1}.
            -shape [batch_num, category_num].
        valid_mask:
            The valid mask, is 0 when the sample is ignored, {0, 1}.
            -shape [batch_num, category_num].
        bins:
            The number of bins for region approximation.
        alpha:
            The moving average parameter.
        dtype:
            The dtype for all operations.
    
    Returns:
        weights:
            The beta value of each sample described in paper.
    """