def get_losses()

in train.py [0:0]

• 17 lines of code
• 6 McCabe index (conditional complexity)

def get_losses(logits, labels, mask=None):
    with tf.name_scope('loss'):

        n, t, nx = shape_list(logits)
        ln, lt = shape_list(labels)
        assert lt == t
        labels = tf.reshape(labels, [-1])
        logits = tf.reshape(logits, [-1, nx])

        if H.float16 and logits.shape[-1].value <= 65536 and logits.dtype == tf.float16:
            # much faster fused fp16 implementation that also saves memory
            losses = bs.softmax_cross_entropy(logits=logits, labels=labels)
        else:
            logits = tf.cast(logits, tf.float32)
            losses = tf.nn.sparse_softmax_cross_entropy_with_logits(
                logits=logits, labels=labels)

        losses = tf.reshape(losses, [n, t])
        if mask is not None:
            # X_mask can be either boolean or scalar (weighted) mask
            return (tf.reduce_sum(losses * mask) / tf.reduce_sum(mask)), losses
        return tf.reduce_mean(losses), losses