# def drop_path(inputs, keep_prob, is_training=True, scope=None):
def drop_path(inputs, keep_prob, is_training=True):
    """Drops out a whole example hiddenstate with the specified probability.
    """
    # with tf.name_scope(scope, 'drop_path', [inputs]):
    net = inputs
    if is_training:
        batch_size = tf.shape(net)[0]
        noise_shape = [batch_size, 1, 1, 1]
        random_tensor = keep_prob
        random_tensor += tf.random_uniform(noise_shape, dtype=tf.float32)
        binary_tensor = tf.floor(random_tensor)
        net = tf.div(net, keep_prob) * binary_tensor
    return net


class DropPath(keras.layers.Layer):
    def __init__(self, drop_prob=None):
        super(DropPath, self).__init__()
        self.drop_prob = drop_prob

    def call(self, inputs,training=None):
        return drop_path(inputs, self.drop_prob, training)

搜索

复制