How to generate a random size crop using tensorflow.dataset module

I want to generate a random size crop when using tensorflow dataset API with tfrecord file. Here is my code

dataset = tf.data.TFRecordDataset(filenames)
dataset = dataset.map(parse_func)
dataset.shuffle(buffer_size=10000).batch(batch_size).repeat()
iterator = dataset.make_initializable_iterator()
data_in_tsr, data_gt_tsr,  = iterator.get_next()

def parse_func(example_proto):
    features = {
        'in': tf.FixedLenFeature(image_size, tf.float32),
        'gt': tf.FixedLenFeature(image_size, tf.float32)}

    parsed_features = tf.parse_single_example(example_proto, features)
    mag_in = parsed_features['mag_in']
    mag_gt = parsed_features['mag_gt']

    # random crop 128x128
    crop_sz = random.randint(mag_in.shape[0]//2, mag_in.shape[0])
    mag_in = tf.random_crop(mag_in, (crop_sz, crop_sz, mag_in.shape[2]))

    return mag_in, mag_gt

The problem is the map() function is called ONLY once, so a random crop of fix size is accquired each time. How to generate a random size crop?