如何在tensorflow中正确使用tf.layers.batch_normalization()？

Question

我被tensorflow中的tf.layers.batch_normalization搞糊涂了。

我的代码如下：

def my_net(x, num_classes, phase_train, scope):
    x = tf.layers.conv2d(...)
    x = tf.layers.batch_normalization(x, training=phase_train)
    x = tf.nn.relu(x) 
    x = tf.layers.max_pooling2d(...)

    # some other staffs
    ...

    # return 
    return x

def train():
    phase_train = tf.placeholder(tf.bool, name='phase_train')
    image_node = tf.placeholder(tf.float32, shape=[batch_size, HEIGHT, WIDTH, 3])
    images, labels = data_loader(train_set)
    val_images, val_labels = data_loader(validation_set)
    prediction_op = my_net(image_node, num_classes=2,phase_train=phase_train, scope='Branch1')

    loss_op = loss(...)
    # some other staffs
    optimizer = tf.train.AdamOptimizer(base_learning_rate)
    update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
    with tf.control_dependencies(update_ops):
        train_op = optimizer.minimize(loss=total_loss, global_step=global_step)
    sess = ...
    coord = ...
    while not coord.should_stop():
        image_batch, label_batch = sess.run([images, labels])
        _,loss_value= sess.run([train_op,loss_op], feed_dict={image_node:image_batch,label_node:label_batch,phase_train:True})

        step = step+1

        if step==NUM_TRAIN_SAMPLES:
            for _ in range(NUM_VAL_SAMPLES/batch_size):
                image_batch, label_batch = sess.run([val_images, val_labels])
                prediction_batch = sess.run([prediction_op], feed_dict={image_node:image_batch,label_node:label_batch,phase_train:False})
            val_accuracy = compute_accuracy(...)


def test():
    phase_train = tf.placeholder(tf.bool, name='phase_train')
    image_node = tf.placeholder(tf.float32, shape=[batch_size, HEIGHT, WIDTH, 3])
    test_images, test_labels = data_loader(test_set)
    prediction_op = my_net(image_node, num_classes=2,phase_train=phase_train, scope='Branch1')

    # some staff to load the trained weights to the graph
    saver.restore(...)

    for _ in range(NUM_TEST_SAMPLES/batch_size):
        image_batch, label_batch = sess.run([test_images, test_labels])
        prediction_batch = sess.run([prediction_op], feed_dict={image_node:image_batch,label_node:label_batch,phase_train:False})
    test_accuracy = compute_accuracy(...)

训练似乎工作得很好，val_accuracy也是合理的(比如0.70)。问题是:当我尝试使用经过训练的模型进行测试(即test函数)时，如果phase_train设置为False，则test_accuracy非常低(比如0.000270)，但是当phase_train设置为True时，test_accuracy看起来是正确的(比如0.69)。

据我所知，测试阶段的phase_train应该是False，对吧？我不确定问题出在哪里。我误解了批处理规范化吗？

Answer 1

这可能是你的代码中的一些bug，或者仅仅是过度匹配。如果你在训练数据上进行评估，准确率是否和训练时一样高？如果问题是批量规范，那么没有训练的训练错误会比训练模式下的训练错误更高。如果问题是过度拟合，那么批处理规范可能不会导致这种情况，根本原因是其他地方。