batch normalization - tf.layers.batch_normalization large test error

Question

I'm trying to use batch normalization. I tried to use tf.layers.batch_normalization on a simple conv net for mnist.

I get high accuracy for train step (>98%) but very low test accuracy (<50%). I tried to change momentum values (I tried 0.8,0.9,0.99,0.999) and to play to with batch sizes but it always behaves basically the same way. I train it on 20k iterations.

my code

# Input placeholders
x = tf.placeholder(tf.float32, [None, 784], name='x-input')
y_ = tf.placeholder(tf.float32, [None, 10], name='y-input')
is_training = tf.placeholder(tf.bool)

# inut layer
input_layer = tf.reshape(x, [-1, 28, 28, 1])
with tf.name_scope('conv1'):
    #Convlution #1 ([5,5] : [28x28x1]->[28x28x6])
    conv1 = tf.layers.conv2d(
        inputs=input_layer,
        filters=6,
        kernel_size=[5, 5],
        padding="same",
        activation=None
    )   

    #Batch Norm #1
    conv1_bn = tf.layers.batch_normalization(
        inputs=conv1,
        axis=-1,
        momentum=0.9,
        epsilon=0.001,
        center=True,
        scale=True,
        training = is_training,
        name='conv1_bn'
    )

    #apply relu
    conv1_bn_relu = tf.nn.relu(conv1_bn)
    #apply pool ([2,2] : [28x28x6]->[14X14X6])
    maxpool1=tf.layers.max_pooling2d(
        inputs=conv1_bn_relu,
        pool_size=[2,2],
        strides=2,
        padding="valid"
    )

with tf.name_scope('conv2'):
    #convolution #2 ([5x5] : [14x14x6]->[14x14x16]
    conv2 = tf.layers.conv2d(
        inputs=maxpool1,
        filters=16,
        kernel_size=[5, 5],
        padding="same",
        activation=None
    )   

    #Batch Norm #2
    conv2_bn = tf.layers.batch_normalization(
        inputs=conv2,
        axis=-1,
        momentum=0.999,
        epsilon=0.001,
        center=True,
        scale=True,
        training = is_training
    )

    #apply relu
    conv2_bn_relu = tf.nn.relu(conv2_bn)
    #maxpool2 ([2,2] : [14x14x16]->[7x7x16]
    maxpool2=tf.layers.max_pooling2d(
        inputs=conv2_bn_relu,
        pool_size=[2,2],
        strides=2,
        padding="valid"
    )

#fully connected 1 [7*7*16 = 784 -> 120]
maxpool2_flat=tf.reshape(maxpool2,[-1,7*7*16])
fc1 = tf.layers.dense(
    inputs=maxpool2_flat,
    units=120,
    activation=None
)

#Batch Norm #2
fc1_bn = tf.layers.batch_normalization(
    inputs=fc1,
    axis=-1,
    momentum=0.999,
    epsilon=0.001,
    center=True,
    scale=True,
    training = is_training
)
#apply reliu

fc1_bn_relu = tf.nn.relu(fc1_bn)

#fully connected 2 [120-> 84]
fc2 = tf.layers.dense(
    inputs=fc1_bn_relu,
    units=84,
    activation=None
)

#apply relu
fc2_bn_relu = tf.nn.relu(fc2)

#fully connected 3 [84->10]. Output layer with softmax
y = tf.layers.dense(
    inputs=fc2_bn_relu,
    units=10,
    activation=None
)

#loss
cross_entropy = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(labels=y_, logits=y))
tf.summary.scalar('cross entropy', cross_entropy)

correct_prediction = tf.equal(tf.argmax(y,1), tf.argmax(y_,1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
tf.summary.scalar('accuracy',accuracy)

#merge summaries and init train writer
sess = tf.Session()
merged = tf.summary.merge_all()
train_writer = tf.summary.FileWriter(log_dir + '/train' ,sess.graph)
test_writer = tf.summary.FileWriter(log_dir + '/test') 
train_step = tf.train.AdamOptimizer(1e-4).minimize(cross_entropy)
init = tf.global_variables_initializer()
sess.run(init)

with sess.as_default():
    def get_variables_values():
        variables = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES)
        values = {}
        for variable in variables:
            values[variable.name[:-2]] = sess.run(variable, feed_dict={
                x:batch[0], y_:batch[1], is_training:True
                })
        return values


    for i in range(t_iter):
        batch = mnist.train.next_batch(batch_size)
        if i%100 == 0: #test-set summary
            print('####################################')
            values = get_variables_values()
            print('moving variance is:')
            print(values["conv1_bn/moving_variance"])
            print('moving mean is:')
            print(values["conv1_bn/moving_mean"])
            print('gamma is:')
            print(values["conv1_bn/gamma/Adam"])
            print('beta is:')
            print(values["conv1_bn/beta/Adam"])
            summary, acc = sess.run([merged,accuracy], feed_dict={
                x:mnist.test.images, y_:mnist.test.labels, is_training:False

            })

        else:
            summary, _ = sess.run([merged,train_step], feed_dict={
                x:batch[0], y_:batch[1], is_training:True
            })
            if i%10 == 0:
                train_writer.add_summary(summary,i)

I think the problem is that that the moving_mean/var is not being updated. I print the moving_mean/var during the run and I get: moving variance is: [ 1. 1. 1. 1. 1. 1.] moving mean is: [ 0. 0. 0. 0. 0. 0.] gamma is: [-0.00055969 0.00164391 0.00163301 -0.00206227 -0.00011434 -0.00070161] beta is: [-0.00232835 -0.00040769 0.00114277 -0.0025414 -0.00049697 0.00221556]

Anyone has any idea what i'm doing wrong?

Answer 1

The operations which tf.layers.batch_normalization adds to update mean and variance don't automatically get added as dependencies of the train operation - so if you don't do anything extra, they never get run. (Unfortunately, the documentation doesn't currently mention this. I'm opening an issue about it.)

Luckily, the update operations are easy to get at, since they're added to the tf.GraphKeys.UPDATE_OPS collection. Then you can either run the extra operations manually:

extra_update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
sess.run([train_op, extra_update_ops], ...)

Or add them as dependencies of your training operation, and then just run your training operation as normal:

extra_update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
with tf.control_dependencies(extra_update_ops):
    train_op = optimizer.minimize(loss)
...
sess.run([train_op], ...)