Batch Normalization in tf.keras does not calculate average mean and average variance

Question

A similar unanswered question was asked here. I am testing one deep reinforcement learning algorithm which uses keras backend in tensorflow. I am not very familiar with tf.keras, nevertheless would like to add batch normalization layers. Therefore, I am trying to use tf.keras.layers.BatchNormalization(), but it does not update average means and variances because update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS) is empty.

Using the regular tf.layers.batch_normalization seem to work fine. However, because the complete algorithm is somewhat complicated, I need to find a way to use tf.keras.

A standard tf layer batch_normed = tf.layers.batch_normalization(hidden, training=True) updates the averages since update_ops is not empty:

[
    <tf.Operation 'batch_normalization/AssignMovingAvg' type=AssignSub>, 
    <tf.Operation 'batch_normalization/AssignMovingAvg_1' type=AssignSub>, 
    <tf.Operation 'batch_normalization_1/AssignMovingAvg' type=AssignSub>, 
    <tf.Operation 'batch_normalization_1/AssignMovingAvg_1' type=AssignSub>
]

Minimal example that does not work:

import tensorflow as tf
import numpy as np

tf.reset_default_graph()
graph = tf.get_default_graph()
tf.keras.backend.set_learning_phase(True)

input_shapes = [(3, )]
hidden_layer_sizes = [16, 16]

inputs = [
    tf.keras.layers.Input(shape=input_shape)
    for input_shape in input_shapes
]

concatenated = tf.keras.layers.Lambda(
    lambda x: tf.concat(x, axis=-1)
)(inputs)

out = concatenated
for units in hidden_layer_sizes:      
    hidden = tf.keras.layers.Dense(
    units, activation=None
    )(out)
    batch_normed = tf.keras.layers.BatchNormalization()(hidden, training=True)
    #batch_normed = tf.layers.batch_normalization(hidden, training=True)
    out = tf.keras.layers.Activation('relu')(batch_normed)

out = tf.keras.layers.Dense(
    units=1, activation='linear'
)(out)


data = np.random.rand(100,3)
with tf.Session(graph=graph) as sess:
    sess.run(tf.global_variables_initializer())

    for i in range(10):

    update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)

    sess.run(update_ops,  {inputs[0]: data})
    sess.run(out, {inputs[0]: data})

    variables = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES,
                          scope='batch_normalization')
    bn_gamma, bn_beta, bn_moving_mean, bn_moving_variance = [], [], [], []
    for variable in variables:
        val = sess.run(variable)
        nv = np.linalg.norm(val)
        if 'gamma' in variable.name:
            bn_gamma.append(nv)
        if 'beta' in variable.name:
            bn_beta.append(nv)
        if 'moving_mean' in variable.name:
            bn_moving_mean.append(nv)
        if 'moving_variance' in variable.name:
            bn_moving_variance.append(nv)

        diagnostics = {
            'bn_Q_gamma': np.mean(bn_gamma),
            'bn_Q_beta': np.mean(bn_beta),
            'bn_Q_moving_mean': np.mean(bn_moving_mean),
            'bn_Q_moving_variance': np.mean(bn_moving_variance),
        }

    print(diagnostics)

The output is the following (you can see moving_mean and moving_variance not changing):

{'bn_Q_gamma': 4.0, 'bn_Q_beta': 0.0, 'bn_Q_moving_mean': 0.0, 'bn_Q_moving_variance': 4.0}
{'bn_Q_gamma': 4.0, 'bn_Q_beta': 0.0, 'bn_Q_moving_mean': 0.0, 'bn_Q_moving_variance': 4.0}
{'bn_Q_gamma': 4.0, 'bn_Q_beta': 0.0, 'bn_Q_moving_mean': 0.0, 'bn_Q_moving_variance': 4.0}
{'bn_Q_gamma': 4.0, 'bn_Q_beta': 0.0, 'bn_Q_moving_mean': 0.0, 'bn_Q_moving_variance': 4.0}
{'bn_Q_gamma': 4.0, 'bn_Q_beta': 0.0, 'bn_Q_moving_mean': 0.0, 'bn_Q_moving_variance': 4.0}
{'bn_Q_gamma': 4.0, 'bn_Q_beta': 0.0, 'bn_Q_moving_mean': 0.0, 'bn_Q_moving_variance': 4.0}
{'bn_Q_gamma': 4.0, 'bn_Q_beta': 0.0, 'bn_Q_moving_mean': 0.0, 'bn_Q_moving_variance': 4.0}
{'bn_Q_gamma': 4.0, 'bn_Q_beta': 0.0, 'bn_Q_moving_mean': 0.0, 'bn_Q_moving_variance': 4.0}
{'bn_Q_gamma': 4.0, 'bn_Q_beta': 0.0, 'bn_Q_moving_mean': 0.0, 'bn_Q_moving_variance': 4.0}
{'bn_Q_gamma': 4.0, 'bn_Q_beta': 0.0, 'bn_Q_moving_mean': 0.0, 'bn_Q_moving_variance': 4.0}

While the expected output is something like the following (comment the line with batch_normed calculus using tf.keras and uncomment the one below it):

{'bn_Q_gamma': 4.0, 'bn_Q_beta': 0.0, 'bn_Q_moving_mean': 0.0148749575, 'bn_Q_moving_variance': 3.966927}
{'bn_Q_gamma': 4.0, 'bn_Q_beta': 0.0, 'bn_Q_moving_mean': 0.029601166, 'bn_Q_moving_variance': 3.934192}
{'bn_Q_gamma': 4.0, 'bn_Q_beta': 0.0, 'bn_Q_moving_mean': 0.04418011, 'bn_Q_moving_variance': 3.9017918}
{'bn_Q_gamma': 4.0, 'bn_Q_beta': 0.0, 'bn_Q_moving_mean': 0.05861327, 'bn_Q_moving_variance': 3.8697228}
{'bn_Q_gamma': 4.0, 'bn_Q_beta': 0.0, 'bn_Q_moving_mean': 0.0729021, 'bn_Q_moving_variance': 3.8379822}
{'bn_Q_gamma': 4.0, 'bn_Q_beta': 0.0, 'bn_Q_moving_mean': 0.08704803, 'bn_Q_moving_variance': 3.8065662}
{'bn_Q_gamma': 4.0, 'bn_Q_beta': 0.0, 'bn_Q_moving_mean': 0.10105251, 'bn_Q_moving_variance': 3.7754717}
{'bn_Q_gamma': 4.0, 'bn_Q_beta': 0.0, 'bn_Q_moving_mean': 0.11491694, 'bn_Q_moving_variance': 3.7446957}
{'bn_Q_gamma': 4.0, 'bn_Q_beta': 0.0, 'bn_Q_moving_mean': 0.12864274, 'bn_Q_moving_variance': 3.7142346}
{'bn_Q_gamma': 4.0, 'bn_Q_beta': 0.0, 'bn_Q_moving_mean': 0.14223127, 'bn_Q_moving_variance': 3.6840856}

Note

There is still something fishy even with tf.layers.batch_normalization. The standard tf approach of tf.control_dependencies:

    with tf.control_dependencies(update_ops):
        sess.run(out, {inputs[0]: data})

which I place instead of the following two lines in the code above:

    sess.run(update_ops,  {inputs[0]: data})
    sess.run(out, {inputs[0]: data})

produces bn_Q_moving_mean = 0.0 and bn_Q_moving_variance = 4.0

Answer 1

This is because tf.keras.layers.BatchNormalization inherits from tf.keras.layers.Layer. Keras API handle update ops as part of its fit and evaluate loops. This in turn means that it won't update tf.GraphKeys.UPDATE_OPS collection without it.

So in order to make it work, you need to update it manually

hidden = tf.keras.layers.Dense(units, activation=None)(out)
batch_normed = tf.keras.layers.BatchNormalization(trainable=True) 
layer = batch_normed(hidden)

This creates separate class instance

tf.add_to_collection(tf.GraphKeys.UPDATE_OPS, batch_normed.updates)

And this updates needed collection. Also take a look https://github.com/tensorflow/tensorflow/issues/25525

Answer 2

tf.add_to_collection(tf.GraphKeys.UPDATE_OPS, bn1.updates[0])
tf.add_to_collection(tf.GraphKeys.UPDATE_OPS, bn1.updates[1])
updates_op = tf.get_collection(tf.GraphKeys.UPDATE_OPS)

this can solve

tf.control_dependencies(update_ops)

error problem.

if use

tf.add_to_collection(tf.GraphKeys.UPDATE_OPS, batch_normed.updates)

the return of

tf.get_collection(tf.GraphKeys.UPDATE_OPS)

is a list in list just like [[something]]

and use

tf.add_to_collection(tf.GraphKeys.UPDATE_OPS, bn1.updates[0])
tf.add_to_collection(tf.GraphKeys.UPDATE_OPS, bn1.updates[1])
updates_op = tf.get_collection(tf.GraphKeys.UPDATE_OPS)

the return of

tf.get_collection(tf.GraphKeys.UPDATE_OPS)

is [something1,something2,...]

i thinks this is the solution.

but the out put is different,and i don't know which is true.