Batch normalization in tensorflow: variables and performance

Question

I would like to add conditional operations on the variables of a batch normalization layer. Specifically, train in float, then quantize in a fine-tuning secondary training phase. For this, I want to add a tf.cond operation on the variables (scale, shift and exp moving averages of mean and var).

I replaced the tf.layers.batch_normalization with a batchnorm layer I wrote (see below).

This function works perfectly (i.e. I get the same metrics with both functions), and I can add whatever pipeline to the variables (before the batchnorm operation). The problem is that the performance (runtime) dropped dramatically (i.e. there's a x2 factor by simply replacing the layers.batchnorm with my own function, as written below).

def batchnorm(self, x, name, epsilon=0.001, decay=0.99):
    epsilon = tf.to_float(epsilon)
    decay = tf.to_float(decay)
    with tf.variable_scope(name):
        shape = x.get_shape().as_list()
        channels_num = shape[3]
        # scale factor
        gamma = tf.get_variable("gamma", shape=[channels_num], initializer=tf.constant_initializer(1.0), trainable=True)
        # shift value
        beta = tf.get_variable("beta", shape=[channels_num], initializer=tf.constant_initializer(0.0), trainable=True)
        moving_mean = tf.get_variable("moving_mean", channels_num, initializer=tf.constant_initializer(0.0), trainable=False)
        moving_var = tf.get_variable("moving_var", channels_num, initializer=tf.constant_initializer(1.0), trainable=False)
        batch_mean, batch_var = tf.nn.moments(x, axes=[0, 1, 2]) # per channel

        update_mean = moving_mean.assign((decay * moving_mean) + ((1. - decay) * batch_mean))
        update_var = moving_var.assign((decay * moving_var) + ((1. - decay) * batch_var))

        tf.add_to_collection(tf.GraphKeys.UPDATE_OPS, update_mean)
        tf.add_to_collection(tf.GraphKeys.UPDATE_OPS, update_var)

        bn_mean = tf.cond(self.is_training, lambda: tf.identity(batch_mean), lambda: tf.identity(moving_mean))
        bn_var = tf.cond(self.is_training, lambda: tf.identity(batch_var), lambda: tf.identity(moving_var))

        with tf.variable_scope(name + "_batchnorm_op"):
            inv = tf.math.rsqrt(bn_var + epsilon)
            inv *= gamma
            output = ((x*inv) - (bn_mean*inv)) + beta

    return output

I would appreciate help in any of the following questions:

• Any ideas on how to improve the performance (reduce runtime) of my solution?
• Is it possible to add my own operators to the variables pipeline of layers.batchnorm before the batchnorm operation?
• Any other solution to the same problem?

Answer 1

tf.nn.fused_batch_norm is optimized and did the trick.

I had to create two subgraphs, one per mode, since fused_batch_norm's interface does not take a conditional training/test mode (is_training is bool and not a tensor, so it's graph is not conditional). I added the condition after (see below). However, even with the two subgraphs, this has about the same runtime of tf.layers.batch_normalization.

Here's the final solution (I'd still appreciate any comment or advice for improvements):

def batchnorm(self, x, name, epsilon=0.001, decay=0.99):
    with tf.variable_scope(name):
        shape = x.get_shape().as_list()
        channels_num = shape[3]
        # scale factor
        gamma = tf.get_variable("gamma", shape=[channels_num], initializer=tf.constant_initializer(1.0), trainable=True)
        # shift value
        beta = tf.get_variable("beta", shape=[channels_num], initializer=tf.constant_initializer(0.0), trainable=True)
        moving_mean = tf.get_variable("moving_mean", channels_num, initializer=tf.constant_initializer(0.0), trainable=False)
        moving_var = tf.get_variable("moving_var", channels_num, initializer=tf.constant_initializer(1.0), trainable=False)

        (output_train, batch_mean, batch_var) = tf.nn.fused_batch_norm(x,
                                                                 gamma,
                                                                 beta,  # pylint: disable=invalid-name
                                                                 mean=None,
                                                                 variance=None,
                                                                 epsilon=epsilon,
                                                                 data_format="NHWC",
                                                                 is_training=True,
                                                                 name="_batchnorm_op")
        (output_test, _, _) = tf.nn.fused_batch_norm(x,
                                                     gamma,
                                                     beta,  # pylint: disable=invalid-name
                                                     mean=moving_mean,
                                                     variance=moving_var,
                                                     epsilon=epsilon,
                                                     data_format="NHWC",
                                                     is_training=False,
                                                     name="_batchnorm_op")

        output = tf.cond(self.is_training, lambda: tf.identity(output_train), lambda: tf.identity(output_test))

        update_mean = moving_mean.assign((decay * moving_mean) + ((1. - decay) * batch_mean))
        update_var = moving_var.assign((decay * moving_var) + ((1. - decay) * batch_var))
        tf.add_to_collection(tf.GraphKeys.UPDATE_OPS, update_mean)
        tf.add_to_collection(tf.GraphKeys.UPDATE_OPS, update_var)

    return output