python - Does batch normalization in tensorflow use running averages during training?

Question

I am using a tensorflow neural net to figure out how batch normalization works and replicate it in my own library. I've run into this strange issue:

When you initialize a neural net layer, all biases (or in case of batchnorm - betas) are set to 0, so the layer should just multiply the input values by the weights, and that's about it. Now, from what I understand about batchnorm, during training it calculates the means and the variances for the layer inputs based on the minibatch it is being fed, and then does this to the input: output = (input - mean) / sqrt(variance + eps).

So, if all the input values of your minibatch are the same, then during training batchnorm will subtract the mean (equal to each value) from the input value, so the net should output 0, regardless of input, right?

And, it doesn't. In fact, it looks like all the means during calculation are 0, and the variances are 1 as if it is using the running averages of those values. So, either I don't understand how batchnorm works or batchnorm is just used incorrectly. Here is how it is initialized in the code I'm using:

    layer= tflearn.fully_connected(layer, 10, weights_init=w_init)
    layer= tflearn.layers.normalization.batch_normalization(layer)
    layer= tflearn.activations.leaky_relu(layer)

The other option is that it is used incorrectly during training, but I would like to eliminate the other possible explanations first.

标签： pythontensorflowneural-networktflearnbatch-normalization

Answer 1

The TensorFlow batch norm implementation has some update ops that are not included in the training op's dependencies by default. You have to add the dependencies explicitly. Quoting the docs:

[W]hen training, the moving_mean and moving_variance need to be updated. By default the update ops are placed in tf.GraphKeys.UPDATE_OPS, so they need to be added as a dependency to the train_op. Also, be sure to add any batch_normalization ops before getting the update_ops collection. Otherwise, update_ops will be empty, and training/inference will not work properly. For example:

  x_norm = tf.layers.batch_normalization(x, training=training)

  # ...

  update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
  with tf.control_dependencies(update_ops):
    train_op = optimizer.minimize(loss)