Correct dim parameter in tf.nn.l2_normalize(x, dim, epsilon=1e-12, name=None) for RGB picture

Question

I want to use tensorflows

tf.nn.l2_normalize(x, dim, epsilon=1e-12, name=None)

method to normalize each channel seperately in a (32x32x3) RGB image.

How do I have to set the

dim

parameter to achieve that?

Answer 1

Quoting the official documentation, tf.nn.l2_normalize,

For x with more dimensions, independently normalizes each 1-D slice along dimension dim.

Since your channel is the third dimension, you can pass in dim=2 (since dimensions start from 0).

Answer 2

You should set dim = [0, 1]. So each channel [32, 32] would be normalized as [32*32], a 1-D tensor.