Curious to know whether others have same experience as me when using Keras experimental preprocessing layers. I am finding the ‘adapt’ method take an unusually long time to sweep the entire dataset and calculate state information (eg mean and variance).
I create my dataset like so:
ds = tf.data.experimental.CsvDataset(
filenames = data_files, …
Then I simply create the preprocessing layer (outside the model)
layer_norm = tensorflow.keras.layers.experimental.preprocessing.Normalization(axis=-1)
layer_norm.adapt(ds)
The adapt process takes well over a 90 seconds working on a 25 MB file.
Admittedly, my machine isnt high spec, and i and pulling the data files from a remote server. But even still, this feels very slow.