[Guides] Pre-processing of images for Xception model about keras-io HOT 6 CLOSED

We need preprocessing to be part of the model, as a best practice.

I will look into adding a new offset argument to Rescaling to support this use case.

For the time being I would recommend using a Normalization layer (from layers.experimental.preprocessing). It can do both scaling and offsetting. Just set the weights correctly.

from keras-io.

Let's move the discussion to the PR.

from keras-io.

Yes, that seems like an issue. We should fix it.

from keras-io.

I'll be working on this and hope to get back to you with a PR soon.
Thanks for addressing the issue.

from keras-io.

It appears to me that the Rescaling layer used in the code can only support multiplicative op(s). Instead of using 1/255., I can use 1/127.5 but that will scale the data in range (0, 2) which is undesired.

In order to attain required normalisation range of (-1, 1), the inputs should be scaled as: x /= 127.5; x -= 1 or similar.

Should we replace the Rescaling layer with a new preprocess tf.function or use keras_applications.xception.preprocess_input that does exactly the same operation as above and pass it to ds.map?

@tf.function
def preprocess(x):
     x /= 127.5
     x -= 1
     return x

ds = ds.map(lambda x, y: (preprocess(x), y), AUTOTUNE)

or

preprocess = tf.keras.applications.xception.preprocess_input
ds = ds.map(lambda x, y: (preprocess(x), y), AUTOTUNE)

Although, it wouldn't be a good idea to let go away the Rescaling layer from a documentation example as it'd then indirectly discourage users from using the new Preprocessing Layers; one workaround would be to use the Rescaling layer combined with a Lambda layer to scale inputs appropriately.

Something like this:

# x has range (0, 255)
x = tf.keras.layers.experimental.preprocessing.Rescaling(1 / 127.5)(x) # x has range (0, 2)
x = tf.keras.layers.Lambda(lambda inp: inp - 1.)(x) # x has range (-1, 1)

from keras-io.

Thanks @fchollet.

I'll make the code changes to the script, use the Normalization layer and submit a PR as of now. Also, hope to get back to you soon regarding the offset argument for Rescaling as well.

/cc: @tanzhenyu

We need preprocessing to be part of the model, as a best practice.

I will look into adding a new offset argument to Rescaling to support this use case.

We can have a discussion regarding the same. Thanks.

from keras-io.