Preprocessing for mean standardization.

[akirasosa]

I use Keras and use featurewise_center and featurewise_std_normalization. Then, I have mean and std for each channels.

But coreml provides only redBias, greenBias, blueBias and imageScale, so that I can not scale for each channels respectively.

Could you provide options something like redScale, greenScale and blueScale?

[aseemw]

@akirasosa To add this functionality a change needs to the made in the CoreML proto and Framework.
For now, I have a few suggestions for you to do this:

• Do the processing in your app on the MLMutltiArray before feeding it to mlmodel
• The "scale" layer in CoreML can do this kind of processing. You can modify the .mlmodel file and add the scale layer in the beginning with the appropriate channel scales.
• You can also modify the Keras model before conversion. Keras does not have a scale layer but you can use the Batch normalization layer and set mean to 0, variance to 1, beta to 0 and gamma (a length 3 vector) to the channel scales.

[akirasosa]

@aseemw
Thanks! The first option is handy for me.

To add this functionality a change needs to the made in the CoreML proto and Framework.

Is there a public github project for it? I would like to watch it.

[yonatanbitton]

Hey, is there a documented solution for it? I'm in the same situation. Thanks!