Subclassing API

[RaphaelMeudec]

The current version of tf-explain does not support subclassing API because of the way subgraphs are created.

See: https://stackoverflow.com/questions/52826134/keras-model-subclassing-examples

[dtransposed]

Hi! Thanks of all thank you for this great library! Is this functionality high on your priority list?

[RaphaelMeudec]

Hi! I'm on holiday for a few days, but plan to tackle this as soon as I get back. There's an open issue on Tensorflow to help fix this

[dtransposed]

Excellent, thank you!

[RaphaelMeudec]

@dtransposed looks like getting support for subclassing API is going to be harder than expected: https://github.com/tensorflow/tensorflow/issues/30955#issuecomment-520291616

[dtransposed]

Yes, I have tried to hack your code to use it with my models built using subclassing API and failed. It's a pity there is no way to "translate" subclassing model into functional one so far...

[leeor-langer]

@RaphaelMeudec is there any update on this issue? A workaround is also fine ...

[RaphaelMeudec]

@leeor-langer No fix as of today. Best option is to define subclass layer with strict minimum operations and use it in a sequential/functional model, instead of defining the whole model in a subclass layer.

For example:

class MySpecialPadding(tf.keras.Layer):
  def call(self, inputs, *args, **kwargs):
     x = tf.pad(inputs, my_padding)
     return x

model = tf.keras.Model([
  tf.keras.layers.Conv2D(..),
  MySpecialPadding(),
])

[ramshi236]

is there an update on the problem ?