What will happen with layers which are in tfa, but not in other keras frameworks and which do not work with Keras 3 (I'm intrested in WeightNormalization layer)

[Kurdakov]

trafficstars

System information

• OS Platform and Distribution (e.g., Linux Ubuntu 20.04): Linux Ubuntu 22.04
• TensorFlow version and how it was installed (source or binary): Tensorflow 2.6.1 binary installation
• TensorFlow-Addons version and how it was installed (source or binary): master source
• Python version: 3.10
• Is GPU used? (yes/no): no

Describe the bug while master branch has fixed imports for Keras 3 class WeightNormalization(tf.keras.layers.Wrapper) won't work with Keras 3

Code to reproduce the issue

import tensorflow as tf
import tensorflow_addons as tfa

from tensorflow.keras.layers import Conv1D, Embedding, MaxPooling1D, Dense, Flatten
from tensorflow.keras.models import Sequential
from tensorflow.keras.datasets import imdb
from tensorflow.keras.preprocessing import sequence
from tensorflow.keras.optimizers import Adam

max_words = 800

(Xtrain, ytrain), (Xtest, ytest) = imdb.load_data(num_words=1000)

Xtrain = sequence.pad_sequences(Xtrain, maxlen=max_words)
Xtest = sequence.pad_sequences(Xtest, maxlen=max_words)        

model = Sequential()
model.add(Embedding(1000, 500, input_length=max_words))
model.add(tfa.layers.WeightNormalization(Conv1D(64, 3, activation='relu')))
model.add(MaxPooling1D(2,2))
model.add(tfa.layers.WeightNormalization(Conv1D(32, 3, activation='relu')))
model.add(MaxPooling1D(2,2))
model.add(Flatten())
model.add(Dense(10, activation='relu'))
model.add(Dense(1, activation='sigmoid'))
model.summary()

model.compile(optimizer=Adam(.0001), metrics=['accuracy'], loss='binary_crossentropy')
model.fit(Xtrain, ytrain, validation_split=.2, epochs=10)

problems:

def compute_output_shape(self, input_shape):

uses as_list(), Keras 3 does not support it, removal of as_list helps.

other problems which I failed to resolve are in creation of self._naked_clone_layer

the problem is essentially is that class WeightNormalization is absent in other keras frameworks, but it does not work in tfa with Keras 3 either.

I understand that tfa is near end of support (and already almost an year in minimal support mode), but then the question is - what to use in place of WeightNormalization layer in Keras 3?

[Kurdakov]

what I tried: instead of serialization/deserialization

               layer_config = self.layer.get_config()
               layer_config["trainable"] = False
               self._naked_clone_layer = type(self.layer).from_config(layer_config)

that avoids build issues, but then weight could not be set

[Kurdakov]

with functional Model while weights could be set (considering fix to serialization in previous comment so that _naked_clone_layer could be built and also removal of as_list in computed shape )

import tensorflow as tf
import tensorflow_addons as tfa

from tensorflow.keras.layers import Conv1D, Embedding, MaxPooling1D, Dense, Flatten,Input
from tensorflow.keras.models import Model
from tensorflow.keras.datasets import imdb
from tensorflow.keras.preprocessing import sequence
from tensorflow.keras.optimizers import Adam

max_words = 800

(Xtrain, ytrain), (Xtest, ytest) = imdb.load_data(num_words=1000)

Xtrain = sequence.pad_sequences(Xtrain, maxlen=max_words)
Xtest = sequence.pad_sequences(Xtest, maxlen=max_words)  

input = Input(shape=(800,))
x = Embedding(1000,500)(input)
x = tfa.layers.WeightNormalization(Conv1D(64, 3, activation='relu'))(x)
x = MaxPooling1D(2,2)(x)
x = tfa.layers.WeightNormalization(Conv1D(32, 3, activation='relu'))(x)
x = MaxPooling1D(2,2)(x)
x = Flatten()(x)
x = Dense(10, activation='relu')(x)
out = Dense(1, activation='sigmoid')(x)
model = Model(inputs=input, outputs=[out])
model.summary()

model.compile(optimizer=Adam(.0001), metrics=['accuracy'], loss='binary_crossentropy')
model.fit(Xtrain, ytrain, validation_split=.2, epochs=10)

but there is exception Can not convert a NoneType into a Tensor or Operation which happens in

        def _update_weights():
            # Ensure we read `self.g` after _update_weights.
            with tf.control_dependencies(self._initialize_weights(inputs)):

[Kurdakov]

finally changing

with tf.control_dependencies(self._initialize_weights(inputs)):

to with tf.control_dependencies([self._initialize_weights(inputs)]):

I see tensorflow running