Specifying shape of normal sampling with partially empty keras tensor about keras HOT 3 CLOSED

That worked because tf.random.normal can operate over symbolic TF tensors. To write equivalent code, just make a Normal() Keras layer -- since layers can operate over symbolic tensors too. But in practice I think you're going to want to put more logic in your layer than a single RNG call, for good factoring.

from keras.

You cannot create an eager random tensor with an undefined shape (here, (None, 1024, 12)). This has never worked because it is conceptually impossible. I am not sure what setup you are referring to when you say, "this used to work", but in any case that's not it.

You can simply use a layer to do what you describe. Like this.

import keras


class MyLayer(keras.Layer):
    def __init__(self, **kwargs):
        super().__init__()
        self.seed_generator = keras.random.SeedGenerator(seed=1337)

    def call(self, z_mean):
        eps = keras.random.normal(shape=keras.ops.shape(z_mean), seed=self.seed_generator)
        return z_mean + eps


z_mean = keras.Input(shape=(1024, 12))
z_mean_plus_eps = MyLayer()(z_mean)

from keras.

Thank you for your reply and the proposed solution.
I realised that perhaps I simplified my question too much.
When I said that

Sampling a partially None-shaped tensor worked fine in Keras 2

I was refering to the following "old" code:
import tensorflow as tf # ==2.10.1
Z_mean = tf.keras.layers.Input(shape=(1024, 12))
eps = tf.random.normal(shape=tf.shape(Z_mean))

If I didn't make a mistake, the keras 3 equivalent should be:
import keras # ==3.3.3
Z_mean = keras.Input(shape=(1024, 12))
eps = keras.random.normal(shape=keras.ops.shape(Z_mean))
which, as you pointed out, does not work.
Why so? What would be the working equivalent of the "old" code?

from keras.