BatchNormalization gives incorrect output with masked inputs > 3 dimensions

[drasmuss]

The mean/variance calculations are incorrect, which means the inputs are not normalized correctly. E.g.

import keras

x = keras.ops.ones((1, 2, 3, 4))
x._keras_mask = keras.ops.ones((1, 2, 1))

y = keras.layers.BatchNormalization()(x, training=True)

print(keras.ops.mean(y, axis=-1))

gives output

tf.Tensor([-0.57732624 -0.57732624 -0.57732624 -0.57732624], shape=(4,), dtype=float32)

instead of the correct normalized output ([0, 0, 0, 0]).

The basic issue is that this calculation is incorrect:

keras/keras/src/layers/normalization/batch_normalization.py

Lines 310 to 314 in efaaf85

	sum_of_weights = ops.sum(
	mask_weights_broadcasted,
	self._reduction_axes,
	keepdims=True,
	)

because it doesn't account for the broadcasting (i.e. it gives a value of 2 in the above example, when it should be 2 * 3 * 4).

See #19818 for more discussion/background.

[Grvzard]

I think a better workaround is to validate the shape of the mask in keras.

[drasmuss]

The shape of the mask is correct in this example (according to #19818 (comment)), so validation wouldn't help in this case.

[Grvzard]

because it doesn't account for the broadcasting (i.e. it gives a value of 2 in the above example, when it should be 2 * 3 * 4).

broadcasting from (2,) to (2, 3, 4) makes sense here, but elsewhere, "broadcasting" may starts with the rightmost dimension, i.e. broadcast (4,) to (2, 3, 4)

[google-ml-butler]

Are you satisfied with the resolution of your issue?
Yes
No