Prevent setting attributes on KerasTensor

[james77777778]

This PR should improve consistency when using KerasTensor

[codecov-commenter]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 78.98%. Comparing base (558d38c) to head (93b191f).
Report is 10 commits behind head on master.

Additional details and impacted files

@@            Coverage Diff             @@
##           master   #19920      +/-   ##
==========================================
- Coverage   79.02%   78.98%   -0.05%     
==========================================
  Files         499      499              
  Lines       46436    46513      +77     
  Branches     8548     8561      +13     
==========================================
+ Hits        36695    36737      +42     
- Misses       8015     8044      +29     
- Partials     1726     1732       +6     

Flag	Coverage Δ
keras	78.84% <100.00%> (-0.05%)	⬇️
keras-jax	62.43% <100.00%> (+0.01%)	⬆️
keras-numpy	57.33% <100.00%> (+0.11%)	⬆️
keras-tensorflow	63.62% <100.00%> (-0.03%)	⬇️
keras-torch	62.39% <100.00%> (+0.01%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[fchollet]

Thanks for the PR. In what cases would we set these attributes? Shouldn't these properties be immutable, like they are for a concrete tensor?

[james77777778]

Thanks for the PR. In what cases would we set these attributes? Shouldn't these properties be immutable, like they are for a concrete tensor?

Sometimes, it is convenient to change them inplace in compute_output_spec:

keras/keras/src/ops/core.py

Lines 105 to 119 in d318fc7

	def compute_output_spec(self, f, init, xs, length):
	if xs is None:
	n = int(length)
	x = None
	else:
	n = (
	int(length)
	if length is not None
	else tree.flatten(xs)[0].shape[0]
	)
	x = xs[0]
	carry_spec, y_spec = backend.compute_output_spec(f, init, x)
	y_spec.shape = (n,) + y_spec.shape
	return carry_spec, y_spec

keras/keras/src/dtype_policies/dtype_policy.py

Lines 162 to 165 in d318fc7

	elif backend.is_keras_tensor(x):
	if self._should_cast(x, autocast, dtype):
	x.dtype = dtype
	return x

Or should these be considered inappropriate?

[fchollet]

Or should these be considered inappropriate?

I would say it's better to create a new KerasTensor instance for these (the existing instance might be in use elsewhere). I'd expect these properties to be immutable.

[james77777778]

Got it. Let me try to fix them.

[fchollet]

Thanks for the update!

[james77777778]

I've added a clearer error msg for the setters of KerasTensor.
This should guide anyone who tries to modify these attributes to create a new instance.

[fchollet]

LGTM, thank you!