YOLOV8Backbone inconsistent output tensor shapes on Torch backend

[DLumi]

Current Behavior:

Variable output shapes are not handled correctly when build() is called by a Sequential model (Torch backend only)

Expected Behavior:

Variable shapes remain variable

Steps To Reproduce:

This shape inconsistency is only for Torch backend. It works fine on TF. (not tested with JAX, though)

0. Define an input shape with at least one variable dimension, say (224, None, 3)
1. Make a subclass keras.Model called, say, MyModel, instantiate YOLOV8Backbone, and add it as an attribute of this subclass
2. Make a subclass keras.Sequential called, say, AnotherModel , instantiate MyModel and self.add it as a layer of this subclass
3. Instantiate AnotherModel and check its output shape. Based off the input shape we defined, it should be [None, 14, None, 512], but it's [None, 14, 6, 512] instead

Colab notebook

Version:

keras_cv==0.9.0
keras==3.4.0
torch==2.3.0+cu121
tensorflow==2.15.0

Anything else:

N/A