Type of model.input is KerasTensor in 3.9 and list in 3.10

[mmtrebuchet]

This seems to be a breaking change and I don't see it documented in the release notes.

If you save a model and then load it, the type of model.input was KerasTensor in 3.9.2 (and as far back as I can remember), but it's now list in 3.10.0. If you don't save and then re-load the model, the input type is KerasTensor, so I think this qualifies as a bug.

Here is a minimal example:

#!/usr/bin/env python3

import keras as k
import keras.layers as kl
from keras import models as km
def runTest():
    il = kl.Input((1024, 4))
    cl = kl.Conv1D(filters=5, kernel_size=10, padding="valid", activation="relu")(il)
    m = km.Model(inputs=il, outputs=[cl])
    print(m.input.shape)
    m.save("test.keras")

def runLoad():
    m = km.load_model("test.keras")
    print(type(m.input))
    print(m.input.shape)

runTest()
runLoad()

Under Keras 3.9.2, I get the following:

(None, 1024, 4)
<class 'keras.src.backend.common.keras_tensor.KerasTensor'>
(None, 1024, 4)

Whereas with 3.10.0, I get:

(None, 1024, 4)
<class 'list'>
Traceback (most recent call last):
  File "/dev/shm/./testModel.py", line 22, in <module>
    runLoad()
  File "/dev/shm/./testModel.py", line 16, in runLoad
    print(m.input.shape)
          ^^^^^^^^^^^^^
AttributeError: 'list' object has no attribute 'shape'

I'm using the Tensorflow backend (TF 2.19.0 in both environments I'm using).

[dhantule]

Hi @mmtrebuchet, thanks for reporting this.

I've tried to run this code and I faced the same error with keras version 3.10.0 in this colab gist. We'll look into this and update you.