return of compute_output_shape

[innat]

The compute_output_shape return is changed. Following sample code, works in Keras 2 but breaks in Keras 3. The thing is, in the following code, as stated there, what are the difference of those returns.

class ReshapeLayer(keras.layers.Layer):
    def __init__(self, new_shape, **kwargs):
        super(ReshapeLayer, self).__init__(**kwargs)
        self.new_shape = new_shape
    def build(self, input_shape):
        super(ReshapeLayer, self).build(input_shape)
    def call(self, inputs):
        return tf.reshape(inputs, self.new_shape)
    def compute_output_shape(self, input_shape):
        # option 1 # works in keras 2 and 3.
        # return (input_shape[0],) + self.new_shape 

        # option 2 #  < ---- works in keras 2 but not 3
        return ( 
            input_shape[0],
            self.new_shape
        )
    
input_shape = (None, 10)  
new_shape = (5, 2)
model = keras.Sequential([
    keras.layers.Dense(20, input_shape=input_shape),
    ReshapeLayer(new_shape)
])
model.summary()

Also, In some official layers, for example multi-head-layer, or conv3d layer, and also from docstring, the tf.TensorShape is used in compute_output_shape func. But doesn't it bring requirement to have tensorflow for other backends or any equivalent keras.ops?

[sampathweb]

Option 2 seems to return a Tuple of tuples instead of single value. So not sure if they are equivalent.

Also, the code references, you pointed to for tf.TensorShape don't exist in the main Keras Repo. It looks like its pointing to a forked repo that may not be updated. Can you check again?

[innat]

@sampathweb

Option 2 seems to return a Tuple of tuples instead of single value. So not sure if they are equivalent.

What I meant is their type (tuple). Regardless any approaches, option 2 used to work with keras 2. Now, in keras 3, I faced an issue with this approaches, specifically on tf and jax backend but successful on torch backend. Please, check this workflow.

# with tensorflow

        if dtype not in ALLOWED_DTYPES:
>           raise ValueError(f"Invalid dtype: {dtype}")
E           ValueError: Exception encountered when calling VideoSwinBasicLayer.call().
E           
E           Invalid dtype: <property object at 0x7fe799453e00>
E           
E           Arguments received by VideoSwinBasicLayer.call():
E             • args=('<KerasTensor shape=(None, 4, 56, 56, 96), dtype=float32, sparse=False, name=keras_tensor_14086>',)
E             • kwargs=<class 'inspect._empty'>

# with jax

        if dtype not in ALLOWED_DTYPES:
>           raise ValueError(f"Invalid dtype: {dtype}")
E           ValueError: Exception encountered when calling VideoSwinBasicLayer.call().
E           
E           Invalid dtype: ArrayImpl
E           
E           Arguments received by VideoSwinBasicLayer.call():
E             • args=('<KerasTensor shape=(None, 4, 56, 56, 96), dtype=float32, sparse=False, name=keras_tensor_14075>',)
E             • kwargs=<class 'inspect._empty'>

To address it, returning the ouput shape info wrapping around tf.TensorShape did work for all backend, like this. Not sure, if it's the ideal approach.

Also, the code references, you pointed to for tf.TensorShape don't exist in the main Keras Repo. It looks like its pointing to a forked repo that may not be updated.

You're right, my bad. I should double check.

[google-ml-butler]

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