Missing Output Shape

[frankmollard]

Hello,

when I build a model I see the following problem:

does anyone see a mistake?

[ZPZhou-lab]

If you want to build a model with 10 features as input tensor, maybe you could try set input_shape = (None, 10) like:

model = tf.keras.models.Sequential([
    DenseKAN(4),
    DenseKAN(1)
])
model.build(input_shape=(None, 10))

we leave None as a batch_size axis.

[frankmollard]

Thank you for your response.

When I use:
model.build(input_shape=(None, 10))

I get the following Error:

TypeError Traceback (most recent call last)
File /opt/conda/lib/python3.10/site-packages/keras/src/models/sequential.py:180, in Sequential.build(self, input_shape)
179 try:
--> 180 x = layer(x)
181 except NotImplementedError:
182 # Can happen if shape inference is not implemented.
183 # TODO: consider reverting inbound nodes on layers processed.

File /opt/conda/lib/python3.10/site-packages/keras/src/utils/traceback_utils.py:122, in filter_traceback..error_handler(*args, **kwargs)
120 # To get the full stack trace, call:
121 # keras.config.disable_traceback_filtering()
--> 122 raise e.with_traceback(filtered_tb) from None
123 finally:

File /opt/conda/lib/python3.10/site-packages/tfkan/layers/dense.py:101, in DenseKAN.call(self, inputs, *args, **kwargs)
99 def call(self, inputs, *args, **kwargs):
100 # check the inputs, and reshape inputs into 2D tensor (-1, in_size)
--> 101 inputs, orig_shape = self._check_and_reshape_inputs(inputs)
102 output_shape = tf.concat([orig_shape, [self.units]], axis=0)

File /opt/conda/lib/python3.10/site-packages/tfkan/layers/dense.py:125, in DenseKAN._check_and_reshape_inputs(self, inputs)
124 shape = tf.shape(inputs)
--> 125 ndim = len(shape)
126 try:

TypeError: Exception encountered when calling DenseKAN.call().

len is not well defined for a symbolic Tensor (Shape:0). Please call x.shape rather than len(x) for shape information.

Arguments received by DenseKAN.call():
• args=('<KerasTensor shape=(None, 10), dtype=float32, sparse=None, name=keras_tensor>',)
• kwargs=<class 'inspect._empty'>

During handling of the above exception, another exception occurred:

ValueError Traceback (most recent call last)
Cell In[3], line 6
1 # create model using KAN
2 model = tf.keras.models.Sequential([
3 DenseKAN(4),
4 DenseKAN(1)
5 ])
----> 6 model.build(input_shape=(None, 10))

File /opt/conda/lib/python3.10/site-packages/keras/src/layers/layer.py:222, in Layer.new..build_wrapper(*args, **kwargs)
219 @wraps(original_build_method)
220 def build_wrapper(*args, **kwargs):
221 with obj._open_name_scope():
--> 222 original_build_method(*args, **kwargs)
223 # Record build config.
224 signature = inspect.signature(original_build_method)

File /opt/conda/lib/python3.10/site-packages/keras/src/models/sequential.py:193, in Sequential.build(self, input_shape)
187 positional_args = [
188 param
189 for param in signature.parameters.values()
190 if param.default == inspect.Parameter.empty
191 ]
192 if len(positional_args) != 1:
--> 193 raise ValueError(
194 "Layers added to a Sequential model "
195 "can only have a single positional argument, "
196 f"the input tensor. Layer {layer.class.name} "
197 f"has multiple positional arguments: {positional_args}"
198 )
199 raise e
200 outputs = x

ValueError: Layers added to a Sequential model can only have a single positional argument, the input tensor. Layer DenseKAN has multiple positional arguments: [<Parameter "inputs">, <Parameter "*args">, <Parameter "**kwargs">]

[ZPZhou-lab]

could you please tell me the tensorflow version you use, I will check and try to fix this.
And I tested with Tensorflow >= 2.10 which will work fine.

[frankmollard]

It is 2.15.0

[ZPZhou-lab]

It is 2.15.0

I have find the reason, the tfkan uses the API of keras < 3, and keras 3 changes some fundamental API I use.

So, if you want to use tfkan now, you need just downgrade your keras version to below keras 3(like keras == 2.15)

I will take some time to fix this or alternatively add a restriction on keras < 3 in requirements.

For Kaggle users, the default Kaggle notebook environment install keras == 3.3.3, the following steps can help you change the keras version in the Kaggle notebook

Hope this can help you~😊

[frankmollard]

Thank you, I will check this very soon.