I am working with a simple TensorFlow model given by

    model = models.Sequential()
    model.add(InputLayer(input_shape = d.shape))
    model.add(Ident(NlastDim = d.shape[-1]))
    model.add(Projection(Nproj = Nproj))
    model.add(Dense(Nout, activation = 'linear'))
    model.compile(optimizer = 'adam', loss = 'mse', metrics=[tf.keras.metrics.RootMeanSquaredError()])

    prior = MyPrior(simModel['prior'])
    trainingGenerator = DataGenerator(simModel, N, NperBatch, Nbatches, prior)
    validationGenerator = DataGenerator(simModel, N, Nho, 1, prior)
    model.summary()
    model.fit(trainingGenerator, validation_data = validationGenerator)

The Ident and Projection layers are custom layers. Ident is defined to be an identity layer for debugging purposes. model.summary prints the model as follows

Model: "sequential"                                                                  _________________________________________________________________
 Layer (type)                Output Shape              Param #   
=================================================================                     ident (Ident)               (None, 2000, 3)           0         
                                                                  
projection (Projection)     (None, 4)                 0                                                                                               
dense (Dense)               (None, 2)                 10        
                                                                           
=================================================================                    Total params: 10 (40.00 Byte)   
Trainable params: 10 (40.00 Byte)                                                    
Non-trainable params: 0 (0.00 Byte)

However after this the Ident layer is called with

[Ident] call in-shape: (None, None, None)                   

After poking around it turns out that also a Dense layer, put first, gets called with this all None-shape. However, it can still provide an output defining the inner-most dimension based on the number of features specified. In my custom layers, I could not find a way to properly deal with this case. Using a placeholder results in:

    ValueError: Exception encountered when calling layer 'ident' (type Ident).
    
    in user code:
    
        File "/home/pingu/Documents/Projekte/2024-02-AutomaticStatistician self/src/./freqFlow.py", line 205, in call  *
            return tf.reshape(inputs, shape = inputs.shape[:-1] + (self.NlastDim,))
    
        ValueError: Tried to convert 'shape' to a tensor and failed. Error: Cannot convert a partially known TensorShape (None, None, 3) to a Tensor.
    
    
    Call arguments received by layer 'ident' (type Ident):
      • inputs=tf.Tensor(shape=(None, None, None), dtype=float32)

The Ident layer is defined as follows

class Ident(tf.keras.layers.Layer):
    def __init__(self, NlastDim = 1, **kwarks):
        super().__init__(**kwarks)
        self.NlastDim = NlastDim
        Log(2, f'[Ident] NlastDim={NlastDim}');
    def build(self, input_shape):
        Log(2, f'[Projection] build input_shape={input_shape}');
        super().build(input_shape)
    def call(self, inputs):
        Log(2, f'[Ident] call in-shape: {inputs.shape}')
        dimLast = inputs.shape[-1]
        if dimLast is None:
            return tf.reshape(inputs, shape = inputs.shape[:-1] + (self.NlastDim,))
        return inputs

I tried various solutions for dealing with the all-Nones by trying to implement compute_output_shape (which never gets called) or intercepting special cases in the code, trying to impose a defined inner-most dimension.
I did not expect a input tensor with all-Nones. I do use data generators for this model.

Thank you in advance.