[Feature request] Complex dtype input support for layers

[refraction-ray]

See the original issue in tensorflow's repo: https://github.com/tensorflow/tensorflow/issues/65306

Current behavior?

The following code works well for tf2.15 but fails tf2.16.1 with the introduction of keras3,

import tensorflow as tf
print(tf.__version__)

class MyDenseLayer(tf.keras.layers.Layer):
    def __init__(self, num_outputs):
        super(MyDenseLayer, self).__init__()
        self.num_outputs = num_outputs

    def build(self, input_shape):
        self.kernel = self.add_weight(shape=[int(input_shape[-1]),
                                         self.num_outputs], )

    def call(self, inputs):
        kernel = tf.cast(self.kernel, tf.complex64)
        return tf.matmul(inputs, kernel)

layer = MyDenseLayer(10)
layer(tf.zeros([10, 5], dtype=tf.complex64))

For short, tf2.16.1 with newly introduced keras 3 seems not supporting the input tensor for a layer in complex dtype

Relevant log output

[usr/local/lib/python3.10/dist-packages/keras/src/backend/common/variables.py](https://localhost:8080/#) in standardize_dtype(dtype)
    428 
    429     if dtype not in dtypes.ALLOWED_DTYPES:
--> 430         raise ValueError(f"Invalid dtype: {dtype}")
    431     return dtype
    432 

ValueError: Invalid dtype: complex64

It would be better that keras3 can support complex valued input for layers as keras did before. Complex valued input is very common in quantum machine learning use cases.

[mehtamansi29]

Hi @refraction-ray -

We have raised an internal fix for this issue. Soon it will be reflected in the documentation as well. Thanks.