Model from MNIST example does not train

[puli-92]

I am trying to run the mnist example from the notebooks. Is it that the repository is not maintained anymore?

I first has errors from multiplication of float32 and complex64 datatypes. After explicitly casting one of the multiplication operands to complex64 in the MeshLayer's transform function, I don't get that error, but the model from the example notebook, does not train.

A warning is displayed that the "model does not have any trainable weights". The loss and accuracy remain constant for 200 iterations.

Is it maybe because the implementattions of custom keras layers (MeshLayer for e.g.) are outdated, i.e., updates to TF and keras made the current definitions of custom layers non-functional?