Saving models with DenseVariational layers

[wbeardall]

Currently, models utilising DenseVariational fail to save (in my experience), because of them lacking a get_config method. Is there any timeline on when this might be implemented?

[bram-wiggers-viacon]

I'm also experiencing this problem, is there a solution found already?

[nils-fl]

I am facing the same problem, and I think the problem is the weights names. When I use one DenseVariational layer and print out the weight names, I get a duplicated entry (3 and 4):

0 lstm/lstm_cell/kernel:0
1 lstm/lstm_cell/recurrent_kernel:0
2 lstm/lstm_cell/bias:0
3 dense_variational/constant:0
4 dense_variational/constant:0

Saving weights_only yields:

RuntimeError: Unable to create link (name already exists)

And full model saving yields the get_config error:

NotImplementedError: Layer DenseVariational has arguments in `__init__` and therefore must override `get_config`.

[KazukiOhta]

I think we can circumvent this problem by using the save_weights method.