Demo of serialisation bug in keras with loss_weights as variables

serialisation_test.py

from keras import losses 
from keras.models import Model, load_model
from keras.layers import Input, Dense
import keras.backend as K
from keras import optimizers

def build_dense_model_two_outputs(dim_in, dim_hidden, dim_out):

    model_input = Input(
              shape=(dim_in,),
              name='Input')

    hidden = Dense(units=dim_hidden,
                   activation=K.relu)(model_input)

    model_output_1 = Dense(units=dim_out)(hidden)
    model_output_2 = Dense(units=dim_out)(hidden)

    model = Model(inputs=model_input,
                  outputs=[model_output_1, model_output_2])

    return model


def save_and_reload_model(model):
  # Tests only that saving and loading doesn't throw, not
  # that it's done so correctly :)
  model.save('test.h5')
  del model
  model = load_model('test.h5')
  

def test_scalar_weights():
  # This works:
  model = build_dense_model_two_outputs(2, 3, 4)

  loss_weights = [1.0, 0.5]
  model.compile(loss=[losses.mse, losses.mse], 
                loss_weights=loss_weights,
                optimizer='adam')    

  save_and_reload_model(model)


def test_variable_weights():
  # This does not work (but training with these weights
  # would work)
  model = build_dense_model_two_outputs(2, 3, 4)

  loss_weights = [K.variable(1.0), K.variable(0.5)]
  model.compile(loss=[losses.mse, losses.mse], 
                loss_weights=loss_weights,
                optimizer='adam')      

  save_and_reload_model(model)