loss_hack.py

def some_mode():
  auto_input = K.Input((whatever,))
  auto_z = encoder(auto_input)
  auto_output = decoder(auto_z)
  
  model = Model(input_layer, output_layer)
  
  loss1_ = K.mean(K.square(auto_input - auto_output))
  loss2_ = K.mean(K.square(auto_z))
  model.hp_loss2_weight = K.variable(1.0, name='loss2_weight')
  
  # instead of using model.add_loss(), we have to make a
  # "dummy" loss function that looks like a function of
  # true vs. predicted outputs
  def loss_total(y_true, y_pred):
    return loss1_ + model.hp_loss2_weight * loss_2_
  
  # also, we have to make similar dummy functions for each
  # metric we want to make
  def loss1(y_true, y_pred):
    return loss1_
  def loss2(y_true, y_pred):
    return loss2_
  def loss2_weighted(y_true, y_pred):
    return model.hp_loss2_weight * loss2_
  
  # finally, we compile the model using these dummy functions
  # as loss and metrics:
  model.compile(optimizer='adam', loss=loss_total, metrics=[loss1, loss2, loss2_weighted])
  
model = some_model()
# The final caveat is that now we have to specify inputs and outputs for the model when training it:
model.fit(X_train, X_train, validation_data=(X_val, X_val))

# now all the new monitored loss terms should appear in the history object in addition to the
# usuall "loss" and "val_loss" items:
print(model.history.history.keys())