neural network - Keras throws a DisconnectedInputError due to custom loss function -

i trying implement regularization term norm of gradient improved wgan training in keras theano backend. want penalize l2 norm of gradient based on how far away 1.

i implementing custom loss this:

def get_gradient_norm(model, y_pred):     weights = model.trainable_weights     gradients = model.optimizer.get_gradients(k.mean(y_pred), weights)     acc = none     g in gradients:         s = k.sum(k.square(g))         if acc == none:             acc = s         else:             acc = s + acc     return k.sqrt(acc)  def make_w_reg_loss(model):     lvar = k.variable(lamb, name="lambda")      def foo(y_true, y_pred):         gnorm = get_gradient_norm(model, y_pred)         return lvar * k.square(gnorm - 1)  return foo  [...]  critic.compile(loss=make_w_reg_loss(critic), optimizer=rmsprop(learn_rate)) 

it throws disconnectedinputerror once training process tries attempt gradient of custom loss function.

why?

replacing loss standard loss works. error loss function defined.

see gist a minimal not-working example of attempt

edit:

so think know how make work now. first randomly added term loss, directly before returning foo(y_true, y_pred):

k.mean(y_pred) - k.mean(y_pred) 

clearly constant zero, , if use term loss zero. if add "constant zero" regularization loss works fine. loss, non-zero comes regularization, , optimization on many train_on_batch reduce loss well.

so odd issue theano being bit overzealous in throwing exceptions? question still stands: why throw in original code. since adding constant 0 term fixes it, looks bug me?

i implement improved wgan in keras too, , i'm surprised see how solved "issue". did verify trought experiments wgan-gp loss working intended? should easy check, stable training enable use of deep discriminator ;) same work done tensorflow backend, , try looking @ code , code here: keras improved wgan

i happy hear updates, i'll write again here have working code of wgan-gp in keras/tensorflow! p.s. link above implementing procedure in tensorflow code, forcing use tf training functions. approach, can define keras loss, using of our usual keras high level api training ;)

edit: code seems work totally k backend, code should run tensorflow backend too. did try changing backend check if issue/bug related theano?

2nd edit: calculating gradient w.r.t weights, in wgan-gp paper gradient penalty calculated starting gradient w.r.t average samples between generated , real samples. bring different results. in following link can find nice improved wgan loss implementation, could work on theano too: https://github.com/farizrahman4u/keras-contrib/