So, that. I get the following error:
theano.gof.fg.MissingInputError: ("An input of the graph, used to compute DimShuffle{0,1,2,x}(convolution1d_input_1), was not provided and not given a value.Use the Theano flag exception_verbosity='high',for more information on this error.", convolution1d_input_1)
I dug around and found that when training the whole thing, the discriminator had .trainable=True. So, even though it was compiled with False, when you turn it on afterwards it becomes trainable, and then the Theano graph is not well defined.
I was able to fix this by setting discriminator.trainable=True and recompiling it before its training, and afterwards setting it to False and recompiling the discriminator_on_generator.
Now I need to understand why the discriminator converges very fast while the generator diverges. Any ideas on this? I am not using images, but midi files expressed as vector sequences.
So, that. I get the following error:
theano.gof.fg.MissingInputError: ("An input of the graph, used to compute DimShuffle{0,1,2,x}(convolution1d_input_1), was not provided and not given a value.Use the Theano flag exception_verbosity='high',for more information on this error.", convolution1d_input_1)I dug around and found that when training the whole thing, the discriminator had .trainable=True. So, even though it was compiled with False, when you turn it on afterwards it becomes trainable, and then the Theano graph is not well defined.
I was able to fix this by setting discriminator.trainable=True and recompiling it before its training, and afterwards setting it to False and recompiling the discriminator_on_generator.
Now I need to understand why the discriminator converges very fast while the generator diverges. Any ideas on this? I am not using images, but midi files expressed as vector sequences.