Find a good way to back propagate the error of the Discriminator to the Generator.

m-stoeckel / TrumpTweetGAN

1 stars 0 forks source link

Find a good way to back propagate the error of the Discriminator to the Generator. #1

Open m-stoeckel opened 4 years ago

m-stoeckel commented 4 years ago

Ideas:

SeqGAN treats their RNN Generator as an agent of Reinforcement Leraning and uses Monte Carlo Search to find the best possible next word.
RelGAN: ?
Treat the output as a single 1D character tensor that is fed into a Bi-LSTM classifier and see if that is easier to back propagate

GeneratingGraphs93 commented 4 years ago

Ich glaube das passt auch dazu: https://arxiv.org/abs/1905.01976 Die verwenden einen Autoencoder um den diskreten Text in eine kontinuierliche Darstellung umzuwandeln. Der Generator generiert auch den Text in dieser Darstellung und beides wird so an den Discriminator weitergegeben.