search

Pascalson / Conditional-Seq-GANs

GANs for Conditional Sequence Generation. Tensorflow. Including the code of paper "Improving Conditional Sequence Generative Adversarial Networks by Stepwise Evaluation" IEEE/ACM TASLP, 2019.
MIT License
34 stars 6 forks source link

Apply to my own data #2

Open dberma15 opened 5 years ago

dberma15 commented 5 years ago

Hi, I'm looking to apply this to my own data after reading the paper "Improving Conditional Sequence Generative Adversarial Networks by Stepwise Evaluation". I'm a bit confused in looking at the code, though. Where exactly are the for the input sequence and the ground-truth is read into the code? I see the training and test sequences, but those appear to be the input sequences. I'm not sure where the ground-truth responses are. As a follow up, what is the dev data?

Thanks

Pascalson commented 5 years ago

Hi,

1) The input sequences are the odd lines and the paired ground-truths are the even lines. 2) The dev data is "development set" or "validation set". It is used to early-stop the training.

Hope you find it helpful!

dberma15 commented 4 years ago

Thanks so much. That did help. I was wondering if you could explain a bit about how the reward works in this? What are the rewards at each time step and how are they determined?

Pascalson commented 4 years ago

The rewards for the GAN-based models are predicted by the discriminator (D). Each reward can be interpreted as the expected D's score the current sub-sequence will obtain. Then, the generator will tend to generate sequence s that can achieve higher expected D's scores.

dberma15 commented 4 years ago

Is the algorithm using teacher forcing when training the decoder?