Is there a paper about this codes?

[lucaskingjade]

Dear Ofir Nachum,

Do you have a paper about this RNN and GAN to explain your codes?

Best
Qi

[ofirnachum]

Hi Qi,

Sorry, there isn't any additional written explanation beyond the readme. Is there a particular part you don't understand? I'm happy to explain.. and perhaps improve the readme in the process.

[Mariewelt]

Hi, can you please provide formulas for unsupervised generator loss, reward loss and reward itself? Thank you.

[ofirnachum]

Hi - I would suggest that you look into this paper: https://arxiv.org/pdf/1609.05473v4.pdf
It is based on the work in this repo (although the authors added a number of things to it), and so the explanation/equations in it may be helpful.

[herbertchen1]

so he use your code to get his paper?????? ehhhhhh, it seems .......

[speculaas]

Hi All,
Anyone know where the design of "D_STEPS" comes from? As in the following code:

I came all the way from https://arxiv.org/pdf/1805.11973 because in my GAN training, D_STEPS seemed to turn out to be an important hyperparameter.

sequence_gan/train.py

Line 22 in 6f0feff

d_steps: number of discriminator training steps per iteration

sequence_gan/simple_demo.py

Line 34 in 6f0feff

D_STEPS = 2 # how many times to train the discriminator per generator step

Best regards,