Which arg should I follow to train SEQGAN?

pclucas14 / GansFallingShort

Code for "Language GANs Falling Short"

58 stars 11 forks source link

Which arg should I follow to train SEQGAN? #5

Closed cloudygoose closed 5 years ago

cloudygoose commented 5 years ago

Hi,

Sorry one more question, Which arg should I follow to train SEQGAN(not leakGAN)? Is it the one in real_data_experiments/trained_models/news/word/best_gan/args.json? Note that "seqgan_reward": 0, so I guess that's for leakGAN?

Thanks!

pclucas14 commented 5 years ago

since the execution structure for leakgan is different, we train it in a separate file main_leak.py The --seqgan reward flag gives the same reward to all timesteps, while setting this flag to false gives a per timestep reward (as described in MaskGan).

If you want to do SEQGAN (not leakGAN), you can just set --seqgan_reward 1. Note that our implementation does not do K Monte Carlo rollouts at every timestep, only 1

cloudygoose commented 5 years ago

That's good to know, thanks! Other than seqgan_reward, is there any other hyper-parameter needed to be changed(For the results in your paper, are you using the same hyper-parameters for seqgan and leakgan)?

pclucas14 commented 5 years ago

Our best performing GANs were all without LeakGAN. (In my personal opinion), having a leaked signal does not help training. This finding is concurrent with https://openreview.net/pdf?id=rJMcdsA5FX.

For a complete list of the GAN hparams, please have a look at https://github.com/pclucas14/GansFallingShort/blob/master/real_data_experiments/trained_models/news/word/best_gan/args.json