Ａ question relate to the D structure

[zxiaomzxm]

Hi, I realize a rnn-gan on a sequence data rencently.
But I found the D_loss is close to zero fast and G_loss not converged for a long time when I
add a RNN in Ｄ-net.And everything is OK when I discard the rnn structure in D-net.Why that could be
happened?Or should Ｉ use the bidirectional rnn structure as your paper used?

[olofmogren]

Hi, What do you mean by "everything is ok"? And what do you mean by "discarding the RNN structure"? What do you have in its pace? Best regards, Olof

…

-- Phone: +46-703-969624 Web: http://mogren.one Email: mogren@chalmers.se, olof@mogren.one

-------- Ursprungligt meddelande -------- Ämne: [olofmogren/c-rnn-gan] Ａ question relate to the D structure (#4) Från: zxiaomzxm <notifications@github.com> Till: olofmogren/c-rnn-gan <c-rnn-gan@noreply.github.com> Kopia: Subscribed <subscribed@noreply.github.com>

Hi, I realize a rnn-gan on a sequence data rencently. But I found the D_loss is close to zero fast and G_loss not converged for a long time when I add a RNN in Ｄ-net.And everything is OK when I discard the rnn structure in D-net.Why that could be happened?Or should Ｉ use the bidirectional rnn structure as your paper used? -- You are receiving this because you are subscribed to this thread. Reply to this email directly or view it on GitHub: #4

[zxiaomzxm]

Thanks for your reply.
"everything is ok" means the generate image looks good but the sequence seems has no temporal correlation.
"discarding the RNN structure" means I just use a vanilla fully connected net instead of vanilla rnn.
I test a bidirectional rnn ,but fails again.It seems that the generator can't deceive the discriminator and generate bad results.

[olofmogren]

I see. Does any of the tricks in Salismans et.al. help? https://arxiv.org/abs/1606.03498

…

-- Phone: +46-703-969624 Web: http://mogren.one Email: mogren@chalmers.se, olof@mogren.one

-------- Ursprungligt meddelande -------- Ämne: Re: [olofmogren/c-rnn-gan] Ａ question relate to the D structure (#4) Från: zxiaomzxm <notifications@github.com> Till: olofmogren/c-rnn-gan <c-rnn-gan@noreply.github.com> Kopia: Olof Mogren <olof@mogren.one>,Comment <comment@noreply.github.com>

Thanks for your reply. "everything is ok" means the generate image looks good but the sequence seems has no temporal correlation. "discarding the RNN structure" means I just use a vanilla fully connected net instead of vanilla rnn. I test a bidirectional rnn ,but fails again.It seems that the generator can't deceive the discriminator and generate bad results. -- You are receiving this because you commented. Reply to this email directly or view it on GitHub: #4 (comment)

[zxiaomzxm]

I will read this paper.Thanks