Open Noahsark opened 6 years ago
Hi Li,
I still have some problems with fine-tuning so still cannot get the same results as in the paper, but inserting cluster vectors + using GMM prior certainly gives better result, comparing to baseline. Though, I hope to made fine-tune work and compare.
Best regards, Nikolai.
On 26 Mar 2018, at 5:17 AM, Li REN notifications@github.com wrote:
Hi Nikolai,
Did you reproduce the result that was reported in the original paper? May I ask the exact parameters?
Best Regards, Li
— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/yiyang92/vae_captioning/issues/1, or mute the thread https://github.com/notifications/unsubscribe-auth/AQhmUwXX8dbgdyByXKq3IQu8STyzMD9lks5tiAlXgaJpZM4S6T4g.
Thanks for the response.
By the way, the Zhusuan was updated I believe and some of your code is no longer compatible. Could you please make an update or share the workable Zhusuan that you are using?
Thank you so much.
Hi Li,
Wow, thats interesting. I`m using zhusuan 0.3.0 and it apparently was updated recently so it doesnt work. I will try myself later and coorect it or just note to install 0.3.0 version in readme. I will commit some update with a bit improve fine-tuning soon, but I want to make sure that it works really better.
Best regards, Nikolai
On 26 Mar 2018, at 10:05 AM, Li REN notifications@github.com wrote:
Thanks for the response.
By the way, the Zhusuan was updated I believe and some of your code is no longer compatible. Could you please make an update or share the workable Zhusuan that you are using?
Thank you so much.
— You are receiving this because you commented. Reply to this email directly, view it on GitHub https://github.com/yiyang92/vae_captioning/issues/1#issuecomment-376026228, or mute the thread https://github.com/notifications/unsubscribe-auth/AQhmU_wLlupN3f1T6zTMeXrpjIPisUyvks5tiE0CgaJpZM4S6T4g.
Hello. Have you try to reproduce the CVAE baseline? In my experiment, this vanilla CVAE (means a N(0,1) prior distribution) could achieve much better result than he reported, by simply tuning the parameter (the ratio between two losses). Thank you very much!
Hi Nikolai,
Did you reproduce the result that was reported in the original paper? May I ask the exact parameters?
Best Regards, Li