Open liujiyuan13 opened 4 years ago
Hi Jiyuan,
Thank you for paying attention to our work.
To answer your question, yes, the first iteration does affect the (middle) 'result' of the second iteration, because Variational Inference (VI) uses an iterative manner to optimize the model for inferring the approximate posteriors.
Actually, for different data sets, you may need to tune different values for itrs. I would suggest you plotting the Evidence Lower BOund (ELBO) curve W.R.T. iterations. When its value does not change too much as the iteration increasing, it means the VI converges.
Bye, Zhen
On Tue, Sep 8, 2020 at 11:30 AM Jiyuan notifications@github.com wrote:
Hello Zhen Wang,
I found your paper on IJCAI20 and very interested in the proposed Bayesian model to detection outliers. Next, I plan to compare your method in my ongoing paper, but have some questions about your code:
Do the variables "itr" and "itrs" take effects in each iteration? In other words, would the results when itr=1 affect the calculation when itr=2? If no, I prefer to set itrs=1 for accelerating the computation. If yes, does them keep the same for all datasets?
Could you tell me something more about this ? Thanks so much !
— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/zwang-datascience/MVAD_Bayesian/issues/1, or unsubscribe https://github.com/notifications/unsubscribe-auth/AHOPVIK2RAACSCMBBHELRW3SEWQORANCNFSM4Q7J7I2Q .
So many thanks to you!
Are there any tricks to judge whether the procedure converge ? For example, (ELBO^{t} - ELBO^{t-1}) / ELBO^{t} < threshold, or something like this. I would like to run the code on multiple datasets, so it's much more convenient if the code stop automatically.
Yours, Jiyuan
Hello Zhen Wang,
I found your paper on IJCAI20 and very interested in the proposed Bayesian model to detection outliers. Next, I plan to compare your method in my ongoing paper, but have some questions about your code:
Do the variables "itr" and "itrs" take effects in each iteration? In other words, would the results when itr=1 affect the calculation when itr=2? If no, I prefer to set itrs=1 for accelerating the computation. If yes, does them keep the same for all datasets?
Could you tell me something more about this ? Thanks so much !