Page 2 of 3

Posted: Fri Aug 25, 2006 5:09 am
by Behrendt
Slow down m8 :lol: Dont wanna have 60 beta versions again :oops:

Enhanced Metropolis

Posted: Fri Aug 25, 2006 6:52 am
by clipi
Hello Nicholas

Have you ever considered using Delayed Rejection or some other
Adaptive Metropolis method such the one introduced by Heikki Haario?
it seem that they can greately improve efficiency of MTL.

Re: Enhanced Metropolis

Posted: Fri Aug 25, 2006 9:29 am
by Jedrzej_s
clipi wrote:Hello Nicholas

Have you ever considered using Delayed Rejection or some other
Adaptive Metropolis method such the one introduced by Heikki Haario?
it seem that they can greately improve efficiency of MTL.
clipi <- Can You show any examples of these methods ? Rendering time comparison or something ???

Posted: Fri Aug 25, 2006 10:55 am
by clipi
Well this method as far as I know never were used in any GI application. But since Indigo is more an experimental render then it might be good to try some inovative aproach.
The concept of delayed rejection is like this acording to MIRA paper
(On Metropolis-Hastings algorithms with delayed rejection )

In a Metropolis-Hastings algorithm, rejection of proposed moves is an intrinsic part of ensuring that the chain converges to the intended target distribution. However, persistent rejection, perhaps in particular parts of the state space, may indicate that locally the proposal distribution is badly calibrated to the target. As an alternative to careful off-line tuning of state-dependent proposals, the basic algorithm can be modified so that on rejection, a second attempt to move is made. A different proposal can be generated from a new distribution, that is allowed to depend on the previously rejected proposal.


An adaptive Metropolis-Hastings scheme: sampling and optimization
Authors: David H. Wolpert, Chiu Fan Lee

We propose an adaptive Metropolis-Hastings algorithm in which sampled data are used to update the proposal distribution. We use the samples found by the algorithm at a particular step to form the information-theoretically optimal mean-field approximation to the target distribution, and update the proposal distribution to be that approximation.

Posted: Fri Aug 25, 2006 11:03 am
by Jedrzej_s
interesting... :). We will see what Ono say ;).

Posted: Fri Aug 25, 2006 11:32 am
by OnoSendai
Very interesting Clipi... will read the paper later

Posted: Sat Aug 26, 2006 10:52 am
by clipi
Thanks Nicholas.. hope this will be usefull for GI!

Posted: Sun Aug 27, 2006 8:12 am
by scaron
could someone repost the first image that took 24hrs? i can't see the comparison :(

so at the moment it just looks like you have 13mins of an incomplete render

Posted: Sun Aug 27, 2006 11:59 am
by Jedrzej_s
Here is comparison:

- Render time: 13 min (new code):
Image

- Render time: 10 hours on 2 cores for each... (Indigo 0.5):
Image

Here is topic with this appartements final(s):
http://www.indigorenderer.com/forum/viewtopic.php?t=52

Posted: Mon Aug 28, 2006 1:16 am
by PakX
Hi,

I think that this is not a very good comparison because the rendertimes a tooo diferent.

I will be look more impressive, when nick post a 13 minutes rendering rendered in an older indigoversion.

But anyway. Good job Nick!

Posted: Mon Aug 28, 2006 7:15 am
by Jedrzej_s
Pak-X wrote:I will be look more impressive, when nick post a 13 minutes rendering rendered in an older indigoversion.
Yes, this would be better comparison or 2 finals renderings (first on old code and second on new code).

Posted: Thu Aug 31, 2006 2:48 am
by OnoSendai
Ok.. another image

Image

Render time: 55 mins, single thread.
Yeah, there is a lot of noise, but it shows that bidir + MLT is usable for sunlit scenes now (cross fingers).
Left overnight with a couple of threads, it should be baked to perfection, I think :)

Posted: Thu Aug 31, 2006 3:10 am
by PakX
Yes it looks very good. But I can just repeat my. I wanna see a Comparison between old and new Indigo :wink:

But it seems now being Maxwell-speed because it looks that its clean after 5 hours of rendering.

I look forward to you :)

Posted: Thu Aug 31, 2006 4:22 am
by manitwo
:lol: :D :D :lol: :D :lol:

Posted: Thu Aug 31, 2006 4:31 am
by Maluminas
Yes i would like to see a comparison too, between the old code and the new code. :D

What i mean is: Render a scene with the old code for a set period of time, lets say 15 minutes. Then render the exact same scene with the new code for the same time, 15 minutes (also on same computer, same image size etc.). Post the two pics so we can compare the new code to the old code with the images. This way we can see and evaluate the difference in performance between the old and the new code. :P