Guenther wrote: ↑Tue Oct 20, 2020 10:19 pm
Thanks for the quick fix - I also thought it might be some thread polling issue - compiled the current repo again and all is ok now.
Great. My intention was for that fix to be included in the initial release. I'm not exactly sure as to what happened to it, but rather than update the existing 1.0 release with the fix, my current plan is just to make a 1.1 release incorporating the fix and the new "phased" networks I've been working on. Hopefully 1.1 will be ready by the end of the week. The master branch you compiled should already be +10 elo to Seer 1.0, though the new phased networks are looking to be +100 elo with any luck.
As it stands for 1.0, the extra CLI thread, in my testing, is very light and shouldn't influence the performance of other engines.
Thanks for thinking of reviving the engines list page, if you can do it. If there were several capable and responsible volunteers allowed to update it, then it would be unlikely that it would be abandoned, as is the case if just one person can make updates. A single individual can suddenly be overwhelmed by outside factors and leave everything nonessential aside.
Engine Releases to be continued. If several capable and responsible volunteers like to contribute, send me a pm with username/e-mail address.
Thanks for thinking of reviving the engines list page, if you can do it. If there were several capable and responsible volunteers allowed to update it, then it would be unlikely that it would be abandoned, as is the case if just one person can make updates. A single individual can suddenly be overwhelmed by outside factors and leave everything nonessential aside.
Engine Releases to be continued. If several capable and responsible volunteers like to contribute, send me a pm with username/e-mail address.
Thanks,
Gerd
Maybe make a separate thread for this topic, both to keep Seer thread about Seer, and maybe easier to get volunteers and collect dates/links
Thanks for thinking of reviving the engines list page, if you can do it. If there were several capable and responsible volunteers allowed to update it, then it would be unlikely that it would be abandoned, as is the case if just one person can make updates. A single individual can suddenly be overwhelmed by outside factors and leave everything nonessential aside.
Engine Releases to be continued. If several capable and responsible volunteers like to contribute, send me a pm with username/e-mail address.
Thanks,
Gerd
Maybe make a separate thread for this topic, both to keep Seer thread about Seer, and maybe easier to get volunteers and collect dates/links
Does Seer do the int8 and int16 qunatization done in NNUE ?
If so is it done post-training or during training ? From a quick glance at the code, it doesn't look like the model
contains quantized layers.
Daniel Shawul wrote: ↑Tue Oct 27, 2020 1:47 pm
Does Seer do the int8 and int16 qunatization done in NNUE ?
If so is it done post-training or during training ? From a quick glance at the code, it doesn't look like the model
contains quantized layers.
Seer uses all floating point weights and does no weight quantization. On modern processors, single precision floating point performance is pretty similar to 32 bit integer performance. One advantage of using floating point weights is that it's much easier to experiment with different architectures (skip connections and mixing layer types complicates quantization schemes). The biggest downside is that incremental updates can't easily be undone as the floating point error would grow (to resolve this, I plan to explore quantizing just the feature transformer layer sometime in the future). The solution, at least for now, is just to copy the vectors on the stack for every ply.
I've released Seer 1.1 here: https://github.com/connormcmonigle/seer ... s/tag/v1.1 (this time with binaries!). It should prove substantially stronger than Seer 1.0 and also corrects an issue where an extra thread would be used. Estimated strength is around 3000 elo, though this a quite poor estimate due to limited testing. In any case, I'd greatly appreciate if it could be tested. More info can be found by following the link. Let me know if you have difficulty running it.
connor_mcmonigle wrote: ↑Mon Nov 02, 2020 7:03 am
I've released Seer 1.1 here: https://github.com/connormcmonigle/seer ... s/tag/v1.1 (this time with binaries!). It should prove substantially stronger than Seer 1.0 and also corrects an issue where an extra thread would be used. Estimated strength is around 3000 elo, though this a quite poor estimate due to limited testing. In any case, I'd greatly appreciate if it could be tested. More info can be found by following the link. Let me know if you have difficulty running it.