SF-NN embedded

[Gabor Szots]

How come? You have fiber optic much sooner than me (I have optic cable to my house only this year) but cable/ADSL services (not a dial-up) about 10-12 years. Download speed of ADSL is not that bad.

You are right Pham. My memory is starting to worry me. It was an ADSL connection.

[MikeB]

I know that madly downloading everything must seem really stupid to a lot of people. Even I often question what the hell I am doing. It's like an addiction. I guess in a way it makes me feel like I am preserving a piece of history. Engine sites tend not to have a long shelf life, so I try and get while the getting is good, before a site goes offline and downloads disappear forever.

A few years ago, I had read that Robert Hyatt had a problem with his backups of Crafty, and as a result we are no longer able to download those ancient version of Crafty, because no one had preserved them. That story pained me, as Crafty really is an important part of computer chess history. Incidents like this are what motivate me to do what I do.

If that is what you like to do, it's not stupid. Or to say it another way , we all do some stupid things, because we can and we want too. You are not alone.

[peter]

I agree that in general it is solution for a problem which doesn't really exist (much like hybrid eval). It wastes bandwidth, storage space, and even makes things more cumbersome in some situations.

For example, say you wanted to use an old net with a new compile. You have been using the (old) baked in net, but if you want to test it with the new compile, you will have to download it again anyway.

While I can see the appeal of embedding the net for the masses, I think this is something which should be reserved for official releases only. If people are competent enough to seek out and use the dev builds, we should expect them to be competent enough to download 1 more file.

Well said!
To me main reason of embedding seems to be anyhow, that SF dev.- versions in framework will be tested extensively with one net only and developers want them to be tested outside the framework with this only one net too.

I can understand that well, testing new versions with more than one net makes things much more complicated and ressources- consuming.
Yet of course you can never be sure, that a new version wouldn't even score better with an older net, can you?
Or an older version with a new net, even worse, because that way round you'll miss possible progress of some nets now and then completely, won't you?

Computerchess nowadays is probably less a problem of developing but of testing and judging the developments, at least with statistically significant tests fighting draw- death and more and more differing hardware- time- performance of different engines and nets, mainly LC0- like and SF- like ones.

What I see as the main disadvantage in embedding like it works now, you cannot delete or replace the embedded net without deleting the compile. Of course that doesn't matter as for byo , but downloading offered embedded compiles gets somewhat devalued, doesn't it?

[Dann Corbit]

But why do we need embedding at all?

I had no trouble with downloading the eval file and to adjust the EvalFile setting correctly.

Because the world is entirely peopled with those who cannot browse to a file.
One may wonder how they will install the engine itself, but they will probably get a friend to help with that.

[cma6]

Once one has downloaded embedded SF-NN and installed it in the chess GUI, what else does one have to do, if anything, with the .bin (network) file? Or does one completely ignore the location for the parameter (in GUI) of the network file?
I don't see the advantage of embedding.

[jhellis3]

Downloading a network file and installing it is the exact same process as downloading an engine and installing it.

Apparently, the first time is quite easy, but the second time is exceedingly difficult....

[corres]

At first (06.08.2020) there were the SF+NNUE.exe and the net file of NNUE.
Now, in the "embedded" version there are the part of SF+NNUE.exe and the part of NNUE-net.
The required place on HDD/SSD near the same so we would not complain about the place requirement.
This situation make possible for the developer making stable connection between "classical" part of Stockfish and NN part of the Stockfish. furthermore the controlling and the checking of the used NNUE-net what is important for the good working of SF+NNUE. Moreover the built-in NNUE-net make possible to generate a buffer for NNUE-net, what speeding up the working of NNUE. Without buffer the SF+NNUE would be weaker about 5-10 Elo, I think.

[cma6]

Thanks for explanation of buffer. Bottom line: is is correct that one no longer has to work about a separate net file?

[syzygy]

Thanks for explanation of buffer. Bottom line: is is correct that one no longer has to work about a separate net file?

Yes, the point of embedding the file in the Stockfish binary is that the user does not have to fiddle with separate files or the EvalFile option at all.

[corres]

Thanks for explanation of buffer. Bottom line: is is correct that one no longer has to work about a separate net file?

Yes. You has not to work about the net file, but if you want to use another net file you can do it:
In that folder where you use the Stockfish.exe create a new folder with eval name and copy the other net file into the eval folder and at last on the side of Engine Options write in the long box
./eval/nn-xyz...nnue where the nn-xyz...nnue is the name of the net what you want to use.
The dot is important before the / !!