Introducing Flabby Fred 0.5

[Dann Corbit]

Instead of doubling the size of the net, I cut it in half.
You will have to make your own net to use it.
The one that is crammed into his helmet is the 20MB size and he needs 10MB.
Could possibly be useful for cell phones.

Whatever it is you have to do to make a net (starting with nothing) is what you will have to do for ff-0.5 to use it.

[gaard]

Instead of doubling the size of the net, I cut it in half.
You will have to make your own net to use it.
The one that is crammed into his helmet is the 20MB size and he needs 10MB.
Could possibly be useful for cell phones.

Whatever it is you have to do to make a net (starting with nothing) is what you will have to do for ff-0.5 to use it.

Could you please be a little more specific? It seems that your post is intentionally vague. What do you mean by "helmet"? Why would it be useful for cellphones? Is this a scavenger hunt or similar?

[towforce]

Could be a sarcastic response to Fat Fritz, maybe?

[Dann Corbit]

Fat Fritz doubled the size of the net, so why not cut it in half.
But that is just the start of the work, and I am way too lazy to complete it.

[Dann Corbit]

Made a correction, same link.

Of course, it won't do you any good without a weights file.
Unless you build your own.

[hgm]

If you want to save on the size of the weights file in a more innovative way, you could try replacing all the tables with King location outside the area where it is usally found (i.e. the white King at or next to c1 or g1, etc.) by a single table (admittedly of about 4 times larger size) accessed in a King-relative way (i.e. as PRST[n][pieceType][sqr - kingSqr] instead of KPST[n][pieceType][kingSqr][sqr]).

This would replace 52 tables (the 64 board squares minus the two 3x2 areas around the castling targets) by a single table roughly 4 times the size (15x15 instead of 8x8), for a total savings of a factor 4. And it is not obvious that it would help much to distinguish positions with the King 'out in the open'. Especially since the weights for these cases will mostly be very poorly trained. So it might be better to 'pool' all info that the training material contains on such positions, and generalize the conclusions amongst them.

[Dann Corbit]

If there is any value in reduction of the weights file, I do not think it relies on size but computation.
I guess that there are some very underpowered systems for which the 20MB weights file is a burden to compute with (e.g. cell phones).
A half sized file might lift the burden a little.
But I use my computers for other things than generation of weights files.
So it may never be tried.

[Daniel Shawul]

I already have half the number of weights in the input layer by taking advantage of vertical symmetry.
But that doesn't decrease the inference time unlike your ff-0.5 idea.
Who knows a NNUE eval that is faster than classic could result in a better engine.

[towforce]

If there is any value in reduction of the weights file, I do not think it relies on size but computation.
I guess that there are some very underpowered systems for which the 20MB weights file is a burden to compute with (e.g. cell phones).
A half sized file might lift the burden a little.

Obviously, if a file size is halved, then it cannot store as much data. If there is a linear relationship between the file size and amount of chess knowledge, then the half sized file would only know about half as much about chess.

A quick reference to the shallow/deep knowledge thread though: there might be a lot to learn about something, but then you realise there's an underlying concept (deep knowledge), and that underlying concept might enable you to "know" everything you knew before with a lot less learning (which would correspond to a smaller file size).

[mar]

Obviously, if a file size is halved, then it cannot store as much data. If there is a linear relationship between the file size and amount of chess knowledge, then the half sized file would only know about half as much about chess.

you fail to understand a simple fact that there's a correlation between engine speed and strength, so the extra knowledge has to outweigh the loss of performance