Fat Fritz 2 nets: Github posted vs commercial

[Tibono]

Hello,

we know both differ, according to their different MD5 hash code (ref. this post)
We as well know, thanks to Stefan's work, the free one is weaker, not that much (40 Elo points or so, under Stefan's testing condition)

But how much do they differ?
Can someone who would have the commercial net available, check on a byte-to-byte ground, and report please?

Github link for the free net
Thanks,
Tibono

[pohl4711]

A.Silver told me about the FF2 free github net:
"Out of 34 nets in succession, this one was the 4th"

So, I would say, the github net is 30 nets (or evolution steps/stairs) behind the commercial FF2 net. Because the commercial net is around +35 Elo stronger, this seems realistic to me.

[Tibono]

Thanks Stefan.
Kind regards,
Eric

[cpeters]

Nicensteyn!

Questions to the Fritz-tester:

Is there such a thing as a license-file attached (now?) to the binary/net? If any yes, what does it say?

@Tibono
"But how much do they differ?
Can someone who would have the commercial net available, check on a byte-to-byte ground, and report please?"

I've got a diff-file from hexdumperies of both nets; it's 361 Megabytl...

[Tibono]

I've got a diff-file from hexdumperies of both nets; it's 361 Megabytl...

OK, so a massive difference in data (consistent with the several versions mentionned above). Thank you for this test and outcome, Christian.
KR,
Eric

[cpeters]

Well the hexdumped/human readable form of one net is each ~170 mb in size/2626679 lines of text

Code: Select all

00000000: 162f f37a caa0 5a3e b400 0000 4665 6174  ./.z..Z>....Feat
00000010: 7572 6573 3d48 616c 664b 5028 4672 6965  ures=HalfKP(Frie
00000020: 6e64 295b 3431 3032 342d 3e35 3132 7832  nd)[41024->512x2
00000030: 5d2c 4e65 7477 6f72 6b3d 4166 6669 6e65  ],Network=Affine
00000040: 5472 616e 7366 6f72 6d5b 313c 2d31 365d  Transform[1<-16]
00000050: 2843 6c69 7070 6564 5265 4c55 5b31 365d  (ClippedReLU[16]
00000060: 2841 6666 696e 6554 7261 6e73 666f 726d  (AffineTransform
00000070: 5b31 363c 2d31 365d 2843 6c69 7070 6564  [16<-16](Clipped
00000080: 5265 4c55 5b31 365d 2841 6666 696e 6554  ReLU[16](AffineT
00000090: 7261 6e73 666f 726d 5b31 363c 2d31 3032  ransform[16<-102
000000a0: 345d 2849 6e70 7574 536c 6963 655b 3130  4](InputSlice[10
000000b0: 3234 2830 3a31 3032 3429 5d29 2929 2929  24(0:1024)])))))

(This is from the net with GIT origins.)

They are totally different in that very sense if that would matter (it doesn't make any sense to compare them like this methinks).

[Ozymandias]

Yes, I don't see much practical use to differences on that side. I did ask about a sim test, but no one seems to care about that aspect of it?