Compare weights - to see what is (un)learned

[MelleKoning]

This adds some code to compare two weightfiles in detail.

When running a testsuite over two nets as defined in nets.txt, the standard output will contain the positions where the result was different for.

Example output for two nets:

 1000  problems...  C:\lc0\lc0.exe
  nodes:0 nodes
  weight:61826
 {
     "logFile": "C:\\lc0\\logs.txt",
     "Threads": 2,
     "minibatchsize": 32,
     "SmartPruningFactor": 1,
     "VerboseMoveStats": false,
     "HistoryFill": "always",
     "WeightsFile": "C:\\lc0\\nets\\61826"
}
*** if you need to pause all instances runing just create a file named 'pause-Position-tester.txt'

Run 1 of 2: 61826, 0 nodes
107/1000,  Agree:10.70%, End of set: 2020-01-09 18:20, avg npm: 1
Run 2 of 2: 61890, 0 nodes
109/1000,  Agree:10.90%, End of set: 2020-01-09 18:21, avg npm: 1

r1bqk2r/2p1bppp/p1np1n2/1p2p3/3PP3/1B3N2/PPP2PPP/RNBQ1RK1 w kq - bm c3 ; 317
 w:61826 True w: 61890 False
3rkbr1/1b1n1p2/p1p1p1q1/4P2p/PppP1BpP/2N3P1/1P2BP2/R1Q1R1K1 w - - bm Bxc4 ; 400
 w:61826 False w: 61890 True
1qrb1rk1/3b1p1p/p2p2p1/3Pp1P1/2N2n2/P3BP2/KP1Q3P/2R2BR1 w - - bm Bxf4 ; 3303
 w:61826 True w: 61890 False
1r3rk1/2q2pp1/1nn1p1p1/p2p2P1/2pPPP1P/P1P5/3NBR2/R4QK1 w - - bm h5 ; 7681
 w:61826 False w: 61890 True
1rb2rk1/1p2bp2/pN2pn1p/4N1q1/7p/3B4/P1P3PP/1R2QR1K w - - bm Nf3 Qc3 ; 10955
 w:61826 False w: 61890 True
2b4k/6pp/4pb2/p1q1p3/1pBrP2P/1P3P2/PNP4Q/1K4R1 w - - bm h5 ; 13795
 w:61826 True w: 61890 False
r1bq1r1k/p2nbpp1/2p1p2p/3nP3/PppPN1Q1/6P1/1PN2PBP/R1B2RK1 w - - bm Qh5 ; 13946
 w:61826 True w: 61890 False
4r1k1/1p3p2/p2b2p1/5nP1/3pBP2/2PP1K2/1P1B4/7R w - - bm c4 ; 18668
 w:61826 True w: 61890 False
2r1r1k1/1bqn1pbp/p1p2np1/1p4N1/1P1P4/P1NB3P/2QB1PP1/2RR2K1 w - - bm Qa2 ; 18732
 w:61826 False w: 61890 True
2rq1rk1/p2nbppb/Ppp3np/3pP3/3P4/1N2BN2/1P2BPPP/R2Q1RK1 w - - bm Ne1 ; 23524
 w:61826 False w: 61890 True
5b2/1N3Qpk/p7/1pp1p3/P7/1P2q2P/1B4P1/6K1 w - - bm Qf2 ; 32131
 w:61826 True w: 61890 False
3rr1k1/1pp2pp1/2npqb1p/pB6/P3b1P1/2P1B2P/1PPQ1P2/1K1RN1R1 w - - bm g5 ; 34782
 w:61826 True w: 61890 False
r1bq1rk1/4bppp/pnn1p3/3pP3/1ppP1PP1/4BN2/PPPQBN1P/R4RK1 w - - bm g5 ; 37750
 w:61826 True w: 61890 False
5rk1/p5p1/P4r1p/1Np1R1n1/3P2qN/1P1Q1bP1/5P1P/2R1K3 w - - bm Rcxc5 ; 43780
 w:61826 False w: 61890 True
r3kb1r/1p1b2pp/2nq1p2/8/p1BN4/3Q3P/PB3PP1/R4K1R w kq - bm Re1+ ; 49844
 w:61826 True w: 61890 False
q7/2nnb1kp/6p1/3pp1P1/1p2P2P/1P2B1N1/1PK2Q2/5B2 w - - bm Kb1 ; 52702
 w:61826 False w: 61890 True
r1b1r1k1/1pq2pp1/p2pp2p/2n5/3NN1Pb/1B2Q3/PPP2P1P/2KR2R1 w - - bm f4 ; 53457
 w:61826 False w: 61890 True
1kr5/8/3Pp3/2np4/p2N4/1p6/1PP3PP/1K1R4 b - - bm e5 ; 62903
 w:61826 True w: 61890 False
2r1k1r1/1q2bpP1/p2p4/1p6/4N2P/PP2B3/2P1N1bK/R3QR2 b - - bm Bxe4 ; 76208
 w:61826 False w: 61890 True
r6r/1pq2pk1/p2p1np1/3Pn2p/5Q2/1N4PP/PPP5/1K1R1B1R b - - bm Rac8 ; 77092
 w:61826 False w: 61890 True
3r1rk1/pb2q1p1/2p4p/1p4p1/1P2P3/2N2BP1/P1QB1RKP/8 b - - bm Qxb4 ; 87576
 w:61826 False w: 61890 True
4k3/8/1p5R/p4N2/n4PP1/5K2/8/6r1 b - - bm Rf1+ ; 89788
 w:61826 False w: 61890 True
r1bqk2r/2p1bppp/p1np1n2/1p2p3/3PP3/1B3N2/PPP2PPP/RNBQ1RK1 w kq - bm c3 ; 113059
 w:61826 True w: 61890 False
3rkbr1/1b1n1p2/p1p1p1q1/4P2p/PppP1BpP/2N3P1/1P2BP2/R1Q1R1K1 w - - bm Bxc4 ; 113092
 w:61826 False w: 61890 True
 positions learned:13 positions unlearned:11

I just tested on 200 nodes instead of 0 nodes as well, although the number of found moves is higher (of course) the difference between nets has also increased:

Run 1 of 2: 61826, 200 nodes
141/1000,  Agree:14.10%, End of set: 2020-01-09 18:57, avg npm: 172
Run 2 of 2: 61890, 200 nodes
142/1000,  Agree:14.20%, End of set: 2020-01-09 19:22, avg npm: 172
 positions learned:26 positions unlearned:25