Lc0 test40 training fork
 Share
The version of the browser you are using is no longer supported. Please upgrade to a supported browser.Dismiss

View only
 
 
ABCDEFGHIJKLMNOPQRSTUVWXYZ
1
An Lc0 fork of test40 training by jhorthos, using test40 games but different training parameters. For questions or comments, ping jhorthos on Leela Chess Zero Discord.
2
I extend special thanks to aart and jjosh (Discord names) for help setting up training and rescoring, plus the general Leela community for various advice and of course the training games.
3
4
Starting net is Leela Id 42000 and all games used are T40 training games. Nets are numbered by training step, starting at the step for Id 42000 (500000).
5
I don't provide full tensorboard support but graphs below provide the main tensorboard data, with numbers found on training log tab. I also have full tensorboard records if anyone wants them.
6
See 800 node Elo test and Policy Position test tabs for periodic game tests. After step 530000 I reduced frequency of direct Elo test as Policy Position is much faster.
7
See training log tab for training numbers and source of training games.
8
See start yaml tab for the configuration at the start of training
9
This sheet is all updated manually, so be patient. Maybe someday I will figure out how to automate it - hasn't seemed a priority.
10
The first training game window is 1 million T40 training games starting on April 7 2019 with the window gradually growing to 3 million games.1 million new test40 training games are added every 10k training steps, with the new games being sequential by game deposit time-date stamp at http://data.lczero.org/files/training_data/
11
After step 570000 I changed to adding new games every 15k steps because I was running out of new test40 games. This will result in a slightly higher rate of position sampling.
12
13
Note on graphs below: large groups of new games are added every 10k training steps and often cause slight shifts in policy loss and MSE. Main T40 uses a smoother training window shift.
14
If you squint and use your imagination you can see policy loss bumping a bit every 10k steps (when new games are added to the training window) and then dropping as learning happens. MSE is noisier but probably has the same trends.
15
Brought to you by looserboard (TM):
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
Loading...