mirror of
https://github.com/sockspls/badfish
synced 2025-05-07 03:39:35 +00:00
2 commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
Tomasz Sobczyk
|
d61d38586e |
New NNUE architecture and net
Introduces a new NNUE network architecture and associated network parameters
The summary of the changes:
* Position for each perspective mirrored such that the king is on e..h files. Cuts the feature transformer size in half, while preserving enough knowledge to be good. See https://docs.google.com/document/d/1gTlrr02qSNKiXNZ_SuO4-RjK4MXBiFlLE6jvNqqMkAY/edit#heading=h.b40q4rb1w7on.
* The number of neurons after the feature transformer increased two-fold, to 1024x2. This is possibly mostly due to the now very optimized feature transformer update code.
* The number of neurons after the second layer is reduced from 16 to 8, to reduce the speed impact. This, perhaps surprisingly, doesn't harm the strength much. See https://docs.google.com/document/d/1gTlrr02qSNKiXNZ_SuO4-RjK4MXBiFlLE6jvNqqMkAY/edit#heading=h.6qkocr97fezq
The AffineTransform code did not work out-of-the box with the smaller number of neurons after the second layer, so some temporary changes have been made to add a special case for InputDimensions == 8. Also additional 0 padding is added to the output for some archs that cannot process inputs by <=8 (SSE2, NEON). VNNI uses an implementation that can keep all outputs in the registers while reducing the number of loads by 3 for each 16 inputs, thanks to the reduced number of output neurons. However GCC is particularily bad at optimization here (and perhaps why the current way the affine transform is done even passed sprt) (see https://docs.google.com/document/d/1gTlrr02qSNKiXNZ_SuO4-RjK4MXBiFlLE6jvNqqMkAY/edit# for details) and more work will be done on this in the following days. I expect the current VNNI implementation to be improved and extended to other architectures.
The network was trained with a slightly modified version of the pytorch trainer (https://github.com/glinscott/nnue-pytorch); the changes are in https://github.com/glinscott/nnue-pytorch/pull/143
The training utilized 2 datasets.
dataset A - https://drive.google.com/file/d/1VlhnHL8f-20AXhGkILujnNXHwy9T-MQw/view?usp=sharing
dataset B - as described in
|
||
|
Tomasz Sobczyk
|
e8d64af123 |
New NNUE architecture and net
Introduces a new NNUE network architecture and associated network parameters, as obtained by a new pytorch trainer. The network is already very strong at short TC, without regression at longer TC, and has potential for further improvements. https://tests.stockfishchess.org/tests/view/60a159c65085663412d0921d TC: 10s+0.1s, 1 thread ELO: 21.74 +-3.4 (95%) LOS: 100.0% Total: 10000 W: 1559 L: 934 D: 7507 Ptnml(0-2): 38, 701, 2972, 1176, 113 https://tests.stockfishchess.org/tests/view/60a187005085663412d0925b TC: 60s+0.6s, 1 thread ELO: 5.85 +-1.7 (95%) LOS: 100.0% Total: 20000 W: 1381 L: 1044 D: 17575 Ptnml(0-2): 27, 885, 7864, 1172, 52 https://tests.stockfishchess.org/tests/view/60a2beede229097940a03806 TC: 20s+0.2s, 8 threads LLR: 2.93 (-2.94,2.94) <0.50,3.50> Total: 34272 W: 1610 L: 1452 D: 31210 Ptnml(0-2): 30, 1285, 14350, 1439, 32 https://tests.stockfishchess.org/tests/view/60a2d687e229097940a03c72 TC: 60s+0.6s, 8 threads LLR: 2.94 (-2.94,2.94) <-2.50,0.50> Total: 45544 W: 1262 L: 1214 D: 43068 Ptnml(0-2): 12, 1129, 20442, 1177, 12 The network has been trained (by vondele) using the https://github.com/glinscott/nnue-pytorch/ trainer (started by glinscott), specifically the branch https://github.com/Sopel97/nnue-pytorch/tree/experiment_56. The data used are in 64 billion positions (193GB total) generated and scored with the current master net d8: https://drive.google.com/file/d/1hOOYSDKgOOp38ZmD0N4DV82TOLHzjUiF/view?usp=sharing d9: https://drive.google.com/file/d/1VlhnHL8f-20AXhGkILujnNXHwy9T-MQw/view?usp=sharing d10: https://drive.google.com/file/d/1ZC5upzBYMmMj1gMYCkt6rCxQG0GnO3Kk/view?usp=sharing fishtest_d9: https://drive.google.com/file/d/1GQHt0oNgKaHazwJFTRbXhlCN3FbUedFq/view?usp=sharing This network also contains a few architectural changes with respect to the current master: Size changed from 256x2-32-32-1 to 512x2-16-32-1 ~15-20% slower ~2x larger adds a special path for 16 valued ClippedReLU fixes affine transform code for 16 inputs/outputs, buy using InputDimensions instead of PaddedInputDimensions this is safe now because the inputs are processed in groups of 4 in the current affine transform code The feature set changed from HalfKP to HalfKAv2 Includes information about the kings like HalfKA Packs king features better, resulting in 8% size reduction compared to HalfKA The board is flipped for the black's perspective, instead of rotated like in the current master PSQT values for each feature the feature transformer now outputs a part that is fowarded directly to the output and allows learning piece values more directly than the previous network architecture. The effect is visible for high imbalance positions, where the current master network outputs evaluations skewed towards zero. 8 PSQT values per feature, chosen based on (popcount(pos.pieces()) - 1) / 4 initialized to classical material values on the start of the training 8 subnetworks (512x2->16->32->1), chosen based on (popcount(pos.pieces()) - 1) / 4 only one subnetwork is evaluated for any position, no or marginal speed loss A diagram of the network is available: https://user-images.githubusercontent.com/8037982/118656988-553a1700-b7eb-11eb-82ef-56a11cbebbf2.png A more complete description: https://github.com/glinscott/nnue-pytorch/blob/master/docs/nnue.md closes https://github.com/official-stockfish/Stockfish/pull/3474 Bench: 3806488 |
Renamed from src/nnue/features/half_kp.cpp (Browse further)