Failed VVLTC non-regression
https://tests.stockfishchess.org/tests/view/69d562b84088e069540a2288
LLR: -2.96 (-2.94,2.94) <-1.75,0.25>
Total: 386998 W: 99181 L: 99760 D: 188057
Ptnml(0-2): 35, 35792, 122429, 35203, 40
Failed STC non-regression
https://tests.stockfishchess.org/tests/view/69f3c6601e5788938e86a99e
LLR: -2.93 (-2.94,2.94) <-1.75,0.25>
Total: 33696 W: 8492 L: 8795 D: 16409
Ptnml(0-2): 124, 4209, 8504, 3868, 143
Many thanks to Dubslow, Torom, ces42, Shawn, vondele, Disservin and others for discussion.
## Summary
The venerable small net has been around for quite some time now, and while the big net architecture has substantially advanced with TI, the small net has stayed with plain HalfKA. It therefore presents a few burdens: multiple net architectures to maintain, multiple nets to train, and a whole lot of templates to deal with the variable L1 size.
Locally I measure a slowdown of -2.5% in NPS with this branch – and it's probably more on non-AVX512 architectures – but a pure slowdown of that magnitude would lead to more dramatic losses (even at VVLTC) than exhibited in the above tests, suggesting that the small net's lower eval quality is deleterious.
Bonus: Shawn found this interesting PGN among the VVLTC games: https://lichess.org/study/hvo8jflc/OeTOityv `master` seems to misevaluate the fortress because all positions go to small net (the material difference is larger than the threshold).
closes https://github.com/official-stockfish/Stockfish/pull/6796
Bench: 2877007
[STC](https://tests.stockfishchess.org/tests/live_elo/69cb991bc025c305b7daa219)
LLR: 2.97 (-2.94,2.94) <0.00,2.00>
Total: 47520 W: 12558 L: 12219 D: 22743
Ptnml(0-2): 190, 5474, 12102, 5795, 199
[LTC](https://tests.stockfishchess.org/tests/live_elo/69cc6b42731dfb72d2e92128)
LLR: 2.94 (-2.94,2.94) <0.50,2.50>
Total: 123756 W: 31996 L: 31497 D: 60263
Ptnml(0-2): 81, 13512, 34221, 13955, 109
Adds features for pawns pushing against other pawns. For indexing, we
use the existing threats feature set by adding the forward square to the
squares threatened by a pawn. This threat is only enabled when another
pawn occupies that square. We then add appropriate logic changes in
`update_piece_threats` and `features/full_threats.cpp`.
The wonderful Ciekce (of Stormphrax) and peregrine (of Reckless) also
had similar ideas with adding forward pawn movements to TI, and I
appreciate their giving me motivation to figure out `nnue-pytorch` and
test it.
closes https://github.com/official-stockfish/Stockfish/pull/6694
Bench: 2926703
Created from 2 distinct spsa tunes of the latest main net (nn-31337bea577c.nnue)
and applying the params to the prior main net (nn-e8bac1c07a5a.nnue). This
effectively reverts the modifications to output weights and biases in
https://github.com/official-stockfish/Stockfish/pull/5509
SPSA:
A: 6000, alpha: 0.602, gamma: 0.101
1st - 437 feature transformer biases where values are < 25
54k / 120k games at 180+1.8
https://tests.stockfishchess.org/tests/view/66af98ac4ff211be9d4edad0
nn-808259761cca.nnue
2nd - 208 L2 weights where values are zero
112k / 120k games at 180+1.8
https://tests.stockfishchess.org/tests/view/66b0c3074ff211be9d4edbe5
nn-a56cb8c3d477.nnue
When creating the above 2 nets (nn-808259761cca.nnue, nn-a56cb8c3d477.nnue),
spsa params were unintentionally applied to nn-e8bac1c07a5a.nnue rather
than nn-31337bea577c.nnue due to an issue in a script that creates nets
by applying spsa results to base nets.
Since they both passed STC and were neutral or slightly positive at LTC,
they were combined to see if the elo from each set of params was additive.
The 2 nets can be merged on top of nn-e8bac1c07a5a.nnue with:
https://github.com/linrock/nnue-tools/blob/90942d3/spsa/combine_nnue.py
```
python3 combine_nnue.py \
nn-e8bac1c07a5a.nnue \
nn-808259761cca.nnue \
nn-a56cb8c3d477.nnue
```
Merging yields nn-87caa003fc6a.nnue which was renamed to nn-1111cefa1111.nnue
with an updated nnue-namer around 10x faster than before by:
- using a prefix trie for efficient prefix matches
- modifying 4 non-functional bytes near the end of the file instead of 2
https://github.com/linrock/nnue-namer
Thanks to @MinetaS for pointing out in #nnue-dev what the non-functional bytes are:
L3 is 32, 4 bytes for biases, 32 bytes for weights. (fc_2)
So -38 and -37 are technically -2 and -1 of fc_1 (type AffineTransform<30, 32>)
And since InputDimension is padded to 32 there are total 32 of 2 adjacent bytes padding.
So yes, it's non-functional whatever values are there.
It's possible to tweak bytes at -38 - 32 * N and -37 - 32 * N given N = 0 ... 31
The net renamed with the new method passed non-regression STC vs. the original net:
https://tests.stockfishchess.org/tests/view/66c0f0a821503a509c13b332
To print the spsa params with nnue-pytorch:
```
import features
from serialize import NNUEReader
feature_set = features.get_feature_set_from_name("HalfKAv2_hm")
with open("nn-31337bea577c.nnue", "rb") as f:
model = NNUEReader(f, feature_set).model
c_end = 16
for i,ft_bias in enumerate(model.input.bias.data[:3072]):
value = int(ft_bias * 254)
if abs(value) < 25:
print(f"ftB[{i}],{value},-1024,1024,{c_end},0.0020")
c_end = 6
for i in range(8):
for j in range(32):
for k in range(30):
value = int(model.layer_stacks.l2.weight.data[32 * i + j, k] * 64)
if value == 0:
print(f"twoW[{i}][{j}][{k}],{value},-127,127,{c_end},0.0020")
```
New params found with the same method as:
https://github.com/official-stockfish/Stockfish/pull/5459
Passed STC:
https://tests.stockfishchess.org/tests/view/66b4d4464ff211be9d4edf6e
LLR: 2.94 (-2.94,2.94) <0.00,2.00>
Total: 136416 W: 35753 L: 35283 D: 65380
Ptnml(0-2): 510, 16159, 34416, 16597, 526
Passed LTC:
https://tests.stockfishchess.org/tests/view/66b76e814ff211be9d4ee1cc
LLR: 2.95 (-2.94,2.94) <0.50,2.50>
Total: 159336 W: 40753 L: 40178 D: 78405
Ptnml(0-2): 126, 17497, 43864, 18038, 143
closes https://github.com/official-stockfish/Stockfish/pull/5534
bench 1613043
Created by updating output weights (256) and biases (8)
of the previous main net with values found with spsa around
101k / 120k games at 140+1.4.
264 spsa params: output weights and biases in nn-e8bac1c07a5a.nnue
A: 6000, alpha: 0.602, gamma: 0.101
weights: [-127, 127], c_end = 6
biases: [-8192, 8192], c_end = 64
Among the 264 params, 189 weights and all 8 biases were changed.
Changes in the weights:
- mean: -0.111 +/- 3.57
- range: [-8, 8]
Found with the same method as:
https://github.com/official-stockfish/Stockfish/pull/5459
Due to the original name (nn-ea8c9128c325.nnue) being too similar
to the previous main net (nn-e8bac1c07a5a.nnue) and creating confusion,
it was renamed by making non-functional changes to the .nnue file
the same way as past nets with:
https://github.com/linrock/nnue-namer
To verify that bench is the same and view the modified non-functional bytes:
```
echo -e "setoption name EvalFile value nn-ea8c9128c325.nnue\nbench" | ./stockfish
echo -e "setoption name EvalFile value nn-31337bea577c.nnue\nbench" | ./stockfish
cmp -l nn-ea8c9128c325.nnue nn-31337bea577c.nnue
diff <(xxd nn-ea8c9128c325.nnue) <(xxd nn-31337bea577c.nnue)
```
Passed STC:
https://tests.stockfishchess.org/tests/view/669564154ff211be9d4ec080
LLR: 2.93 (-2.94,2.94) <0.00,2.00>
Total: 57280 W: 15139 L: 14789 D: 27352
Ptnml(0-2): 209, 6685, 14522, 6995, 229
Passed LTC:
https://tests.stockfishchess.org/tests/view/669694204ff211be9d4ec1b4
LLR: 2.94 (-2.94,2.94) <0.50,2.50>
Total: 63030 W: 16093 L: 15720 D: 31217
Ptnml(0-2): 47, 6766, 17516, 7139, 47
closes https://github.com/official-stockfish/Stockfish/pull/5509
bench 1371485
Created by modifying L2 weights from the previous main net (nn-74f1d263ae9a.nnue)
with params found by spsa around 9k / 120k games at 120+1.2.
370 spsa params - L2 weights in nn-74f1d263ae9a.nnue where |val| >= 50
A: 6000, alpha: 0.602, gamma: 0.101
weights: [-127, 127], c_end = 6
To print the spsa params with nnue-pytorch:
```
import features
from serialize import NNUEReader
feature_set = features.get_feature_set_from_name("HalfKAv2_hm")
with open("nn-74f1d263ae9a.nnue", "rb") as f:
model = NNUEReader(f, feature_set).model
c_end = 6
for i in range(8):
for j in range(32):
for k in range(30):
value = int(model.layer_stacks.l2.weight[32 * i + j, k] * 64)
if abs(value) >= 50:
print(f"twoW[{i}][{j}][{k}],{value},-127,127,{c_end},0.0020")
```
Among the 370 params, 229 weights were changed.
avg change: 0.0961 ± 1.67
range: [-4, 3]
The number of weights changed, grouped by layer stack index,
shows more weights were modified in the lower piece count buckets:
[54, 52, 29, 23, 22, 18, 14, 17]
Found with the same method described in:
https://github.com/official-stockfish/Stockfish/pull/5459
Passed STC:
https://tests.stockfishchess.org/tests/view/668aec9a58083e5fd88239e7
LLR: 3.00 (-2.94,2.94) <0.00,2.00>
Total: 52384 W: 13569 L: 13226 D: 25589
Ptnml(0-2): 127, 6141, 13335, 6440, 149
Passed LTC:
https://tests.stockfishchess.org/tests/view/668af50658083e5fd8823a0b
LLR: 2.94 (-2.94,2.94) <0.50,2.50>
Total: 46974 W: 12006 L: 11668 D: 23300
Ptnml(0-2): 25, 4992, 13121, 5318, 31
closes https://github.com/official-stockfish/Stockfish/pull/5466
bench 1300471
Created by setting output weights (256) and biases (8) of the previous main net
nn-ddcfb9224cdb.nnue to values found around 12k / 120k spsa games at 120+1.2
This used modified fishtest dev workers to construct .nnue files from
spsa params, then load them with EvalFile when running tests:
https://github.com/linrock/fishtest/tree/spsa-file-modified-nnue/worker
Inspired by researching loading spsa params from files:
https://github.com/official-stockfish/fishtest/pull/1926
Scripts for modifying nnue files and preparing params:
https://github.com/linrock/nnue-pytorch/tree/no-gpu-modify-nnue
spsa params:
weights: [-127, 127], c_end = 6
biases: [-8192, 8192], c_end = 64
Example of reading output weights and biases from the previous main net using
nnue-pytorch and printing spsa params in a format compatible with fishtest:
```
import features
from serialize import NNUEReader
feature_set = features.get_feature_set_from_name("HalfKAv2_hm")
with open("nn-ddcfb9224cdb.nnue", "rb") as f:
model = NNUEReader(f, feature_set).model
c_end_weights = 6
c_end_biases = 64
for i in range(8):
for j in range(32):
value = round(int(model.layer_stacks.output.weight[i, j] * 600 * 16) / 127)
print(f"oW[{i}][{j}],{value},-127,127,{c_end_weights},0.0020")
for i in range(8):
value = int(model.layer_stacks.output.bias[i] * 600 * 16)
print(f"oB[{i}],{value},-8192,8192,{c_end_biases},0.0020")
```
For more info on spsa tuning params in nets:
https://github.com/official-stockfish/Stockfish/pull/5149https://github.com/official-stockfish/Stockfish/pull/5254
Passed STC:
https://tests.stockfishchess.org/tests/view/66894d64e59d990b103f8a37
LLR: 2.94 (-2.94,2.94) <0.00,2.00>
Total: 32000 W: 8443 L: 8137 D: 15420
Ptnml(0-2): 80, 3627, 8309, 3875, 109
Passed LTC:
https://tests.stockfishchess.org/tests/view/6689668ce59d990b103f8b8b
LLR: 2.94 (-2.94,2.94) <0.50,2.50>
Total: 172176 W: 43822 L: 43225 D: 85129
Ptnml(0-2): 97, 18821, 47633, 19462, 75
closes https://github.com/official-stockfish/Stockfish/pull/5459
bench 1120091
Created by first retraining the spsa-tuned main net `nn-ae6a388e4a1a.nnue` with:
- using v6-dd data without bestmove captures removed
- addition of T80 mar2024 data
- increasing loss by 20% when Q is too high
- torch.compile changes for marginal training speed gains
And then SPSA tuning weights of epoch 899 following methods described in:
https://github.com/official-stockfish/Stockfish/pull/5149
This net was reached at 92k out of 120k steps in this 70+0.7 th 7 SPSA tuning run:
https://tests.stockfishchess.org/tests/view/66413b7df9f4e8fc783c9bbb
Thanks to @Viren6 for suggesting usage of:
- c value 4 for the weights
- c value 128 for the biases
Scripts for automating applying fishtest spsa params to exporting tuned .nnue are in:
https://github.com/linrock/nnue-tools/tree/master/spsa
Before spsa tuning, epoch 899 was nn-f85738aefa84.nnue
https://tests.stockfishchess.org/tests/view/663e5c893a2f9702074bc167
After initially training with max-epoch 800, training was resumed with max-epoch 1000.
```
experiment-name: 3072--S11--more-data-v6-dd-t80-mar2024--see-ge0-20p-more-loss-high-q-sk28-l8
nnue-pytorch-branch: linrock/nnue-pytorch/3072-r21-skip-more-wdl-see-ge0-20p-more-loss-high-q-torch-compile-more
start-from-engine-test-net: False
start-from-model: /data/config/apr2024-3072/nn-ae6a388e4a1a.nnue
early-fen-skipping: 28
training-dataset:
/data/S11-mar2024/:
- leela96.v2.min.binpack
- test60-2021-11-12-novdec-12tb7p.v6-dd.min.binpack
- test78-2022-01-to-05-jantomay-16tb7p.v6-dd.min.binpack
- test80-2022-06-jun-16tb7p.v6-dd.min.binpack
- test80-2022-08-aug-16tb7p.v6-dd.min.binpack
- test80-2022-09-sep-16tb7p.v6-dd.min.binpack
- test80-2023-01-jan-16tb7p.v6-sk20.min.binpack
- test80-2023-02-feb-16tb7p.v6-sk20.min.binpack
- test80-2023-03-mar-2tb7p.v6-sk16.min.binpack
- test80-2023-04-apr-2tb7p.v6-sk16.min.binpack
- test80-2023-05-may-2tb7p.v6.min.binpack
# https://github.com/official-stockfish/Stockfish/pull/4782
- test80-2023-06-jun-2tb7p.binpack
- test80-2023-07-jul-2tb7p.binpack
# https://github.com/official-stockfish/Stockfish/pull/4972
- test80-2023-08-aug-2tb7p.v6.min.binpack
- test80-2023-09-sep-2tb7p.binpack
- test80-2023-10-oct-2tb7p.binpack
# S9 new data: https://github.com/official-stockfish/Stockfish/pull/5056
- test80-2023-11-nov-2tb7p.binpack
- test80-2023-12-dec-2tb7p.binpack
# S10 new data: https://github.com/official-stockfish/Stockfish/pull/5149
- test80-2024-01-jan-2tb7p.binpack
- test80-2024-02-feb-2tb7p.binpack
# S11 new data
- test80-2024-03-mar-2tb7p.binpack
/data/filt-v6-dd/:
- test77-dec2021-16tb7p-filter-v6-dd.binpack
- test78-juntosep2022-16tb7p-filter-v6-dd.binpack
- test79-apr2022-16tb7p-filter-v6-dd.binpack
- test79-may2022-16tb7p-filter-v6-dd.binpack
- test80-jul2022-16tb7p-filter-v6-dd.binpack
- test80-oct2022-16tb7p-filter-v6-dd.binpack
- test80-nov2022-16tb7p-filter-v6-dd.binpack
num-epochs: 1000
lr: 4.375e-4
gamma: 0.995
start-lambda: 0.8
end-lambda: 0.7
```
Training data can be found at:
https://robotmoon.com/nnue-training-data/
Local elo at 25k nodes per move:
nn-epoch899.nnue : 4.6 +/- 1.4
Passed STC:
https://tests.stockfishchess.org/tests/view/6645454893ce6da3e93b31ae
LLR: 2.95 (-2.94,2.94) <0.00,2.00>
Total: 95232 W: 24598 L: 24194 D: 46440
Ptnml(0-2): 294, 11215, 24180, 11647, 280
Passed LTC:
https://tests.stockfishchess.org/tests/view/6645522d93ce6da3e93b31df
LLR: 2.95 (-2.94,2.94) <0.50,2.50>
Total: 320544 W: 81432 L: 80524 D: 158588
Ptnml(0-2): 164, 35659, 87696, 36611, 142
closes https://github.com/official-stockfish/Stockfish/pull/5254
bench 1995552
For each thread persist an accumulator cache for the network, where each
cache contains multiple entries for each of the possible king squares.
When the accumulator needs to be refreshed, the cached entry is used to more
efficiently update the accumulator, instead of rebuilding it from scratch.
This idea, was first described by Luecx (author of Koivisto) and
is commonly referred to as "Finny Tables".
When the accumulator needs to be refreshed, instead of filling it with
biases and adding every piece from scratch, we...
1. Take the `AccumulatorRefreshEntry` associated with the new king bucket
2. Calculate the features to activate and deactivate (from differences
between bitboards in the entry and bitboards of the actual position)
3. Apply the updates on the refresh entry
4. Copy the content of the refresh entry accumulator to the accumulator
we were refreshing
5. Copy the bitboards from the position to the refresh entry, to match
the newly updated accumulator
Results at STC:
https://tests.stockfishchess.org/tests/view/662301573fe04ce4cefc1386
(first version)
https://tests.stockfishchess.org/tests/view/6627fa063fe04ce4cefc6560
(final)
Non-Regression between first and final:
https://tests.stockfishchess.org/tests/view/662801e33fe04ce4cefc660a
STC SMP:
https://tests.stockfishchess.org/tests/view/662808133fe04ce4cefc667c
closes https://github.com/official-stockfish/Stockfish/pull/5183
No functional change
Created by retraining the previous main net `nn-b1a57edbea57.nnue` with:
- some of the same options as before:
- ranger21, more WDL skipping, 15% more loss when Q is too high
- removal of the huge 514G pre-interleaved binpack
- removal of SF-generated dfrc data (dfrc99-16tb7p-filt-v2.min.binpack)
- interleaving many binpacks at training time
- training with some bestmove capture positions where SEE < 0
- increased usage of torch.compile to speed up training by up to 40%
```yaml
experiment-name: 2560--S10-dfrc0-to-dec2023-skip-more-wdl-15p-more-loss-high-q-see-ge0-sk28
nnue-pytorch-branch: linrock/nnue-pytorch/r21-more-wdl-skip-15p-more-loss-high-q-skip-see-ge0-torch-compile-more
start-from-engine-test-net: True
early-fen-skipping: 28
training-dataset:
# similar, not the exact same as:
# https://github.com/official-stockfish/Stockfish/pull/4635
- /data/S5-5af/leela96.v2.min.binpack
- /data/S5-5af/test60-2021-11-12-novdec-12tb7p.v6-dd.min.binpack
- /data/S5-5af/test77-2021-12-dec-16tb7p.v6-dd.min.binpack
- /data/S5-5af/test78-2022-01-to-05-jantomay-16tb7p.v6-dd.min.binpack
- /data/S5-5af/test78-2022-06-to-09-juntosep-16tb7p.v6-dd.min.binpack
- /data/S5-5af/test79-2022-04-apr-16tb7p.v6-dd.min.binpack
- /data/S5-5af/test79-2022-05-may-16tb7p.v6-dd.min.binpack
- /data/S5-5af/test80-2022-06-jun-16tb7p.v6-dd.min.unmin.binpack
- /data/S5-5af/test80-2022-07-jul-16tb7p.v6-dd.min.binpack
- /data/S5-5af/test80-2022-08-aug-16tb7p.v6-dd.min.binpack
- /data/S5-5af/test80-2022-09-sep-16tb7p.v6-dd.min.unmin.binpack
- /data/S5-5af/test80-2022-10-oct-16tb7p.v6-dd.min.binpack
- /data/S5-5af/test80-2022-11-nov-16tb7p.v6-dd.min.binpack
- /data/S5-5af/test80-2023-01-jan-16tb7p.v6-sk20.min.binpack
- /data/S5-5af/test80-2023-02-feb-16tb7p.v6-dd.min.binpack
- /data/S5-5af/test80-2023-03-mar-2tb7p.min.unmin.binpack
- /data/S5-5af/test80-2023-04-apr-2tb7p.binpack
- /data/S5-5af/test80-2023-05-may-2tb7p.min.dd.binpack
# https://github.com/official-stockfish/Stockfish/pull/4782
- /data/S6-1ee1aba5ed/test80-2023-06-jun-2tb7p.binpack
- /data/S6-1ee1aba5ed/test80-2023-07-jul-2tb7p.min.binpack
# https://github.com/official-stockfish/Stockfish/pull/4972
- /data/S8-baff1edbea57/test80-2023-08-aug-2tb7p.v6.min.binpack
- /data/S8-baff1edbea57/test80-2023-09-sep-2tb7p.binpack
- /data/S8-baff1edbea57/test80-2023-10-oct-2tb7p.binpack
# https://github.com/official-stockfish/Stockfish/pull/5056
- /data/S9-b1a57edbea57/test80-2023-11-nov-2tb7p.binpack
- /data/S9-b1a57edbea57/test80-2023-12-dec-2tb7p.binpack
num-epochs: 800
lr: 4.375e-4
gamma: 0.995
start-lambda: 1.0
end-lambda: 0.7
```
This particular net was reached at epoch 759. Use of more torch.compile decorators
in nnue-pytorch model.py than in the previous main net training run sped up training
by up to 40% on Tesla gpus when using recent pytorch compiled with cuda 12:
https://github.com/linrock/nnue-tools/blob/7fb9831/Dockerfile
Skipping positions with bestmove captures where static exchange evaluation is >= 0
is based on the implementation from Sopel's NNUE training & experimentation log:
https://docs.google.com/document/d/1gTlrr02qSNKiXNZ_SuO4-RjK4MXBiFlLE6jvNqqMkAY
Experiment 293 - only skip captures with see>=0
Positions with bestmove captures where score == 0 are always skipped for
compatibility with minimized binpacks, since the original minimizer sets
scores to 0 for slight improvements in compression.
The trainer branch used was:
https://github.com/linrock/nnue-pytorch/tree/r21-more-wdl-skip-15p-more-loss-high-q-skip-see-ge0-torch-compile-more
Binpacks were renamed to be sorted chronologically by default when sorted by name.
The binpack data are otherwise the same as binpacks with similar names in the prior
naming convention.
Training data can be found at:
https://robotmoon.com/nnue-training-data/
Passed STC:
https://tests.stockfishchess.org/tests/view/65e3ddd1f2ef6c733362ae5c
LLR: 2.92 (-2.94,2.94) <0.00,2.00>
Total: 149792 W: 39153 L: 38661 D: 71978
Ptnml(0-2): 675, 17586, 37905, 18032, 698
Passed LTC:
https://tests.stockfishchess.org/tests/view/65e4d91c416ecd92c162a69b
LLR: 2.94 (-2.94,2.94) <0.50,2.50>
Total: 64416 W: 16517 L: 16135 D: 31764
Ptnml(0-2): 38, 7218, 17313, 7602, 37
closes https://github.com/official-stockfish/Stockfish/pull/5090
Bench: 1373183
Created by retraining the previous main net `nn-baff1edbea57.nnue` with:
- some of the same options as before: ranger21, more WDL skipping
- the addition of T80 nov+dec 2023 data
- increasing loss by 15% when prediction is too high, up from 10%
- use of torch.compile to speed up training by over 25%
```yaml
experiment-name: 2560--S9-514G-T80-augtodec2023-more-wdl-skip-15p-more-loss-high-q-sk28
training-dataset:
# https://github.com/official-stockfish/Stockfish/pull/4782
- /data/S6-514G-1ee1aba5ed.binpack
- /data/test80-aug2023-2tb7p.v6.min.binpack
- /data/test80-sep2023-2tb7p.binpack
- /data/test80-oct2023-2tb7p.binpack
- /data/test80-nov2023-2tb7p.binpack
- /data/test80-dec2023-2tb7p.binpack
early-fen-skipping: 28
start-from-engine-test-net: True
nnue-pytorch-branch: linrock/nnue-pytorch/r21-more-wdl-skip-15p-more-loss-high-q-torch-compile
num-epochs: 1000
lr: 4.375e-4
gamma: 0.995
start-lambda: 1.0
end-lambda: 0.7
```
Epoch 819 trained with the above config led to this PR. Use of torch.compile
decorators in nnue-pytorch model.py was found to speed up training by at least
25% on Ampere gpus when using recent pytorch compiled with cuda 12:
https://catalog.ngc.nvidia.com/orgs/nvidia/containers/pytorch
See recent main net PRs for more info on
- ranger21 and more WDL skipping: https://github.com/official-stockfish/Stockfish/pull/4942
- increasing loss when Q is too high: https://github.com/official-stockfish/Stockfish/pull/4972
Training data can be found at:
https://robotmoon.com/nnue-training-data/
Passed STC:
https://tests.stockfishchess.org/tests/view/65cd76151d8e83c78bfd2f52
LLR: 2.98 (-2.94,2.94) <0.00,2.00>
Total: 78336 W: 20504 L: 20115 D: 37717
Ptnml(0-2): 317, 9225, 19721, 9562, 343
Passed LTC:
https://tests.stockfishchess.org/tests/view/65ce5be61d8e83c78bfd43e9
LLR: 2.95 (-2.94,2.94) <0.50,2.50>
Total: 41016 W: 10492 L: 10159 D: 20365
Ptnml(0-2): 22, 4533, 11071, 4854, 28
closes https://github.com/official-stockfish/Stockfish/pull/5056
Bench: 1351997
Also remove dead code, `rootSimpleEval` is no longer used since the introduction of dual net.
`iterBestValue` is also no longer used in evaluate and can be reduced to a local variable.
closes https://github.com/official-stockfish/Stockfish/pull/4979
No functional change
This aims to remove some of the annoying global structure which Stockfish has.
Overall there is no major elo regression to be expected.
Non regression SMP STC (paused, early version):
https://tests.stockfishchess.org/tests/view/65983d7979aa8af82b9608f1
LLR: 0.23 (-2.94,2.94) <-1.75,0.25>
Total: 76232 W: 19035 L: 19096 D: 38101
Ptnml(0-2): 92, 8735, 20515, 8690, 84
Non regression STC (early version):
https://tests.stockfishchess.org/tests/view/6595b3a479aa8af82b95da7f
LLR: 2.93 (-2.94,2.94) <-1.75,0.25>
Total: 185344 W: 47027 L: 46972 D: 91345
Ptnml(0-2): 571, 21285, 48943, 21264, 609
Non regression SMP STC:
https://tests.stockfishchess.org/tests/view/65a0715c79aa8af82b96b7e4
LLR: 2.94 (-2.94,2.94) <-1.75,0.25>
Total: 142936 W: 35761 L: 35662 D: 71513
Ptnml(0-2): 209, 16400, 38135, 16531, 193
These global structures/variables add hidden dependencies and allow data
to be mutable from where it shouldn't it be (i.e. options). They also
prevent Stockfish from internal selfplay, which would be a nice thing to
be able to do, i.e. instantiate two Stockfish instances and let them
play against each other. It will also allow us to make Stockfish a
library, which can be easier used on other platforms.
For consistency with the old search code, `thisThread` has been kept,
even though it is not strictly necessary anymore. This the first major
refactor of this kind (in recent time), and future changes are required,
to achieve the previously described goals. This includes cleaning up the
dependencies, transforming the network to be self contained and coming
up with a plan to deal with proper tablebase memory management (see
comments for more information on this).
The removal of these global structures has been discussed in parts with
Vondele and Sopel.
closes https://github.com/official-stockfish/Stockfish/pull/4968
No functional change
Created by retraining the previous main net nn-b1e55edbea57.nnue with:
- some of the same options as before: ranger21 optimizer, more WDL
skipping
- adding T80 aug filter-v6, sep, and oct 2023 data to the previous best
dataset
- increasing training loss for positions where predicted win rates were
higher than estimated match results from training data position scores
```yaml
experiment-name: 2560--S8-r21-more-wdl-skip-10p-more-loss-high-q-sk28
training-dataset:
# https://github.com/official-stockfish/Stockfish/pull/4782
- /data/S6-1ee1aba5ed.binpack
- /data/test80-aug2023-2tb7p.v6.min.binpack
- /data/test80-sep2023-2tb7p.binpack
- /data/test80-oct2023-2tb7p.binpack
early-fen-skipping: 28
start-from-engine-test-net: True
nnue-pytorch-branch: linrock/nnue-pytorch/r21-more-wdl-skip-10p-more-loss-high-q
num-epochs: 1000
lr: 4.375e-4
gamma: 0.995
start-lambda: 1.0
end-lambda: 0.7
```
Training data can be found at:
https://robotmoon.com/nnue-training-data/
Training loss was increased by 10% for positions where predicted win
rates were higher than suggested by the win rate model based on the
training data, by multiplying with: ((qf > pt) * 0.1 + 1). This was a
variant of experiments from Sopel's NNUE training & experimentation log:
https://docs.google.com/document/d/1gTlrr02qSNKiXNZ_SuO4-RjK4MXBiFlLE6jvNqqMkAY
Experiment 302 - increase loss when prediction too high, vondele’s idea
Experiment 309 - increase loss when prediction too high, normalize in a
batch
Passed STC:
https://tests.stockfishchess.org/tests/view/6597a21c79aa8af82b95fd5c
LLR: 2.93 (-2.94,2.94) <0.00,2.00>
Total: 148320 W: 37960 L: 37475 D: 72885
Ptnml(0-2): 542, 17565, 37383, 18206, 464
Passed LTC:
https://tests.stockfishchess.org/tests/view/659834a679aa8af82b960845
LLR: 2.94 (-2.94,2.94) <0.50,2.50>
Total: 55188 W: 13955 L: 13592 D: 27641
Ptnml(0-2): 34, 6162, 14834, 6535, 29
closes https://github.com/official-stockfish/Stockfish/pull/4972
Bench: 1219824
Created by training an L1-128 net from scratch with a wider range of
evals in the training data and wld-fen-skipping disabled during
training. The differences in this training data compared to the first
dual nnue PR are:
- removal of all positions with 3 pieces
- when piece count >= 16, keep positions with simple eval above 750
- when piece count < 16, remove positions with simple eval above 3000
The asymmetric data filtering was meant to flatten the training data
piece count distribution, which was previously heavily skewed towards
positions with low piece counts.
Additionally, the simple eval range where the smallnet is used was
widened to cover more positions previously evaluated by the big net and
simple eval.
```yaml
experiment-name: 128--S1-hse-S7-v4-S3-v1-no-wld-skip
training-dataset:
- /data/hse/S3/leela96-filt-v2.min.high-simple-eval-1k.binpack
- /data/hse/S3/dfrc99-16tb7p-eval-filt-v2.min.high-simple-eval-1k.binpack
- /data/hse/S3/test80-apr2022-16tb7p.min.high-simple-eval-1k.binpack
- /data/hse/S7/test60-2020-2tb7p.v6-3072.high-simple-eval-v4.binpack
- /data/hse/S7/test60-novdec2021-12tb7p-filter-v6-dd.min-mar2023.unmin.high-simple-eval-v4.binpack
- /data/hse/S7/test77-nov2021-2tb7p.v6-3072.min.high-simple-eval-v4.binpack
- /data/hse/S7/test77-dec2021-16tb7p-filter-v6-dd.min-mar2023.unmin.high-simple-eval-v4.binpack
- /data/hse/S7/test77-jan2022-2tb7p.high-simple-eval-v4.binpack
- /data/hse/S7/test78-jantomay2022-16tb7p-filter-v6-dd.min-mar2023.unmin.high-simple-eval-v4.binpack
- /data/hse/S7/test78-juntosep2022-16tb7p-filter-v6-dd.min-mar2023.unmin.high-simple-eval-v4.binpack
- /data/hse/S7/test79-apr2022-16tb7p-filter-v6-dd.min-mar2023.unmin.high-simple-eval-v4.binpack
- /data/hse/S7/test79-may2022-16tb7p-filter-v6-dd.min-mar2023.unmin.high-simple-eval-v4.binpack
- /data/hse/S7/test80-may2022-16tb7p.high-simple-eval-v4.binpack
- /data/hse/S7/test80-jun2022-16tb7p-filter-v6-dd.min-mar2023.unmin.high-simple-eval-v4.binpack
- /data/hse/S7/test80-jul2022-16tb7p.v6-dd.min.high-simple-eval-v4.binpack
- /data/hse/S7/test80-aug2022-16tb7p-filter-v6-dd.min-mar2023.unmin.high-simple-eval-v4.binpack
- /data/hse/S7/test80-sep2022-16tb7p-filter-v6-dd.min-mar2023.unmin.high-simple-eval-v4.binpack
- /data/hse/S7/test80-oct2022-16tb7p.v6-dd.high-simple-eval-v4.binpack
- /data/hse/S7/test80-nov2022-16tb7p-v6-dd.min.high-simple-eval-v4.binpack
- /data/hse/S7/test80-jan2023-3of3-16tb7p-filter-v6-dd.min-mar2023.unmin.high-simple-eval-v4.binpack
- /data/hse/S7/test80-feb2023-16tb7p-filter-v6-dd.min-mar2023.unmin.high-simple-eval-v4.binpack
- /data/hse/S7/test80-mar2023-2tb7p.v6-sk16.min.high-simple-eval-v4.binpack
- /data/hse/S7/test80-apr2023-2tb7p-filter-v6-sk16.min.high-simple-eval-v4.binpack
- /data/hse/S7/test80-may2023-2tb7p.v6.min.high-simple-eval-v4.binpack
- /data/hse/S7/test80-jun2023-2tb7p.v6-3072.min.high-simple-eval-v4.binpack
- /data/hse/S7/test80-jul2023-2tb7p.v6-3072.min.high-simple-eval-v4.binpack
- /data/hse/S7/test80-aug2023-2tb7p.v6.min.high-simple-eval-v4.binpack
- /data/hse/S7/test80-sep2023-2tb7p.high-simple-eval-v4.binpack
- /data/hse/S7/test80-oct2023-2tb7p.high-simple-eval-v4.binpack
wld-fen-skipping: False
start-from-engine-test-net: False
nnue-pytorch-branch: linrock/nnue-pytorch/L1-128
engine-test-branch: linrock/Stockfish/L1-128-nolazy
engine-base-branch: linrock/Stockfish/L1-128
num-epochs: 500
start-lambda: 1.0
end-lambda: 1.0
```
Experiment yaml configs converted to easy_train.sh commands with:
https://github.com/linrock/nnue-tools/blob/4339954/yaml_easy_train.py
Binpacks interleaved at training time with:
https://github.com/official-stockfish/nnue-pytorch/pull/259
FT weights permuted with 10k positions from fishpack32.binpack with:
https://github.com/official-stockfish/nnue-pytorch/pull/254
Data filtered for high simple eval positions (v4) with:
https://github.com/linrock/Stockfish/blob/b9c8440/src/tools/transform.cpp#L640-L675
Training data can be found at:
https://robotmoon.com/nnue-training-data/
Local elo at 25k nodes per move of
L1-128 smallnet (nnue-only eval) vs. L1-128 trained on standard S1 data:
nn-epoch319.nnue : -241.7 +/- 3.2
Passed STC vs. 36db936:
https://tests.stockfishchess.org/tests/view/6576b3484d789acf40aabbfe
LLR: 2.94 (-2.94,2.94) <0.00,2.00>
Total: 21920 W: 5680 L: 5381 D: 10859
Ptnml(0-2): 82, 2488, 5520, 2789, 81
Passed LTC vs. DualNNUE #4915:
https://tests.stockfishchess.org/tests/view/65775c034d789acf40aac7e3
LLR: 2.95 (-2.94,2.94) <0.50,2.50>
Total: 147606 W: 36619 L: 36063 D: 74924
Ptnml(0-2): 98, 16591, 39891, 17103, 120
closes https://github.com/official-stockfish/Stockfish/pull/4919
Bench: 1438336
The primary rationale behind this lies in the fact that enums were not
originally designed to be employed in the manner we currently utilize them.
The Value enum was used like a type alias throughout the code and was often
misused. Furthermore, changing the underlying size of the enum to int16_t broke
everything, mostly because of the operator overloads for the Value enum, were
causing data to be truncated. Since Value is now a type alias, the operator
overloads are no longer required.
Passed Non-Regression STC:
https://tests.stockfishchess.org/tests/view/6593b8bb79aa8af82b95b401
LLR: 2.95 (-2.94,2.94) <-1.75,0.25>
Total: 235296 W: 59919 L: 59917 D: 115460
Ptnml(0-2): 743, 27085, 62054, 26959, 807
closes https://github.com/official-stockfish/Stockfish/pull/4960
No functional change
