Did the newest code can implement stmkomi?

[pangafu]

I notice in patch-39
std::vector Network::gather_features
change to
std::vector<uint8_t> Network::gather_features

So if there is a way to implement stmkomi code? thanks~

[pangafu]

I think gpu worker and batchsize seperate maybe greater for stm komi, can you implement the stm komi to the newest code?

[alreadydone]

In order to change the stm (color) planes in patch-39, you need to modify the fourth and fifth parameters of forward0 (btm and wtm):

lz/src/OpenCLScheduler.cpp

Line 338 in 70a8aff

const float btm, const float wtm,

lz/src/Network.cpp

Lines 815 to 816 in 70a8aff

	auto tomove = state->get_to_move();
	m_forward->forward0(gnum, i, gather_features(state, symmetry), !tomove, tomove, tomove, symmetry, result);

lz/src/UCTSearch.cpp

Line 416 in 70a8aff

m_network.get_output0(gnum, i, bd, Network::Ensemble::RANDOM_SYMMETRY);

lz/src/UCTSearch.cpp

Line 957 in 70a8aff

play_simulation(std::move(rootstate), root, pending_counter, gnum, i);

When I get a chance I'll try to implement dynamic komi over patch-39, and you are certainly welcome to implement it in the meantime.

Regarding workers and batchsizes: the official branch uses search threads that can send positions to any of the GPUs, while my branch (patch-39 etc.) assigns dedicated worker threads for each GPU and allows the number of worker threads and the batch size configured for each GPU separately. My approach reduces contention between threads and allows a higher n/s to be achieved with many GPUs, but I am not seeing why it might be greater for stm komi.

[pangafu]

The offical branch search too wide when batchsize is large, and stm komi is not well training, many low pn search position will cause bad value, so maybe limit worker number will make search more reasonable.

Wait for your stm komi code~ thanks a lot!

[pangafu]

And in my test, in patch-39, when use offical weight, if increase worker number upper than 2 (such as 3), the gpu usage will increase, pos also increase, but can't beat woker number = 2.

So I think think the weight now seem has many fault value in low pn position, because pn is low mean the weight not well training in that way, search too wide maybe mean more fault.

[pangafu]

Also in stm komi test, when I use 4 or 8 gpu, batchsize > 8 in offical branch stm komi code, the handicap capability is lower than 1 gpu, batchsize = 2/3 run in long time.

So maybe stm komi not suitable search that wide.