benediamond / leela-chess Goto Github PK

View Code? Open in Web Editor NEW

14.0 14.0 3.0 206 KB

A chess adaption of GCP's Leela Zero

C++ 95.48% QMake 0.04% Python 3.95% Makefile 0.24% C 0.29%

leela-chess's People

Contributors

Stargazers

Watchers

Forkers

akababa syys96 liujh168

leela-chess's Issues

Write the move_lookup map

leela-chess/src/UCTSearch.h

Line 72 in 3079185

static std::unordered_map<Move, int> move_lookup;

The dictionary defined here will be used during the MCTS search, to use Moves to index into the Network's results, i.e. in:

leela-chess/src/UCTNode.cpp

Lines 116 to 120 in 3079185

 for (Move move : moves) { 

 auto node = raw_netlist.first[UCTSearch::move_lookup[move]]; 

 nodelist.emplace_back(node.first, move); 

 legal_sum += node.first; 

 }

I've already done something like this in my Python repository, see https://github.com/benediamond/chess-alpha-zero/blob/588b4d4f81978805be6a178faf8469948d891721/src/chess_zero/config.py#L19-L86

Integrating UCTSearch with scripts

Though the UCTSearch and UCTNode classes should be all but functional as is (except for adapting the Network, see Issue #1) I haven't integrated these with the scripts handling the command line interface and autogp and so on. As it stands, I have just dumped into main.cpp the beginnings of how to set everything up:

leela-chess/src/main.cpp

Lines 54 to 59 in c92efdb

 Position game; 

 StateListPtr states(new std::deque<StateInfo>(1)); 

 game.set(StartFEN, &states->back()); 

 auto search = std::make_unique<UCTSearch>(game, states); 

 search->think();

But this should be handled by a smoother utility that orchestrates games of self-play, etc. (presumably this code is already written). Note: once this is done, the moves encountered should be properly emplace_backed onto states, as in stockfish:

https://github.com/official-stockfish/Stockfish/blob/2198cd0524574f0d9df8c0ec9aaf14ad8c94402b/src/uci.cpp#L69-L78

Adapting Network and OpenCL

The get_scored_moves_internal routine of Network.cpp, together with the entirety of OpenCL.cpp, are littered with constants specific to Go (19, 391, 392, etc.). Presumably, this is just a matter of changing a few numbers. Yet the whole thing is fairly complicated, so I have held off on doing this for the time being. Here are a few examples:

leela-chess/src/Network.cpp

Lines 356 to 360 in 8675bd2

 Network::Netresult Network::get_scored_moves_internal(Position* pos, NNPlanes& planes) { //--must fix this! 

 assert(INPUT_CHANNELS == planes.size()); 

 constexpr int width = 19; 

 constexpr int height = 19; 

 const auto convolve_channels = conv_pol_w.size() / conv_pol_b.size();

leela-chess/src/Network.cpp

Lines 383 to 393 in 8675bd2

 convolve<1, 2>(output_data, conv_pol_w, conv_pol_b, policy_data_1); 

 batchnorm<2, 361>(policy_data_1, bn_pol_w1, bn_pol_w2, policy_data_2); 

 innerproduct<2*361, 362>(policy_data_2, ip_pol_w, ip_pol_b, policy_out); 

 softmax(policy_out, softmax_data, cfg_softmax_temp); 

 std::vector<float>& outputs = softmax_data; 

 // Now get the score 

 convolve<1, 1>(output_data, conv_val_w, conv_val_b, value_data_1); 

 batchnorm<1, 361>(value_data_1, bn_val_w1, bn_val_w2, value_data_2); 

 innerproduct<361, 256>(value_data_2, ip1_val_w, ip1_val_b, winrate_data); 

 innerproduct<256, 1>(winrate_data, ip2_val_w, ip2_val_b, winrate_out);

leela-chess/src/OpenCL.cpp

Lines 559 to 572 in 3079185

 size_t stripSize; 

 int rowTileSize; 

 int rowTiles; 

 if (filter_size == 3) { 

 stripSize = filter_size * (width + (filter_size - 1)) * sizeof(float); 

 rowTiles = cfg_rowtiles; 

 rowTileSize = (19 + rowTiles - 1) / rowTiles; 

 } else { 

 assert(filter_size == 1); 

 stripSize = width * sizeof(float); 

 rowTiles = 19; 

 rowTileSize = 1; 

 assert(channelGroup == 8); // hardcoded in kernel 

 }

Supervised learning

Hi, @benediamond. I like this approximation very much. I'll try to help in my possibilities.

I don't know yet the source code of this project too much but I think we should implement a supervised learning script able to create input files for checking the model is able to converge.

I thought about importing directly the python script we already have in order just to create the input files for feeding the model. How do we need to create this files? what's its format?

benediamond / leela-chess Goto Github PK

leela-chess's People

Contributors

Stargazers

Watchers

Forkers

leela-chess's Issues

Write the move_lookup map

Integrating UCTSearch with scripts

Adapting Network and OpenCL

Supervised learning

Recommend Projects

React

Vue.js

Typescript

TensorFlow

Django

Laravel

D3

Recommend Topics

javascript

web

server

Machine learning

Visualization

Game

Recommend Org

Facebook

Microsoft

Google

Alibaba

D3

Tencent

Jobs

	for (Move move : moves) {
	auto node = raw_netlist.first[UCTSearch::move_lookup[move]];
	nodelist.emplace_back(node.first, move);
	legal_sum += node.first;
	}

	Position game;
	StateListPtr states(new std::deque<StateInfo>(1));
	game.set(StartFEN, &states->back());

	auto search = std::make_unique<UCTSearch>(game, states);
	search->think();

	Network::Netresult Network::get_scored_moves_internal(Position* pos, NNPlanes& planes) { //--must fix this!
	assert(INPUT_CHANNELS == planes.size());
	constexpr int width = 19;
	constexpr int height = 19;
	const auto convolve_channels = conv_pol_w.size() / conv_pol_b.size();

	convolve<1, 2>(output_data, conv_pol_w, conv_pol_b, policy_data_1);
	batchnorm<2, 361>(policy_data_1, bn_pol_w1, bn_pol_w2, policy_data_2);
	innerproduct<2*361, 362>(policy_data_2, ip_pol_w, ip_pol_b, policy_out);
	softmax(policy_out, softmax_data, cfg_softmax_temp);
	std::vector<float>& outputs = softmax_data;

	// Now get the score
	convolve<1, 1>(output_data, conv_val_w, conv_val_b, value_data_1);
	batchnorm<1, 361>(value_data_1, bn_val_w1, bn_val_w2, value_data_2);
	innerproduct<361, 256>(value_data_2, ip1_val_w, ip1_val_b, winrate_data);
	innerproduct<256, 1>(winrate_data, ip2_val_w, ip2_val_b, winrate_out);

	size_t stripSize;
	int rowTileSize;
	int rowTiles;
	if (filter_size == 3) {
	stripSize = filter_size * (width + (filter_size - 1)) * sizeof(float);
	rowTiles = cfg_rowtiles;
	rowTileSize = (19 + rowTiles - 1) / rowTiles;
	} else {
	assert(filter_size == 1);
	stripSize = width * sizeof(float);
	rowTiles = 19;
	rowTileSize = 1;
	assert(channelGroup == 8); // hardcoded in kernel
	}