Print in the log, something like:
I0514 12:51:32.724236 14467 mcts_engine.cc:157] 1th move(b): dp, winrate=44.110905%, N=654, Q=-0.117782, p=0.079232, v=-0.116534, cost 39042.679688ms, sims=7132, height=11, avg_height=5.782244, global_step=639200
It is possible to display the PV (variation of move path with
continuation of the moves : main move path, second move path,
third move path)
An easy way to do that is for example to increase verbose level,
for example --logtostderr --v=1
(for windows the syntax is different, see
FAQ question for details
result is something like this (there are 30000 simulations in that example) :
mcts_debugger.cc:43] ========== debug info for 1th move(b) begin ==========
mcts_debugger.cc:44] main move path: dd(2959,-0.12,0.08,-0.11),pp(762,0.12,0.22,0.11),dq(142,-0.12,0.16,-0.12),pd(107,0.12,0.70,0.11),nq(21,-0.12,0.17,-0.13),co(6,0.13,0.24,0.14),fp(3,-0.13,0.36,-0.13),cc(2,0.13,0.31,0.11),cd(1,-0.14,0.92,-0.14),dk(0,-nan,0.00,nan)
mcts_debugger.cc:45] second move path: pd(2930,-0.12,0.08,-0.11),dp(745,0.12,0.22,0.11),cd(136,-0.12,0.16,-0.12),pp(102,0.12,0.70,0.11),cn(19,-0.12,0.16,-0.13),ec(6,0.13,0.24,0.13),df(3,-0.13,0.36,-0.11),qc(2,0.13,0.32,0.12),pc(1,-0.14,0.90,-0.14),dj(0,-nan,0.00,nan)
mcts_debugger.cc:46] third move path: dp(2911,-0.12,0.08,-0.12),pd(765,0.12,0.23,0.10),qp(144,-0.12,0.16,-0.12),dd(109,0.12,0.70,0.11),qf(20,-0.12,0.16,-0.12),oq(6,0.13,0.23,0.12),pn(3,-0.13,0.36,-0.13),cq(2,0.13,0.31,0.13),dq(1,-0.14,0.91,-0.14),im(0,-nan,0.00,nan)
mcts_debugger.cc:140] dd: N=2959, W=-346.395, Q=-0.117065, p=0.0802264, v=-0.113951
mcts_debugger.cc:140] pd: N=2930, W=-343.235, Q=-0.117145, p=0.0792342, v=-0.113951
mcts_debugger.cc:140] dp: N=2911, W=-340.83, Q=-0.117083, p=0.0783167, v=-0.116534
mcts_debugger.cc:140] pp: N=2840, W=-332.817, Q=-0.117189, p=0.0770947, v=-0.120119
mcts_debugger.cc:140] dq: N=2363, W=-279.487, Q=-0.118277, p=0.0700366, v=-0.117204
mcts_debugger.cc:140] cd: N=2330, W=-275.263, Q=-0.118138, p=0.0683769, v=-0.116526
mcts_debugger.cc:140] qp: N=2321, W=-273.97, Q=-0.118039, p=0.0676945, v=-0.12072
mcts_debugger.cc:140] dc: N=2295, W=-271.415, Q=-0.118264, p=0.0683262, v=-0.116526
mcts_debugger.cc:140] cp: N=2262, W=-267.68, Q=-0.118338, p=0.0675763, v=-0.12072
mcts_debugger.cc:140] qd: N=2261, W=-268.092, Q=-0.118572, p=0.0687082, v=-0.118771
mcts_debugger.cc:50] model global step: 639200
mcts_debugger.cc:51] ========== debug info for 1th move(b) end ==========
It is also possible to develop all the tree by adding these lines to your config file :
debugger {
print_tree_depth: 20
print_tree_width: 3
}
In this example, a tree of 20 depth 3 width moves will be printed
See this document
Passing --v=0 to mcts_main will turn off many debug log.
Moreover, --minloglevel=1 and --minloglevel=2 could disable
INFO log and WARNING log.
Or, if you just don't want to log to stderr, replace --logtostderr
to --log_dir={log_dir}, then you could read your log from
{log_dir}/mcts_main.INFO.
For windows,
- in config file,
you need to write path with / and not \ in the config file
.conf, for example :
model_config {
train_dir: "c:/users/amd2018/Downloads/PhoenixGo/ckpt"
- in cmd.exe,
Here you need to write paths with \ and not / in command line,
it is the opposite for command line.
Also command format on command line needs a space and not a =, for
example :
mcts_main.exe --gtp --config_path C:\Users\amd2018\Downloads\PhoenixGo\etc\mcts_1gpu_notensorrt.conf
or if you want to show the PV, you need to remove the = too,
for example :
mcts_main.exe --gtp --config_path C:\Users\amd2018\Downloads\PhoenixGo\etc\mcts_1gpu_notensorrt.conf --logtostderr --v 1
See next point below :
This fix works for all systems : Linux, Mac, Windows, only the name of the ckpt file changes. Modify your config file and write the full path of your ckpt directory, for example for linux :
model_config {
train_dir: "/home/amd2018/PhoenixGo/ckpt"
for example, for windows :
model_config {
train_dir: "c:/users/amd2018/Downloads/PhoenixGo/ckpt"
if you use tensorRT (linux only, and compatible nvidia GPU only), also change path of tensorRT, for example :
model_config {
train_dir: "/home/amd2018/PhoenixGo/ckpt/"
enable_tensorrt: 1
tensorrt_model_path: "/home/amd2018/test/PhoenixGo/ckpt/zero.ckpt-20b-v1.FP32.PLAN"
}
Setting GTP engine in Sabaki's menu: Engines -> Manage Engines,
fill Path with path of start.sh.
Click Engines -> Attach to use the engine in your game.
See also #22.
Modify timeout_ms_per_step in your config file.
For example 5000 is 5 seconds per move.
Modify your config file. early_stop, unstable_overtime,
behind_overtime andtime_control are options that affect the
search time, remove them if exist then
each move will cost constant time/simulations.
- In command line, you can add the command flag
--init_moves, for example :
$ bazel-bin/mcts/mcts_main --gtp --config_path=etc/mcts_1gpu.conf --logtostderr --v=1 --init_moves="dp,zz,qe,zz,dd,qp,zz,qg,zz,pc,oq,pn,od,pd,pe,oe,of,ne,pg,md,pi"
In this example we start at move 22, see also #83 for screenshots
For windows the syntax is different, see FAQ question for details
You can also use the GTP command undo to cancel last move only,
the undo command can be repeated many times if want
- In non command line, you can open your .sgf file in a sgf viewer like Sabaki for example
GPU is much faster to compute than CPU (but only nvidia GPU are supported)
TensorRT also increases significantly the speed of computation, but it is only available for linux with a compatible nvidia GPU
Bigger batch size significantly increases the speed of the computation, but a bigger batch size puts a bigger burden on the computation device (in case it is the GPU, higher GPU load, higher VRAM usage), increase it only if your computation device can handle it
Some independent speed benchmarks have been run, they are available in the docs :
-
for GTX 1060 75W (75w power limit) : benchmark testing batch size from 4 to 64, tree size up to 2000M, max children up to 512, with tensorRT ON and OFF
-
for Tesla V100 : benchmark testing batch size from 4 to 128, 4 to 12 vcpu, no tensorrt
Add these lines in your config:
time_control {
enable: 1
c_denom: 20
c_maxply: 40
reserved_time: 1.0
}
c_denomandc_maxplyare parameters for deciding how to use the "main time".reserved_timeis how many seconds should reserved (for network latency) in "byo-yomi time".
Some GTP commands are not supported by PhoenixGo, for example the
showboard command.
To know supported GTP commands, start phoenixgo in GTP mode and enter
the GTP command list_commands
Result as of today is :
version
protocol_version
list_commands
quit
clear_board
boardsize
komi
time_settings
time_left
place_free_handicap
set_free_handicap
play
genmove
final_score
get_debug_info
get_last_move_debug_info
undo
If you use unsupported commands the engine will not work. Make sure your GTP tool does not communicate with PhoenixGo with unsupported GTP commands.
For example, for gtp2ogs server
command line GTP tool, you need to edit the file gtp2ogs.js and manually
remove the existing showboard line if it
is not already done,
see
With the default settings, the only komi value supported is only 7.5, with chinese rules only.If it is not automated, you need to manually set komi value to 7.5 with chinese rules.
If you want to implement PhoenixGo to a server where players play with different komi values (for example 6.5, 0.5, 85.5, 200.5,etc), you need to force komi to 7.5 for the players.
If it is not possible, there is a workaround you can use : configure you GTP tool to tell PhoenixGo engine that the komi for the game is 7.5 even if it is not true
if you do that, the game will not be scored correctly because PhoenixGo will think that the komi is 7.5 while the real komi is different, but at least PhoenixGo will be able to play the game.
- need CUDA 10.0 or higher (so currently, only linux is supported, or windows with your own building)
- If you compile PhoenixGo, it has been tested to work on linux here with CUDA 10.0, cudnn 7.4.2, ubuntu 18.04.
- However currently there is no tensorRT support for PhoenixGo (RTX cards require tensorRT 5.x or more, and this also requires tensorflow 1.9 or more), which is currently not supported by PhoenixGo)
- are compatible with cuda 9.0 and higher (it is recommended to use latest version when possible), cudnn 7.1.x and higher (x is any number)
- has been tested to work successfully on windows
- to use TensorRT with V100, you need to manually build TensorRT model on V100. See: #75 for how to build TensorRT model.
You can find a speed benchmark for Tesla V100 in FAQ question
You may find hard to download, install, run bazel
If that's the case, and if you are using ubuntu or similar operating system, you can use the all-in-one command below instead of the main README commands
Read the main README for explanations and instructions : the all-in-one command below only saves you the time to find how to install bazel and run all the bazel commands one by one, but you still have to read the main README for explanations
The all-in-one command below has been tested to run successfully on ubuntu 16.04 LTS and 18.04 LTS
sudo apt-get -y install pkg-config zip g++ zlib1g-dev unzip python git && \
git clone https://github.com/Tencent/PhoenixGo.git && \
cd PhoenixGo && \
wget https://github.com/bazelbuild/bazel/releases/download/0.19.2/bazel-0.19.2-installer-linux-x86_64.sh && \
chmod +x bazel-0.19.2-installer-linux-x86_64.sh && \
./bazel-0.19.2-installer-linux-x86_64.sh --user && \
echo 'export PATH="$PATH:$HOME/bin"' >> ~/.bashrc && source ~/.bashrc && \
sudo ldconfig && \
wget https://github.com/Tencent/PhoenixGo/releases/download/trained-network-20b-v1/trained-network-20b-v1.tar.gz && \
tar xvzf trained-network-20b-v1.tar.gz && \
rm trained-network-20b-v1.tar.gz bazel-0.19.2-installer-linux-x86_64.sh && \
./configure && \
bazel build //mcts:mcts_main
This all-in-one command will :
- Download and install bazel and PhoenixGo dependencies for ubuntu
and similar systems (need
apt-get) - Clone PhoenixGo from github
- Download and install bazel 0.19.2
- Do the post-install of bazel
- Download and extract trained network (ckpt)
- Cleanup : trained network archive and remove bazel installer
- Run
./configure: at this step, you have to confifure bazel same as explained in main README - When configure is finished, start building automatically, same as explained in main README
If you built with bazel, see : Most common path errors during cuda/cudnn install and bazel configure
See also minimalist bazel configure for an example of build configure
If you are still getting errors, try using an older version of bazel. For example bazel 0.20.0 is known to cause issues, and bazel 0.19.2 is known good
During the bazel building, there are many options that can are not required and can be disabled
This will reduce building time and will have smaller size after the building, see minimalist bazel configure and #76
Increasing batch size in the config file makes the engine compute faster, as explained earlier in FAQ question
However, with default building, you cannot use batch size higher than 4 with tensorRT To increase batch size for example to 32 with tensorRT enabled, you need to build tensorrt model with bazel, See : #75
assuming you installed PhoenixGo in home directory
(~ or /home/yourusername/)
# clean with bazel
cd ~/PhoenixGo && bazel clean
# remove PhoenixGo local directory
sudo rm -rf ~/PhoenixGo
# remove any remaining bazel file
sudo rm -rf ~/.cache/bazel
This will free a few GB (arround 3-6 GB depending on your installation settings)