Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Update resign_analysis script for chess. #427

Merged
merged 3 commits into from
Apr 27, 2018
Merged

Update resign_analysis script for chess. #427

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
73b3093
Update resign_analysis script for chess.
killerducky Apr 24, 2018
5d3991a
Fix Incorrect resigns to include draws.
killerducky Apr 25, 2018
9d365ba
Fix bug in correct_resigns.
killerducky Apr 25, 2018
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
190 changes: 51 additions & 139 deletions scripts/resign_analysis/resign_analysis.py
100644 → 100755
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -17,156 +17,68 @@
# along with Leela Zero. If not, see <http://www.gnu.org/licenses/>.

import sys
import re
import argparse

class GameStats:
def __init__(self, filename):
self.filename = filename
self.total_moves = None
self.resign_movenum = None
self.resign_type = None # "Correct", "Incorrect"
self.winner = None

class TotalStats:
def __init__(self):
self.num_games = 0
self.no_resign_count = 0
self.correct_resign_count = 0
self.incorrect_resign_count = 0
self.game_len_sum = 0
self.resigned_game_len_sum = 0
def calcOverall(self, b, w):
self.num_games = b.num_games + w.num_games
self.no_resign_count = b.no_resign_count + w.no_resign_count
self.correct_resign_count = b.correct_resign_count + w.correct_resign_count
self.incorrect_resign_count = b.incorrect_resign_count + w.incorrect_resign_count
self.game_len_sum = b.game_len_sum + w.game_len_sum
self.resigned_game_len_sum = b.resigned_game_len_sum + w.resigned_game_len_sum

def to_move_str(to_move):
if (to_move): return "W"
else: return "B"

def parseGames(filenames, resignrate, verbose):
gsd = {}
def parseGames(filenames, resignrate):
#print("Begin Analyze resign rate {:0.0f}%".format(resignrate))
games = len(filenames)
correct_resigns = 0
incorrect_resigns = 0
total_plys = 0
resign_plys = 0
for filename in filenames:
training_filename = filename.replace(".debug", "")
gs = GameStats(filename)
gsd[filename] = gs
with open(filename) as fh, open(training_filename) as tfh:
movenum = 0
version = fh.readline()
assert version == "1\n"
while 1:
movenum += 1
for _ in range(16):
line = tfh.readline() # Board input planes
if not line: break
to_move = int(tfh.readline()) # 0 = black, 1 = white
policy_weights = tfh.readline() # 361 moves + 1 pass
side_to_move_won = int(tfh.readline()) # 1 = side to move won, -1 = lost
if not gs.winner:
if side_to_move_won == 1: gs.winner = to_move
else : gs.winner = 1 - to_move
(netwinrate, root_uctwinrate, child_uctwinrate, bestmovevisits) = fh.readline().split()
netwinrate = float(netwinrate)
root_uctwinrate = float(root_uctwinrate)
child_uctwinrate = float(child_uctwinrate)
bestmovevisits = int(bestmovevisits)
if side_to_move_won == 1:
if verbose >= 2: print("+", to_move, movenum, netwinrate, child_uctwinrate, bestmovevisits)
if not gs.resign_type and child_uctwinrate < resignrate:
if verbose >= 1:
print("Incorrect resign -- %s rr=%0.3f wr=%0.3f winner=%s movenum=%d" %
(filename, resignrate, child_uctwinrate, to_move_str(to_move), movenum))
if verbose >= 2:
print("policy_weights", policy_weights)
gs.resign_type = "Incorrect"
gs.resign_movenum = movenum
else:
if verbose >= 2: print("-", to_move, movenum, netwinrate, child_uctwinrate, bestmovevisits)
if not gs.resign_type and child_uctwinrate < resignrate:
if verbose >= 2: print("Correct resign -- %s" % (filename))
gs.resign_type = "Correct"
gs.resign_movenum = movenum
gs.total_moves = movenum
return gsd

def resignStats(gsd, resignrate):
stats = [ TotalStats(), TotalStats(), TotalStats() ] # B wins, W wins, Overall
for gs in gsd.values():
stats[gs.winner].num_games += 1
if not gs.resign_type:
stats[gs.winner].no_resign_count += 1
stats[gs.winner].resigned_game_len_sum += gs.total_moves
elif gs.resign_type == "Correct":
stats[gs.winner].correct_resign_count += 1
stats[gs.winner].resigned_game_len_sum += gs.resign_movenum
else:
assert gs.resign_type == "Incorrect"
stats[gs.winner].incorrect_resign_count += 1
stats[gs.winner].resigned_game_len_sum += gs.resign_movenum
stats[gs.winner].game_len_sum += gs.total_moves
stats[2].calcOverall(stats[0], stats[1])
print("Black won %d/%d (%0.2f%%)" % (
stats[0].num_games,
stats[0].num_games+stats[1].num_games,
100.0 * stats[0].num_games / (stats[0].num_games+stats[1].num_games)))
for winner in (0,1,2):
if winner==0:
print("Of games Black won:")
elif winner==1:
print("Of games White won:")
else:
print("Of all games:")
if stats[winner].num_games == 0:
print(" No games to report")
continue
avg_len = 1.0*stats[winner].game_len_sum/stats[winner].num_games
resigned_avg_len = 1.0*stats[winner].resigned_game_len_sum/stats[winner].num_games
avg_reduction = 1.0*(avg_len-resigned_avg_len)/avg_len
print(" Incorrect uct resigns = %d/%d (%0.2f%%)" % (
stats[winner].incorrect_resign_count,
stats[winner].num_games,
100.0 * stats[winner].incorrect_resign_count / stats[winner].num_games))
print(" Correct uct resigns = %d/%d (%0.2f%%)" % (
stats[winner].correct_resign_count,
stats[winner].num_games,
100.0 * stats[winner].correct_resign_count / stats[winner].num_games))
print(" No Resign = %d/%d (%0.2f%%)" % (
stats[winner].no_resign_count,
stats[winner].num_games,
100.0 * stats[winner].no_resign_count / stats[winner].num_games))
print(" Average game length = %d. Average game length with resigns = %d (%0.2f%% reduction)" % (
avg_len, resigned_avg_len, avg_reduction*100))
#print("debug filename: {}".format(filename))
plynum = 0
who_resigned = None
resign_plynum = None
for line in open(filename).readlines():
m = re.match("^Score: (\S+)", line)
if m:
(score,) = m.groups()
score = float(score)
#print("debug Score", score)
m = re.match("^info string stm\s+(\S+) winrate\s+(\S+)%", line)
if m:
plynum += 1
total_plys += 1
if who_resigned:
# We already found who resigned.
# continue after just counting plys
continue
resign_plys += 1
(stm, winrate) = m.groups()
winrate = float(winrate)
if winrate < resignrate:
who_resigned = stm
resign_plynum = plynum
#print("debug stm, winrate, plynum", stm, winrate, plynum)
#print("debug who_resigned {} resign_playnum {} total_plynum {}".format(who_resigned, resign_plynum, plynum))
if ((score == -1 and who_resigned == "Black") or
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Shouldn't this be score != 1 and score != -1 on the next line?
Converting a resignation in to a draw is also an incorrect resignation?

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

It's only half as bad, so I think it should only have half as much affect on increasing FP rate.

Copy link
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Good catch. I will change it to report more stats like resigned a drawn game and resigned a won game.

(score == 1 and who_resigned == "White")):
#print("debug incorrect resign")
incorrect_resigns += 1
elif who_resigned != None:
#print("debug correct resign")
correct_resigns += 1
print("Analyze resign rate {:0.0f}%".format(resignrate))
print("incorrect {:0.0f}%\ncorrect {:0.0f}%\nplies saved {:0.0f}%".format(
incorrect_resigns/games*100, correct_resigns/games*100, (total_plys-resign_plys)/total_plys*100))
print()

if __name__ == "__main__":
usage_str = """
This script analyzes the debug output from leelaz
This script analyzes the debug output from `client -debug`
to determine the impact of various resign thresholds.

Process flow:
Run autogtp with debug on:
autogtp -k savedir -d savedir

Unzip training and debug files:
gunzip savedir/*.gz
Run `client -debug`

Analyze results with this script:
./resign_analysis.py savedir/*.txt.debug.0

Note the script takes the debug files hash.txt.debug.0
as the input arguments, but it also expects the training
files hash.txt.0 to be in the same directory."""
./resign_analysis.py logs-###/*.log
"""
parser = argparse.ArgumentParser(formatter_class=argparse.RawDescriptionHelpFormatter, description=usage_str)
default_resignrates="0.5,0.2,0.15,0.1,0.05,0.02,0.01"
parser.add_argument("--r", metavar="Resign_rates", dest="resignrates", type=str, default=default_resignrates, help="comma separated resign thresholds (default %s)" % (default_resignrates))
parser.add_argument("--v", metavar="Verbose", dest="verbose", type=int, default=0, help="Verbosity level (default 0)")
parser.add_argument("data", metavar="files", type=str, nargs="+", help="Debug data files (*.txt.debug.0)")
parser.add_argument("data", metavar="files", type=str, nargs="+", help="client debug log files (logs-###/*.log)")
args = parser.parse_args()
resignrates = [float(i) for i in args.resignrates.split(",")]
for resignrate in (resignrates):
print("Resign rate: %0.2f" % (resignrate))
gsd = parseGames(args.data, resignrate, args.verbose)
resignStats(gsd, resignrate)
for rr in (50.0, 20.0, 10.0, 5.0, 2.0, 1.0, 0.0):
parseGames(args.data, rr)
6 changes: 4 additions & 2 deletions src/UCTSearch.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -157,6 +157,10 @@ void UCTSearch::dump_stats(BoardHistory& state, UCTNode& parent) {

myprintf_so("%s\n", pvstring.c_str());
}
// winrate separate info string since it's not UCI spec
float feval = m_root->get_eval(color);
myprintf_so("info string stm %s winrate %5.2f%%\n",
color == Color::WHITE ? "White" : "Black", feval * 100.f);
myprintf("\n");
}

Expand Down Expand Up @@ -257,8 +261,6 @@ void UCTSearch::dump_analysis(int64_t elapsed, bool force_output) {
myprintf_so("info depth %d nodes %d nps %0.f score cp %d time %lld pv %s\n",
depth, visits, 1000.0 * m_playouts / (elapsed + 1),
cp, elapsed, pvstring.c_str());
//winrate separate info string since it's not UCI spec
myprintf_so("info string winrate %5.2f%%\n", feval * 100.f);
}

bool UCTSearch::is_running() const {
Expand Down