Repository navigation
Expand file tree
/
Copy patheval_gating.py
More file actions
469 lines (366 loc) · 17.9 KB
/
Copy patheval_gating.py
File metadata and controls
469 lines (366 loc) · 17.9 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
<
8000
div data-line-number="170" class="react-line-number react-code-text" style="padding-right:16px">170171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
""" revamp: attempt at a more accurate eval of two bots (including kata engine), using new dnn cache
- control randomness in each bot: open move, soft-picks (temperature?), noise
8000
my bot is controlled by configs
- measure game redundancy:
- bot uses a simpler interface than gtp: just move, and some debug info
PlayerInterface: ok, but doesn't feel quite right
"""
import logging
import random
import time
from typing import List, Tuple
import attr
import numpy as np
from absl import flags, app
import coords
import k2net as dual_net
import go
import myconf
import sgf_wrapper
import utils
from model_config import ModelConfig
from katago.analysis_engine import ARequest, KataEngine, MoveInfo, RootInfo, assemble_comment, KataModels
from run_selfplay import InitPositions
from sgfs_stats import run_tournament_report
from strategies import MCTSPlayer
FLAGS = flags.FLAGS
# flags.mark_flags_as_required(['softpick_move_cutoff', 'softpick_topn_cutoff'])
class BasicPlayerInterface:
""" simplest interface for playing an eval game
"""
def id(self) -> str:
""" id for the model config """
def initialize_game(self, position: go.Position):
"""Initializes a new game. May start from a setup position
"""
def play_move(self, c: tuple):
""" play the given move, to advance the game
"""
def suggest_moves(self, position: go.Position) -> List[Tuple]:
""" return a list of ranked moves, together with probs, for the current position.
position should match the internal state if player chooses to track position.
play_move(c) will be called after this, which may not be the top move bot considered.
returns [] if it's ready to resign
"""
def set_result(self, winner, was_resign):
""" Sets the game result. This ends the game
"""
def get_game_comments(self) -> List[str]:
""" player's comment on each move of the game
"""
class KataPlayer(BasicPlayerInterface):
""" backed by kata analysis engine
we don't really maintain game state here
"""
def __init__(self, kata_engine: KataEngine, max_readouts=500):
self.kata_engine = kata_engine
self.max_readouts = max_readouts
def id(self):
return f'{self.kata_engine.model_id()}#{self.max_readouts}'
def initialize_game(self, position):
self.moves = [coords.to_gtp(move.move)
for i, move in enumerate(position.recent)]
self.comments = ['init' for x in self.moves]
self._resp1 = None
self.win_rate = 0.5
def play_move(self, c):
""" this gets called whether or not it's our turn to move """
move = coords.to_gtp(c)
comment = 'not katas turn'
if self._resp1 is not None:
comment = assemble_comment(move, self._resp1)
self.comments.append(comment)
self.moves.append(move)
self._resp1 = None
def suggest_moves(self, position: go.Position) -> List:
assert position.n == len(self.moves)
kmoves = [['B' if i % 2 == 0 else 'W', move]
for i, move in enumerate(self.moves)]
arequest = ARequest(kmoves, [len(kmoves)], maxVisits=self.max_readouts)
responses = self.kata_engine.analyze(arequest)
assert len(responses) == 1
resp1 = responses[0]
self._resp1 = resp1
rinfo = RootInfo.from_dict(resp1.rootInfo)
self.win_rate = rinfo.winrate
top_moves = []
top_n = FLAGS.softpick_topn_cutoff
for minfo_dict in resp1.moveInfos[:top_n]:
minfo = MoveInfo.from_dict(minfo_dict)
# this may not sum up to 1
prob = minfo.visits / rinfo.visits
top_moves.append((minfo.move, prob))
return top_moves
def get_game_comments(self) -> List[str]:
return self.comments
def arg_top_k_moves(pi: np.array, k) -> np.array:
indices = np.argpartition(pi, -k)[-k:]
return indices[np.argsort(-pi[indices])]
def notest_arg_top_k_moves():
arr = np.arange(0, 1, .1)
np.random.shuffle(arr)
print(arr)
top_k_indices = arg_top_k_moves(arr, 3)
print(arr[top_k_indices])
# assert all(arr[top_k_indices] == np.array([0.9, 0.8, 0.7]))
class K2Player(BasicPlayerInterface):
""" k2net w/ mcts """
def __init__(self, model_config: ModelConfig):
self.model_config = model_config
self.dnn = dual_net.load_net(model_config.model_path())
self.mcts_player = MCTSPlayer(self.dnn, num_readouts=model_config.num_readouts)
def id(self):
return self.model_config.model_id()
def initialize_game(self, position: go.Position):
self.mcts_player.initialize_game(position, focus_area=myconf.FULL_BOARD_FOCUS)
first_node = self.mcts_player.root.select_leaf()
prob, val = self.mcts_player.network.run(first_node.position, self.mcts_player.focus_area)
first_node.incorporate_results(prob, val, first_node)
self.mcts_player.comments = ['init' for x in range(position.n)]
self.win_rate = 0.5
def play_move(self, c: tuple):
self.mcts_player.play_move(c, record_pi=False)
def suggest_moves(self, position: go.Position) -> List:
""" return a list of ranked moves, together with probs, for the current position.
"""
active = self.mcts_player
assert active.root.position.n == position.n
current_readouts = active.root.N
while active.root.N < current_readouts + active.num_readouts:
active.tree_search()
if active.should_resign(): # Force resign
return []
self.win_rate = (active.root.Q + 1) / 2 # transform my Q value to winrate
pi = active.root.children_as_pi(squash=False).flatten()
# restrict soft-pick to only top 5 moves
move_indices = arg_top_k_moves(pi, FLAGS.softpick_topn_cutoff)
top_moves = [(coords.flat_to_gtp(x), pi[x]) for x in move_indices]
return top_moves
def set_result(self, winner, was_resign):
""" """
def get_game_comments(self) -> List[str]:
return self.mcts_player.comments
@attr.s
class GameResult:
black_margin: float = attr.ib()
was_resign: bool = attr.ib()
def winner(self):
return 'B' if self.black_margin > 0 else 'W' if self.black_margin < 0 else '-'
def sgf_str(self):
winner = self.winner()
if self.was_resign:
return f'{winner}+R'
if self.black_margin == 0: # tie is always 'B+T'
return 'B+T'
margin = abs(self.black_margin)
return f'{winner}+{margin}'
def notest_game_result():
result = GameResult(0, False)
assert result.winner() == '-'
assert result.sgf_str() == 'B+T'
result = GameResult(1, True)
assert result.sgf_str() == 'B+R'
result = GameResult(-2, True)
assert result.sgf_str() == 'W+R'
result = GameResult(12, False)
assert result.sgf_str() == 'B+12'
result = GameResult(-2.5, False)
assert result.sgf_str() == 'W+2.5'
class Arbiter:
""" singleton: only starts Kata when necessary """
_instance = None
@classmethod
def get_arbiter(cls) -> 'ExpertReviwer':
if cls._instance is not None:
return cls._instance
from expert_eval import ExpertReviewer
cls._instance = ExpertReviewer()
return cls._instance
class EvaluateOneSide:
""" run evaluation, control for randomness """
def __init__(self, black_player: BasicPlayerInterface, white_player: BasicPlayerInterface, sgf_dir: str):
self.black_player = black_player
self.white_player = white_player
self.sgf_dir = sgf_dir
utils.ensure_dir_exists(sgf_dir)
self.init_positions = InitPositions(None, None) #['C2'], [1.0])
self._num_games_so_far = 0
def _end_game(self, result: GameResult):
for player in (self.black_player, self.white_player):
player.set_result(result.black_margin, was_resign=result.was_resign)
def _pick_move(self, cur_pos: go.Position, moves_with_probs: List):
""" evaluator may soft-pick to increase game variety """
# active.pick_move(active.root.position.n < FLAGS.softpick_move_cutoff)
top_move = moves_with_probs[0][0]
if cur_pos.n < FLAGS.softpick_move_cutoff:
chosen_move = utils.choose_moves_with_probs(moves_with_probs, softpick_topn_cutoff=FLAGS.softpick_topn_cutoff)
else:
chosen_move = top_move
return
8000
chosen_move, top_move
def _resolve_final_score(self, pos: go.Position, benson_score: float) -> int:
""" resolve game outcome if Benson score non-final """
reviewer = Arbiter.get_arbiter()
final_winrate = reviewer.calc_black_winrate(pos)
winner = np.sign(final_winrate - 0.5)
if winner != np.sign(benson_score):
black_winrate, white_winrate = self.black_player.win_rate, self.white_player.win_rate
logging.warning(f'Benson score {benson_score:.1f} is non-final -> final {winner}.'
f' Black winrate={black_winrate:.1f}, white {white_winrate:.1f}')
return winner
def _play_one_game(self, init_position: go.Position):
""" """
black, white = self.black_player, self.white_player
for player in [black, white]:
player.initialize_game(init_position)
cur_pos = init_position # type: go.Position
max_game_length = FLAGS.max_game_length
game_result = None
for move_idx in range(init_position.n, max_game_length):
if move_idx % 2:
active, inactive = white, black
else:
active, inactive = black, white
moves_with_probs = active.suggest_moves(cur_pos)
if len(moves_with_probs) == 0: # resigned
game_result = GameResult(-1 * cur_pos.to_play, was_resign=True)
break
# evaluator picks the next move
move, best_move = self._pick_move(cur_pos, moves_with_probs)
# advance game
c = coords.from_gtp(move)
active.play_move(c)
inactive.play_move(c)
cur_pos = cur_pos.play_move(c)
benson_score_details = cur_pos.score_benson()
if benson_score_details.final: # end the game when score is final
game_result = GameResult(benson_score_details.score, was_resign=False)
break
if cur_pos.is_game_over(): # pass-pass: use Benson score, same as self-play
final_score = self._resolve_final_score(cur_pos, benson_score_details.score)
game_result = GameResult(final_score, was_resign=False)
break
if cur_pos.n >= max_game_length:
logging.warning('ending game without a definite winner: %s', benson_score_details.score)
game_result = GameResult(0, was_resign=False)
break
self._end_game(game_result)
return cur_pos, game_result
def _create_sgf(self, final_pos: go.Position, game_result: GameResult, sgf_fname: str):
""" merge comments from black and white """
black_comments = self.black_player.get_game_comments()
white_comments = self.white_player.get_game_comments()
assert len(black_comments) == len(white_comments) and len(black_comments) == final_pos.n
comments = [black_comments[i] if i % 2 == 0 else white_comments[i] for i in range(final_pos.n)]
with open(sgf_fname, 'w') as _file:
sgfstr = sgf_wrapper.make_sgf(final_pos.recent,
game_result.sgf_str(), komi=final_pos.komi,
comments=comments,
black_name=self.black_player.id(), white_name=self.white_player.id())
_file.write(sgfstr)
def _accumulate_stats(self):
""" this needs to be done post-game, in case we run more games later """
def play_a_game(self) -> GameResult:
game_idx = self._num_games_so_far
self._num_games_so_far += 1
init_position = self.init_positions.sample()
final_pos, game_result = self._play_one_game(init_position)
# accu game stats
# self._accumulate_stats(final_pos, game_result)
sgf_fname = f'{self.sgf_dir}/{self.black_player.id()}-vs-{self.white_player.id()}-{game_idx}-%s.sgf' % (
utils.microseconds_since_midnight())
self._create_sgf(final_pos, game_result, sgf_fname)
game_history = final_pos.recent
all_moves = [coords.to_gtp(x.move) for x in game_history]
line = utils.format_game_summary(all_moves, game_result.sgf_str(), sgf_fname=sgf_fname)
logging.info(f'Finished game %3d: %s', game_idx, line)
return game_result
def play_games(self, n: int) -> int:
logging.info(f'eval between {self.black_player.id()} vs {self.white_player.id()}: {n} games')
black_wins = 0
for i in range(n):
game_result = self.play_a_game()
black_wins += game_result.black_margin > 0
logging.info(f' {self.black_player.id()} wins as black {black_wins} / {n} games against {self.white_player.id()}')
return black_wins
def load_player(model_config: ModelConfig) -> BasicPlayerInterface:
if model_config.is_kata_model():
kata_engine = KataEngine(model_config.model_path()).start()
return KataPlayer(kata_engine, model_config.num_readouts)
return K2Player(model_config)
def run_one_side(black_player, white_player, sgf_dir, num_games: int) -> int:
evaluator = EvaluateOneSide(black_player, white_player, f'{sgf_dir}')
num_blackwins = evaluator.play_games(num_games)
return num_blackwins
def main_debug(argv):
black_id = 'model7_6#50'
white_id = 'model7_4#50'
black_player = load_player(ModelConfig(black_id))
white_player = load_player(ModelConfig(white_id))
run_one_side(black_player, white_player, f'{myconf.EXP_HOME}/eval7/epochs', 2)
def main(argv):
logging.info('softpick_move_cutoff = %d, softpick_topn_cutoff = %d, reduce_symmetry_before_move = %d',
FLAGS.softpick_move_cutoff, FLAGS.softpick_topn_cutoff, FLAGS.reduce_symmetry_before_move)
_, black_id, white_id, eval_sgf_dir, num_games = argv
num_games = int(num_games)
black_player = load_player(ModelConfig(black_id))
white_player = load_player(ModelConfig(white_id))
return run_one_side(black_player, white_player, eval_sgf_dir, num_games)
def main_kata(argv):
""" eval against kata remains serial, as it's likely not a good idea to run multiple KataEngine
"""
num_readouts = 200
player2model = 'model7_8'
sgf_dir_root = f'{myconf.EXP_HOME}/eval_gating/#{num_readouts}/{player2model}-vs-elo4k'
player1id = f'{KataModels.MODEL_B6_4k}#{num_readouts}'
player2id = f'{player2model}#{num_readouts}'
num_games_per_side = 20
FLAGS.softpick_move_cutoff = 8
FLAGS.reduce_symmetry_before_move = 3
logging.info('softpick_move_cutoff = %d, softpick_topn_cutoff = %d, reduce_symmetry_before_move = %d resign_thresh = %.1f',
FLAGS.softpick_move_cutoff, FLAGS.softpick_topn_cutoff, FLAGS.reduce_symmetry_before_move, FLAGS.resign_threshold)
# share KataEngine between two runs
player1 = load_player(ModelConfig(player1id))
player2 = load_player(ModelConfig(player2id))
run_one_side(player1, player2, f'{sgf_dir_root}', num_games_per_side)
run_one_side(player2, player1, f'{sgf_dir_root}', num_games_per_side)
run_tournament_report(f'{sgf_dir_root}/*', include_details=True)
def test_run_report_on_sgfs():
""" just read sgfs and report """
sgf_pattern = f'{myconf.EXP_HOME}/eval_gating/#200/model10_2-vs-elo4k/*'
# run_tournament_report(f'{sgf_pattern}')
# run_tournament_report(f'{myconf.EXP_HOME}/eval10/epochs/model10_2#200-vs-model7_10#200-*.sgf')
# run_tournament_report(f'{myconf.EXP_HOME}/eval10/epochs/model7_10#200-vs-model10_2#200-*.sgf')
run_tournament_report(f'{myconf.EXP_HOME}/../9x9-exp2/eval_gating/model12_2-vs-elo4k/*.sgf')
# run_tournament_report(f'{myconf.EXP_HOME}/../9x9-exp2/eval_gating/model6_2/2/*')
def review_kata(argv):
""" generational review of models in exp2 """
num_readouts = 400
num_games_per_side = 20
FLAGS.reduce_symmetry_before_move = 3
kata_id = KataModels.MODEL_B6_4k
kata_id_short = KataModels.model_id(kata_id)
sgf_dir_root = f'{myconf.EXP_HOME}/eval_review/{kata_id_short}'
# share KataEngine across runs
kataplayer = load_player(ModelConfig(f'{kata_id}#{num_readouts}'))
k2models = [f'model{x}.mlpackage#{num_readouts}'
for x in ['1_5', '2_2', '3_3', '4_4', '5_2', '6_2', '7_4', '8_4', '9_4', '10_4', '11_3', '12_2']]
logging.info('softpick_move_cutoff = %d, softpick_topn_cutoff = %d, reduce_symmetry_before_move = %d resign_thresh = %.1f',
FLAGS.softpick_move_cutoff, FLAGS.softpick_topn_cutoff, FLAGS.reduce_symmetry_before_move, FLAGS.resign_threshold)
for player2id in k2models[:11]:
player2 = load_player(ModelConfig(player2id))
sgf_dir = f'{sgf_dir_root}/{player2id}'
FLAGS.softpick_move_cutoff = 0
logging.info(f'running {player2id} softpick_move_cutoff = %d', FLAGS.softpick_move_cutoff)
# play the strongest game
num_wins_as_white = 1 - run_one_side(kataplayer, player2, f'{sgf_dir}', 1)
num_wins_as_black = run_one_side(player2, kataplayer, f'{sgf_dir}', 1)
# play the rest with diversity
FLAGS.softpick_move_cutoff = 8
logging.info(f'running {player2id} softpick_move_cutoff = %d', FLAGS.softpick_move_cutoff)
num_wins_as_white += num_games_per_side - 1 - run_one_side(kataplayer, player2, f'{sgf_dir}', num_games_per_side - 1)
num_wins_as_black += run_one_side(player2, kataplayer, f'{sgf_dir}', num_games_per_side - 1)
logging.info(f'{player2id} wins as black {num_wins_as_black}, as white {num_wins_as_white}')
if __name__ == '__main__':
FLAGS.resign_threshold = -1
# app.run(main_kata)
app.run(main)