def settledness(d, player_sign, player):
ownership_grid = var_to_grid(
d["ownership"], (board_size_x, board_size_y))
return sum([
abs(ownership_grid[s.coords[0]][s.coords[1]])
for s in game.stones if s.player == player
])
def manual_score(self):
rules = self.rules
if (not self.current_node.ownership
or str(rules).lower() not in ["jp", "japanese"]
or not self.current_node.parent
or not self.current_node.parent.ownership):
if not self.current_node.score:
return None
return self.current_node.format_score(
round(2 * self.current_node.score) / 2) + "?"
board_size_x, board_size_y = self.board_size
mean_ownership = [(c + p) / 2 for c, p in zip(
self.current_node.ownership, self.current_node.parent.ownership)]
ownership_grid = var_to_grid(mean_ownership,
(board_size_x, board_size_y))
stones = {m.coords: m.player for m in self.stones}
lo_threshold = 0.15
hi_threshold = 0.85
max_unknown = 10
max_dame = 4 * (board_size_x + board_size_y)
def japanese_score_square(square, owner):
player = stones.get(square, None)
if ((player == "B" and owner > hi_threshold)
or (player == "W" and owner < -hi_threshold)
or abs(owner) < lo_threshold):
return 0 # dame or own stones
if player is None and abs(owner) >= hi_threshold:
return round(owner) # surrounded empty intersection
if (player == "B"
and owner < -hi_threshold) or (player == "W"
and owner > hi_threshold):
return 2 * round(owner) # captured stone
return math.nan # unknown!
scored_squares = [
japanese_score_square((x, y), ownership_grid[y][x])
for y in range(board_size_y) for x in range(board_size_x)
]
num_sq = {
t: sum([s == t for s in scored_squares])
for t in [-2, -1, 0, 1, 2]
}
num_unkn = sum(math.isnan(s) for s in scored_squares)
prisoners = self.prisoner_count
score = sum([t * n for t, n in num_sq.items()
]) + prisoners["W"] - prisoners["B"] - self.komi
self.katrain.log(
f"Manual Scoring: {num_sq} score by square with {num_unkn} unknown, {prisoners} captures, and {self.komi} komi -> score = {score}",
OUTPUT_DEBUG,
)
if num_unkn > max_unknown or (num_sq[0] - len(stones)) > max_dame:
return None
return self.current_node.format_score(score)
def policy_ranking(
self
) -> Optional[List[Tuple[
float, Move]]]: # return moves from highest policy value to lowest
if self.policy:
szx, szy = self.board_size
policy_grid = var_to_grid(self.policy, size=(szx, szy))
moves = [(policy_grid[y][x], Move((x, y), player=self.next_player))
for x in range(szx) for y in range(szy)]
moves.append((self.policy[-1], Move(None,
player=self.next_player)))
return sorted(moves, key=lambda mp: -mp[0])
def analyze_extra(self, mode):
stones = {s.coords for s in self.stones}
cn = self.current_node
engine = self.engines[cn.next_player]
if mode == "extra":
visits = cn.analysis_visits_requested + engine.config["max_visits"]
self.katrain.controls.set_status(i18n._("extra analysis").format(visits=visits))
cn.analyze(engine, visits=visits, priority=-1_000, time_limit=False)
return
elif mode == "sweep":
board_size_x, board_size_y = self.board_size
if cn.analysis_ready:
policy_grid = (
var_to_grid(self.current_node.policy, size=(board_size_x, board_size_y))
if self.current_node.policy
else None
)
analyze_moves = sorted(
[
Move(coords=(x, y), player=cn.next_player)
for x in range(board_size_x)
for y in range(board_size_y)
if (policy_grid is None and (x, y) not in stones) or policy_grid[y][x] >= 0
],
key=lambda mv: -policy_grid[mv.coords[1]][mv.coords[0]],
)
else:
analyze_moves = [
Move(coords=(x, y), player=cn.next_player)
for x in range(board_size_x)
for y in range(board_size_y)
if (x, y) not in stones
]
visits = engine.config["fast_visits"]
self.katrain.controls.set_status(i18n._("sweep analysis").format(visits=visits))
priority = -1_000_000_000
else: # mode=='equalize':
if not cn.analysis_ready:
self.katrain.controls.set_status(i18n._("wait-before-equalize"), self.current_node)
return
analyze_moves = [Move.from_gtp(gtp, player=cn.next_player) for gtp, _ in cn.analysis["moves"].items()]
visits = max(d["visits"] for d in cn.analysis["moves"].values())
self.katrain.controls.set_status(i18n._("equalizing analysis").format(visits=visits))
priority = -1_000
for move in analyze_moves:
cn.analyze(
engine, priority, visits=visits, refine_move=move, time_limit=False
) # explicitly requested so take as long as you need
def analyze_extra(self, mode, **kwargs):
stones = {s.coords for s in self.stones}
cn = self.current_node
if mode == "stop":
for e in set(self.engines.values()):
e.terminate_queries()
self.katrain.idle_analysis = False
return
engine = self.engines[cn.next_player]
Clock.schedule_once(self.katrain.analysis_controls.hints.activate, 0)
if mode == "extra":
if kwargs.get("continuous", False):
visits = min(
1_000_000_000,
max(engine.config["max_visits"],
math.ceil(cn.analysis_visits_requested * 1.25)))
else:
visits = cn.analysis_visits_requested + engine.config[
"max_visits"]
self.katrain.controls.set_status(
i18n._("extra analysis").format(visits=visits),
STATUS_ANALYSIS)
self.katrain.controls.set_status(
i18n._("extra analysis").format(visits=visits),
STATUS_ANALYSIS)
cn.analyze(engine,
visits=visits,
priority=-1_000,
region_of_interest=self.region_of_interest,
time_limit=False)
return
if mode == "game":
nodes = self.root.nodes_in_tree
if "visits" in kwargs:
visits = kwargs["visits"]
else:
min_visits = min(node.analysis_visits_requested
for node in nodes)
visits = min_visits + engine.config["max_visits"]
for node in nodes:
node.analyze(engine,
visits=visits,
priority=-1_000_000,
time_limit=False,
report_every=None)
self.katrain.controls.set_status(
i18n._("game re-analysis").format(visits=visits),
STATUS_ANALYSIS)
return
elif mode == "sweep":
board_size_x, board_size_y = self.board_size
if cn.analysis_exists:
policy_grid = (var_to_grid(self.current_node.policy,
size=(board_size_x, board_size_y))
if self.current_node.policy else None)
analyze_moves = sorted(
[
Move(coords=(x, y), player=cn.next_player)
for x in range(board_size_x)
for y in range(board_size_y)
if (policy_grid is None and
(x, y) not in stones) or policy_grid[y][x] >= 0
],
key=lambda mv: -policy_grid[mv.coords[1]][mv.coords[0]],
)
else:
analyze_moves = [
Move(coords=(x, y), player=cn.next_player)
for x in range(board_size_x) for y in range(board_size_y)
if (x, y) not in stones
]
visits = engine.config["fast_visits"]
self.katrain.controls.set_status(
i18n._("sweep analysis").format(visits=visits),
STATUS_ANALYSIS)
priority = -1_000_000_000
elif mode in ["equalize", "alternative", "local"]:
if not cn.analysis_complete and mode != "local":
self.katrain.controls.set_status(
i18n._("wait-before-extra-analysis"), STATUS_INFO,
self.current_node)
return
if mode == "alternative": # also do a quick update on current candidates so it doesn't look too weird
self.katrain.controls.set_status(
i18n._("alternative analysis"), STATUS_ANALYSIS)
cn.analyze(engine,
priority=-500,
time_limit=False,
find_alternatives="alternative")
visits = engine.config["fast_visits"]
else: # equalize
visits = max(d["visits"]
for d in cn.analysis["moves"].values())
self.katrain.controls.set_status(
i18n._("equalizing analysis").format(visits=visits),
STATUS_ANALYSIS)
priority = -1_000
analyze_moves = [
Move.from_gtp(gtp, player=cn.next_player)
for gtp, _ in cn.analysis["moves"].items()
]
else:
raise ValueError("Invalid analysis mode")
for move in analyze_moves:
if cn.analysis["moves"].get(move.gtp(),
{"visits": 0})["visits"] < visits:
cn.analyze(
engine,
priority=priority,
visits=visits,
refine_move=move,
time_limit=False
) # explicitly requested so take as long as you need
def draw_board_contents(self, *_args):
if not (self.katrain and self.katrain.game):
return
stone_color = STONE_COLORS
outline_color = OUTLINE_COLORS
katrain = self.katrain
board_size_x, board_size_y = katrain.game.board_size
lock_ai = self.trainer_config[
"lock_ai"] and self.katrain.play_analyze_mode == MODE_PLAY
show_n_eval = self.trainer_config["eval_off_show_last"]
self.canvas.clear()
with self.canvas:
# stones
current_node = katrain.game.current_node
game_ended = katrain.game.ended
full_eval_on = katrain.analysis_controls.eval.active
has_stone = {}
drawn_stone = {}
for m in katrain.game.stones:
has_stone[m.coords] = m.player
show_dots_for = {
p: self.trainer_config["eval_show_ai"]
or katrain.players_info[p].human
for p in Move.PLAYERS
}
show_dots_for_class = self.trainer_config["show_dots"]
nodes = katrain.game.current_node.nodes_from_root
realized_points_lost = None
katrain.config("trainer/show_dots")
for i, node in enumerate(nodes[::-1]): # reverse order!
points_lost = node.points_lost
evalsize = 1
if points_lost and realized_points_lost:
if points_lost <= 0.5 and realized_points_lost <= 1.5:
evalsize = 0
else:
evalsize = min(
1, max(0, realized_points_lost / points_lost))
for m in node.move_with_placements:
if has_stone.get(m.coords) and not drawn_stone.get(
m.coords): # skip captures, last only for
move_eval_on = show_dots_for.get(
m.player) and (i < show_n_eval or full_eval_on)
if move_eval_on and points_lost is not None:
evalcol = self.eval_color(points_lost,
show_dots_for_class)
else:
evalcol = None
inner = stone_color[m.opponent] if i == 0 else None
drawn_stone[m.coords] = m.player
self.draw_stone(
m.coords[0],
m.coords[1],
stone_color[m.player],
outline_color[m.player],
inner,
evalcol,
evalsize,
)
realized_points_lost = node.parent_realized_points_lost
if katrain.game.current_node.is_root and katrain.debug_level >= 3: # secret ;)
for y in range(0, board_size_y):
evalcol = self.eval_color(16 * y / board_size_y)
self.draw_stone(0, y, stone_color["B"], outline_color["B"],
None, evalcol, y / (board_size_y - 1))
self.draw_stone(1, y, stone_color["B"], outline_color["B"],
stone_color["W"], evalcol, 1)
self.draw_stone(2, y, stone_color["W"], outline_color["W"],
None, evalcol, y / (board_size_y - 1))
self.draw_stone(3, y, stone_color["W"], outline_color["W"],
stone_color["B"], evalcol, 1)
self.draw_stone(4, y, [*evalcol[:3], 0.5], scale=0.8)
# ownership - allow one move out of date for smooth animation
ownership = current_node.ownership or (
current_node.parent and current_node.parent.ownership)
if katrain.analysis_controls.ownership.active and ownership and not lock_ai:
ownership_grid = var_to_grid(ownership,
(board_size_x, board_size_y))
rsz = self.grid_size * 0.2
for y in range(board_size_y - 1, -1, -1):
for x in range(board_size_x):
ix_owner = "B" if ownership_grid[y][x] > 0 else "W"
if ix_owner != (has_stone.get((x, y), -1)):
Color(*stone_color[ix_owner],
abs(ownership_grid[y][x]))
Rectangle(pos=(self.gridpos_x[x] - rsz / 2,
self.gridpos_y[y] - rsz / 2),
size=(rsz, rsz))
policy = current_node.policy
if (not policy and current_node.parent
and current_node.parent.policy
and katrain.last_player_info.ai
and katrain.next_player_info.ai):
policy = (
current_node.parent.policy
) # in the case of AI self-play we allow the policy to be one step out of date
pass_btn = katrain.board_controls.pass_btn
pass_btn.canvas.after.clear()
if katrain.analysis_controls.policy.active and policy and not lock_ai:
policy_grid = var_to_grid(policy, (board_size_x, board_size_y))
best_move_policy = max(*policy)
for y in range(board_size_y - 1, -1, -1):
for x in range(board_size_x):
if policy_grid[y][x] > 0:
polsize = 1.1 * math.sqrt(policy_grid[y][x])
policy_circle_color = (
*POLICY_COLOR,
GHOST_ALPHA + TOP_MOVE_ALPHA *
(policy_grid[y][x] == best_move_policy),
)
self.draw_stone(x,
y,
policy_circle_color,
scale=polsize)
polsize = math.sqrt(policy[-1])
with pass_btn.canvas.after:
draw_circle(
(pass_btn.pos[0] + pass_btn.width / 2,
pass_btn.pos[1] + pass_btn.height / 2),
polsize * pass_btn.height / 2,
POLICY_COLOR,
)
# pass circle
passed = len(nodes) > 1 and current_node.is_pass
if passed:
if game_ended:
text = katrain.game.manual_score or i18n._(
"board-game-end")
else:
text = i18n._("board-pass")
Color(0.45, 0.05, 0.45, 0.7)
center = (self.gridpos_x[int(board_size_x / 2)],
self.gridpos_y[int(board_size_y / 2)])
size = min(self.width, self.height) * 0.22
Ellipse(pos=(center[0] - size / 2, center[1] - size / 2),
size=(size, size))
Color(0.85, 0.85, 0.85)
draw_text(pos=center,
text=text,
font_size=size * 0.25,
halign="center",
outline_color=[0.95, 0.95, 0.95])
self.draw_hover_contents()
def draw_board_contents(self, *_args):
if not (self.katrain and self.katrain.game):
return
katrain = self.katrain
board_size_x, board_size_y = katrain.game.board_size
if len(self.gridpos_x) < board_size_x or len(self.gridpos_y) < board_size_y:
return # race condition
show_n_eval = self.trainer_config.get("eval_on_show_last", 3)
with self.canvas:
self.canvas.clear()
# stones
current_node = katrain.game.current_node
all_dots_off = not katrain.analysis_controls.eval.active
has_stone = {}
drawn_stone = {}
for m in katrain.game.stones:
has_stone[m.coords] = m.player
show_dots_for = {
p: self.trainer_config["eval_show_ai"] or katrain.players_info[p].human for p in Move.PLAYERS
}
show_dots_for_class = self.trainer_config["show_dots"]
nodes = katrain.game.current_node.nodes_from_root
realized_points_lost = None
for i, node in enumerate(nodes[::-1]): # reverse order!
points_lost = node.points_lost
evalscale = 1
if points_lost and realized_points_lost:
if points_lost <= 0.5 and realized_points_lost <= 1.5:
evalscale = 0
else:
evalscale = min(1, max(0, realized_points_lost / points_lost))
placements = node.placements
for m in node.moves + placements:
if has_stone.get(m.coords) and not drawn_stone.get(m.coords): # skip captures, last only for
move_eval_on = not all_dots_off and show_dots_for.get(m.player) and i < show_n_eval
if move_eval_on and points_lost is not None:
evalcol = self.eval_color(points_lost, show_dots_for_class)
else:
evalcol = None
inner = Theme.STONE_COLORS[m.opponent] if i == 0 and m not in placements else None
drawn_stone[m.coords] = m.player
self.draw_stone(
x=m.coords[0],
y=m.coords[1],
player=m.player,
innercol=inner,
evalcol=evalcol,
evalscale=evalscale,
)
realized_points_lost = node.parent_realized_points_lost
if katrain.game.current_node.is_root and katrain.debug_level >= 3: # secret ;)
for y in range(0, board_size_y):
evalcol = self.eval_color(16 * y / board_size_y)
self.draw_stone(0, y, "B", evalcol=evalcol, evalscale=y / (board_size_y - 1))
self.draw_stone(1, y, "B", innercol=Theme.STONE_COLORS["W"], evalcol=evalcol)
self.draw_stone(2, y, "W", evalcol=evalcol, evalscale=y / (board_size_y - 1))
self.draw_stone(3, y, "W", innercol=Theme.STONE_COLORS["B"], evalcol=evalcol)
# ownership - allow one move out of date for smooth animation
ownership = current_node.ownership or (current_node.parent and current_node.parent.ownership)
if katrain.analysis_controls.ownership.active and ownership:
rsz = self.grid_size * 0.2
if (
current_node.children
and katrain.controls.status_state[1] == STATUS_TEACHING
and current_node.children[-1].auto_undo
and current_node.children[-1].ownership
): # loss
loss_grid = var_to_grid(
[a - b for a, b in zip(current_node.children[-1].ownership, ownership)],
(board_size_x, board_size_y),
)
for y in range(board_size_y - 1, -1, -1):
for x in range(board_size_x):
loss = max(0, (-1 if current_node.children[-1].move.player == "B" else 1) * loss_grid[y][x])
if loss > 0:
Color(*Theme.EVAL_COLORS[self.trainer_config["theme"]][1][:3], loss)
Rectangle(
pos=(self.gridpos_x[x] - rsz / 2, self.gridpos_y[y] - rsz / 2), size=(rsz, rsz)
)
else:
ownership_grid = var_to_grid(ownership, (board_size_x, board_size_y))
for y in range(board_size_y - 1, -1, -1):
for x in range(board_size_x):
ix_owner = "B" if ownership_grid[y][x] > 0 else "W"
if ix_owner != (has_stone.get((x, y), -1)):
Color(*Theme.STONE_COLORS[ix_owner][:3], abs(ownership_grid[y][x]))
Rectangle(
pos=(self.gridpos_x[x] - rsz / 2, self.gridpos_y[y] - rsz / 2), size=(rsz, rsz)
)
policy = current_node.policy
if (
not policy
and current_node.parent
and current_node.parent.policy
and katrain.last_player_info.ai
and katrain.next_player_info.ai
):
policy = (
current_node.parent.policy
) # in the case of AI self-play we allow the policy to be one step out of date
pass_btn = katrain.board_controls.pass_btn
pass_btn.canvas.after.clear()
if katrain.analysis_controls.policy.active and policy:
policy_grid = var_to_grid(policy, (board_size_x, board_size_y))
best_move_policy = max(*policy)
colors = Theme.EVAL_COLORS[self.trainer_config["theme"]]
text_lb = 0.01 * 0.01
for y in range(board_size_y - 1, -1, -1):
for x in range(board_size_x):
move_policy = policy_grid[y][x]
if move_policy < 0:
continue
pol_order = max(0, 5 + int(math.log10(max(1e-9, move_policy - 1e-9))))
if move_policy > text_lb:
draw_circle(
(self.gridpos_x[x], self.gridpos_y[y]),
self.stone_size * Theme.HINT_SCALE * 0.98,
Theme.APPROX_BOARD_COLOR,
)
scale = 0.95
else:
scale = 0.5
draw_circle(
(self.gridpos_x[x], self.gridpos_y[y]),
Theme.HINT_SCALE * self.stone_size * scale,
(*colors[pol_order][:3], Theme.POLICY_ALPHA),
)
if move_policy > text_lb:
Color(*Theme.HINT_TEXT_COLOR)
draw_text(
pos=(self.gridpos_x[x], self.gridpos_y[y]),
text=f"{100 * move_policy :.2f}"[:4] + "%",
font_name="Roboto",
font_size=self.grid_size / 4,
halign="center",
)
if move_policy == best_move_policy:
Color(*Theme.TOP_MOVE_BORDER_COLOR[:3], Theme.POLICY_ALPHA)
Line(
circle=(
self.gridpos_x[x],
self.gridpos_y[y],
self.stone_size - dp(1.2),
),
width=dp(2),
)
with pass_btn.canvas.after:
move_policy = policy[-1]
pol_order = 5 - int(-math.log10(max(1e-9, move_policy - 1e-9)))
if pol_order >= 0:
draw_circle(
(pass_btn.pos[0] + pass_btn.width / 2, pass_btn.pos[1] + pass_btn.height / 2),
pass_btn.height / 2,
(*colors[pol_order][:3], Theme.GHOST_ALPHA),
)
self.redraw_hover_contents_trigger()
def generate_ai_move(game: Game, ai_mode: str,
ai_settings: Dict) -> Tuple[Move, GameNode]:
cn = game.current_node
while not cn.analysis_ready:
time.sleep(0.01)
game.engines[cn.next_player].check_alive(exception_if_dead=True)
ai_thoughts = ""
if (ai_mode
in AI_STRATEGIES_POLICY) and cn.policy: # pure policy based move
policy_moves = cn.policy_ranking
pass_policy = cn.policy[-1]
# dont make it jump around for the last few sensible non pass moves
top_5_pass = any([polmove[1].is_pass for polmove in policy_moves[:5]])
size = game.board_size
policy_grid = var_to_grid(cn.policy, size) # type: List[List[float]]
top_policy_move = policy_moves[0][1]
ai_thoughts += f"Using policy based strategy, base top 5 moves are {fmt_moves(policy_moves[:5])}. "
if (ai_mode == AI_POLICY and cn.depth <= ai_settings["opening_moves"]
) or (ai_mode in [AI_LOCAL, AI_TENUKI]
and not (cn.move and cn.move.coords)):
ai_mode = AI_WEIGHTED
ai_thoughts += f"Strategy override, using policy-weighted strategy instead. "
ai_settings = {
"pick_override": 0.9,
"weaken_fac": 1,
"lower_bound": 0.02
}
if top_5_pass:
aimove = top_policy_move
ai_thoughts += "Playing top one because one of them is pass."
elif ai_mode == AI_POLICY:
aimove = top_policy_move
ai_thoughts += f"Playing top policy move {aimove.gtp()}."
else: # weighted or pick-based
legal_policy_moves = [(pol, mv) for pol, mv in policy_moves
if not mv.is_pass and pol > 0]
board_squares = size[0] * size[1]
if ai_mode == AI_RANK: # calibrated, override from 0.8 at start to ~0.4 at full board
override = 0.8 * (
1 - 0.5 *
(board_squares - len(legal_policy_moves)) / board_squares)
else:
override = ai_settings["pick_override"]
if policy_moves[0][0] > override:
aimove = top_policy_move
ai_thoughts += f"Top policy move has weight > {override:.1%}, so overriding other strategies."
elif ai_mode == AI_WEIGHTED:
aimove, ai_thoughts = policy_weighted_move(
policy_moves, ai_settings["lower_bound"],
ai_settings["weaken_fac"])
elif ai_mode in AI_STRATEGIES_PICK:
if ai_mode != AI_RANK:
n_moves = int(ai_settings["pick_frac"] *
len(legal_policy_moves) +
ai_settings["pick_n"])
else:
n_moves = int(
round(
board_squares / 361 *
10**(-0.05737 * ai_settings["kyu_rank"] + 1.9482)))
if ai_mode in [
AI_INFLUENCE, AI_TERRITORY, AI_LOCAL, AI_TENUKI
]:
if cn.depth > ai_settings["endgame"] * board_squares:
weighted_coords = [(pol, 1, *mv.coords)
for pol, mv in legal_policy_moves]
x_ai_thoughts = (
f"Generated equal weights as move number >= {ai_settings['endgame'] * size[0] * size[1]}. "
)
n_moves = int(
max(n_moves, 0.5 * len(legal_policy_moves)))
elif ai_mode in [AI_INFLUENCE, AI_TERRITORY]:
weighted_coords, x_ai_thoughts = generate_influence_territory_weights(
ai_mode, ai_settings, policy_grid, size)
else: # ai_mode in [AI_LOCAL, AI_TENUKI]
weighted_coords, x_ai_thoughts = generate_local_tenuki_weights(
ai_mode, ai_settings, policy_grid, cn, size)
ai_thoughts += x_ai_thoughts
else: # ai_mode in [AI_PICK, AI_RANK]:
weighted_coords = [(policy_grid[y][x], 1, x, y)
for x in range(size[0])
for y in range(size[1])
if policy_grid[y][x] > 0]
pick_moves = weighted_selection_without_replacement(
weighted_coords, n_moves)
ai_thoughts += f"Picked {min(n_moves,len(weighted_coords))} random moves according to weights. "
if pick_moves:
new_top = [
(p, Move((x, y), player=cn.next_player))
for p, wt, x, y in heapq.nlargest(5, pick_moves)
]
aimove = new_top[0][1]
ai_thoughts += f"Top 5 among these were {fmt_moves(new_top)} and picked top {aimove.gtp()}. "
if new_top[0][0] < pass_policy:
ai_thoughts += f"But found pass ({pass_policy:.2%} to be higher rated than {aimove.gtp()} ({new_top[0][0]:.2%}) so will play top policy move instead."
aimove = top_policy_move
else:
aimove = top_policy_move
ai_thoughts += f"Pick policy strategy {ai_mode} failed to find legal moves, so is playing top policy move {aimove.gtp()}."
else:
raise ValueError(f"Unknown Policy-based AI mode {ai_mode}")
else: # Engine based move
candidate_ai_moves = cn.candidate_moves
top_cand = Move.from_gtp(candidate_ai_moves[0]["move"],
player=cn.next_player)
if top_cand.is_pass: # don't play suicidal to balance score - pass when it's best
aimove = top_cand
ai_thoughts += f"Top move is pass, so passing regardless of strategy."
else:
if ai_mode == AI_JIGO:
sign = cn.player_sign(cn.next_player)
jigo_move = min(
candidate_ai_moves,
key=lambda move: abs(sign * move["scoreLead"] -
ai_settings["target_score"]))
aimove = Move.from_gtp(jigo_move["move"],
player=cn.next_player)
ai_thoughts += f"Jigo strategy found {len(candidate_ai_moves)} candidate moves (best {top_cand.gtp()}) and chose {aimove.gtp()} as closest to 0.5 point win"
elif ai_mode == AI_SCORELOSS:
c = ai_settings["strength"]
moves = [(
d["pointsLost"],
math.exp(min(200, -c * max(0, d["pointsLost"]))),
Move.from_gtp(d["move"], player=cn.next_player),
) for d in candidate_ai_moves]
topmove = weighted_selection_without_replacement(moves, 1)[0]
aimove = topmove[2]
ai_thoughts += f"ScoreLoss strategy found {len(candidate_ai_moves)} candidate moves (best {top_cand.gtp()}) and chose {aimove.gtp()} (weight {topmove[1]:.3f}, point loss {topmove[0]:.1f}) based on score weights."
else:
if ai_mode != AI_DEFAULT:
game.katrain.log(
f"Unknown AI mode {ai_mode} or policy missing, using default.",
OUTPUT_INFO)
ai_thoughts += f"Strategy {ai_mode} not found or unexpected fallback."
aimove = top_cand
ai_thoughts += f"Default strategy found {len(candidate_ai_moves)} moves returned from the engine and chose {aimove.gtp()} as top move"
game.katrain.log(f"AI thoughts: {ai_thoughts}", OUTPUT_DEBUG)
played_node = game.play(aimove)
played_node.ai_thoughts = ai_thoughts
return aimove, played_node
def generate_ai_move(game: Game, ai_mode: str,
ai_settings: Dict) -> Tuple[Move, GameNode]:
cn = game.current_node
if ai_mode == AI_HANDICAP:
pda = ai_settings["pda"]
if ai_settings["automatic"]:
n_handicaps = len(game.root.get_list_property("AB", []))
MOVE_VALUE = 14 # could be rules dependent
b_stones_advantage = max(
n_handicaps - 1, 0) - (cn.komi - MOVE_VALUE / 2) / MOVE_VALUE
pda = min(3, max(
-3, -b_stones_advantage *
(3 /
8))) # max PDA at 8 stone adv, normal 9 stone game is 8.46
handicap_analysis = request_ai_analysis(
game, cn, {
"playoutDoublingAdvantage": pda,
"playoutDoublingAdvantagePla": "BLACK"
})
if not handicap_analysis:
game.katrain.log(f"Error getting handicap-based move",
OUTPUT_ERROR)
ai_mode = AI_DEFAULT
while not cn.analysis_ready:
time.sleep(0.01)
game.engines[cn.next_player].check_alive(exception_if_dead=True)
ai_thoughts = ""
if (ai_mode
in AI_STRATEGIES_POLICY) and cn.policy: # pure policy based move
policy_moves = cn.policy_ranking
pass_policy = cn.policy[-1]
# dont make it jump around for the last few sensible non pass moves
top_5_pass = any([polmove[1].is_pass for polmove in policy_moves[:5]])
size = game.board_size
policy_grid = var_to_grid(cn.policy, size) # type: List[List[float]]
top_policy_move = policy_moves[0][1]
ai_thoughts += f"Using policy based strategy, base top 5 moves are {fmt_moves(policy_moves[:5])}. "
if (ai_mode == AI_POLICY and cn.depth <= ai_settings["opening_moves"]
) or (ai_mode in [AI_LOCAL, AI_TENUKI]
and not (cn.move and cn.move.coords)):
ai_mode = AI_WEIGHTED
ai_thoughts += f"Strategy override, using policy-weighted strategy instead. "
ai_settings = {
"pick_override": 0.9,
"weaken_fac": 1,
"lower_bound": 0.02
}
if top_5_pass:
aimove = top_policy_move
ai_thoughts += "Playing top one because one of them is pass."
elif ai_mode == AI_POLICY:
aimove = top_policy_move
ai_thoughts += f"Playing top policy move {aimove.gtp()}."
else: # weighted or pick-based
legal_policy_moves = [(pol, mv) for pol, mv in policy_moves
if not mv.is_pass and pol > 0]
board_squares = size[0] * size[1]
if ai_mode == AI_RANK: # calibrated, override from 0.8 at start to ~0.4 at full board
override = 0.8 * (
1 - 0.5 *
(board_squares - len(legal_policy_moves)) / board_squares)
overridetwo = 0.85 + max(0, 0.02 *
(ai_settings["kyu_rank"] - 8))
else:
override = ai_settings["pick_override"]
overridetwo = 1.0
if policy_moves[0][0] > override:
aimove = top_policy_move
ai_thoughts += f"Top policy move has weight > {override:.1%}, so overriding other strategies."
elif policy_moves[0][0] + policy_moves[1][0] > overridetwo:
aimove = top_policy_move
ai_thoughts += (
f"Top two policy moves have cumulative weight > {overridetwo:.1%}, so overriding other strategies."
)
elif ai_mode == AI_WEIGHTED:
aimove, ai_thoughts = policy_weighted_move(
policy_moves, ai_settings["lower_bound"],
ai_settings["weaken_fac"])
elif ai_mode in AI_STRATEGIES_PICK:
if ai_mode != AI_RANK:
n_moves = max(
1,
int(ai_settings["pick_frac"] *
len(legal_policy_moves) + ai_settings["pick_n"]))
else:
orig_calib_avemodrank = 0.063015 + 0.7624 * board_squares / (
10**(-0.05737 * ai_settings["kyu_rank"] + 1.9482))
norm_leg_moves = len(legal_policy_moves) / board_squares
modified_calib_avemodrank = (
0.3931 + 0.6559 * norm_leg_moves * math.exp(
-1 *
(3.002 * norm_leg_moves * norm_leg_moves -
norm_leg_moves - 0.034889 *
ai_settings["kyu_rank"] - 0.5097)**2) - 0.01093 *
ai_settings["kyu_rank"]) * orig_calib_avemodrank
n_moves = board_squares * norm_leg_moves / (
1.31165 * (modified_calib_avemodrank + 1) - 0.082653)
n_moves = max(1, round(n_moves))
if ai_mode in [
AI_INFLUENCE, AI_TERRITORY, AI_LOCAL, AI_TENUKI
]:
if cn.depth > ai_settings["endgame"] * board_squares:
weighted_coords = [(pol, 1, *mv.coords)
for pol, mv in legal_policy_moves]
x_ai_thoughts = (
f"Generated equal weights as move number >= {ai_settings['endgame'] * size[0] * size[1]}. "
)
n_moves = int(
max(n_moves,
len(legal_policy_moves) // 2))
elif ai_mode in [AI_INFLUENCE, AI_TERRITORY]:
weighted_coords, x_ai_thoughts = generate_influence_territory_weights(
ai_mode, ai_settings, policy_grid, size)
else: # ai_mode in [AI_LOCAL, AI_TENUKI]
weighted_coords, x_ai_thoughts = generate_local_tenuki_weights(
ai_mode, ai_settings, policy_grid, cn, size)
ai_thoughts += x_ai_thoughts
else: # ai_mode in [AI_PICK, AI_RANK]:
weighted_coords = [(policy_grid[y][x], 1, x, y)
for x in range(size[0])
for y in range(size[1])
if policy_grid[y][x] > 0]
pick_moves = weighted_selection_without_replacement(
weighted_coords, n_moves)
ai_thoughts += f"Picked {min(n_moves,len(weighted_coords))} random moves according to weights. "
if pick_moves:
new_top = [
(p, Move((x, y), player=cn.next_player))
for p, wt, x, y in heapq.nlargest(5, pick_moves)
]
aimove = new_top[0][1]
ai_thoughts += f"Top 5 among these were {fmt_moves(new_top)} and picked top {aimove.gtp()}. "
if new_top[0][0] < pass_policy:
ai_thoughts += f"But found pass ({pass_policy:.2%} to be higher rated than {aimove.gtp()} ({new_top[0][0]:.2%}) so will play top policy move instead."
aimove = top_policy_move
else:
aimove = top_policy_move
ai_thoughts += f"Pick policy strategy {ai_mode} failed to find legal moves, so is playing top policy move {aimove.gtp()}."
else:
raise ValueError(f"Unknown Policy-based AI mode {ai_mode}")
else: # Engine based move
candidate_ai_moves = cn.candidate_moves
if ai_mode == AI_HANDICAP:
candidate_ai_moves = handicap_analysis["moveInfos"]
top_cand = Move.from_gtp(candidate_ai_moves[0]["move"],
player=cn.next_player)
if top_cand.is_pass and ai_mode not in [
AI_DEFAULT, AI_HANDICAP
]: # don't play suicidal to balance score
aimove = top_cand
ai_thoughts += f"Top move is pass, so passing regardless of strategy. "
else:
if ai_mode == AI_JIGO:
sign = cn.player_sign(cn.next_player)
jigo_move = min(
candidate_ai_moves,
key=lambda move: abs(sign * move["scoreLead"] -
ai_settings["target_score"]))
aimove = Move.from_gtp(jigo_move["move"],
player=cn.next_player)
ai_thoughts += f"Jigo strategy found {len(candidate_ai_moves)} candidate moves (best {top_cand.gtp()}) and chose {aimove.gtp()} as closest to 0.5 point win"
elif ai_mode == AI_SCORELOSS:
c = ai_settings["strength"]
moves = [(
d["pointsLost"],
math.exp(min(200, -c * max(0, d["pointsLost"]))),
Move.from_gtp(d["move"], player=cn.next_player),
) for d in candidate_ai_moves]
topmove = weighted_selection_without_replacement(moves, 1)[0]
aimove = topmove[2]
ai_thoughts += f"ScoreLoss strategy found {len(candidate_ai_moves)} candidate moves (best {top_cand.gtp()}) and chose {aimove.gtp()} (weight {topmove[1]:.3f}, point loss {topmove[0]:.1f}) based on score weights."
else:
if ai_mode not in [AI_DEFAULT, AI_HANDICAP]:
game.katrain.log(
f"Unknown AI mode {ai_mode} or policy missing, using default.",
OUTPUT_INFO)
ai_thoughts += f"Strategy {ai_mode} not found or unexpected fallback."
aimove = top_cand
if ai_mode == AI_HANDICAP:
ai_thoughts += f"Handicap strategy found {len(candidate_ai_moves)} moves returned from the engine and chose {aimove.gtp()} as top move. PDA based score {cn.format_score(handicap_analysis['rootInfo']['scoreLead'])} and win rate {cn.format_winrate(handicap_analysis['rootInfo']['winrate'])}"
else:
ai_thoughts += f"Default strategy found {len(candidate_ai_moves)} moves returned from the engine and chose {aimove.gtp()} as top move"
game.katrain.log(f"AI thoughts: {ai_thoughts}", OUTPUT_DEBUG)
played_node = game.play(aimove)
played_node.ai_thoughts = ai_thoughts
return aimove, played_node
def generate_ai_move(game: Game, ai_mode: str,
ai_settings: Dict) -> Tuple[Move, GameNode]:
cn = game.current_node
if ai_mode == AI_HANDICAP:
pda = ai_settings["pda"]
if ai_settings["automatic"]:
n_handicaps = len(game.root.get_list_property("AB", []))
MOVE_VALUE = 14 # could be rules dependent
b_stones_advantage = max(
n_handicaps - 1, 0) - (cn.komi - MOVE_VALUE / 2) / MOVE_VALUE
pda = min(3, max(
-3, -b_stones_advantage *
(3 /
8))) # max PDA at 8 stone adv, normal 9 stone game is 8.46
handicap_analysis = request_ai_analysis(
game, cn, {
"playoutDoublingAdvantage": pda,
"playoutDoublingAdvantagePla": "BLACK"
})
if not handicap_analysis:
game.katrain.log(f"Error getting handicap-based move",
OUTPUT_ERROR)
ai_mode = AI_DEFAULT
while not cn.analysis_complete:
time.sleep(0.01)
game.engines[cn.next_player].check_alive(exception_if_dead=True)
ai_thoughts = ""
if (ai_mode
in AI_STRATEGIES_POLICY) and cn.policy: # pure policy based move
policy_moves = cn.policy_ranking
pass_policy = cn.policy[-1]
# dont make it jump around for the last few sensible non pass moves
top_5_pass = any([polmove[1].is_pass for polmove in policy_moves[:5]])
size = game.board_size
policy_grid = var_to_grid(cn.policy, size) # type: List[List[float]]
top_policy_move = policy_moves[0][1]
ai_thoughts += f"Using policy based strategy, base top 5 moves are {fmt_moves(policy_moves[:5])}. "
if (ai_mode == AI_POLICY and cn.depth <= ai_settings["opening_moves"]
) or (ai_mode in [AI_LOCAL, AI_TENUKI]
and not (cn.move and cn.move.coords)):
ai_mode = AI_WEIGHTED
ai_thoughts += f"Strategy override, using policy-weighted strategy instead. "
ai_settings = {
"pick_override": 0.9,
"weaken_fac": 1,
"lower_bound": 0.02
}
if top_5_pass:
aimove = top_policy_move
ai_thoughts += "Playing top one because one of them is pass."
elif ai_mode == AI_POLICY:
aimove = top_policy_move
ai_thoughts += f"Playing top policy move {aimove.gtp()}."
else: # weighted or pick-based
legal_policy_moves = [(pol, mv) for pol, mv in policy_moves
if not mv.is_pass and pol > 0]
board_squares = size[0] * size[1]
if ai_mode == AI_RANK: # calibrated, override from 0.8 at start to ~0.4 at full board
override = 0.8 * (
1 - 0.5 *
(board_squares - len(legal_policy_moves)) / board_squares)
overridetwo = 0.85 + max(0, 0.02 *
(ai_settings["kyu_rank"] - 8))
else:
override = ai_settings["pick_override"]
overridetwo = 1.0
if policy_moves[0][0] > override:
aimove = top_policy_move
ai_thoughts += f"Top policy move has weight > {override:.1%}, so overriding other strategies."
elif policy_moves[0][0] + policy_moves[1][0] > overridetwo:
aimove = top_policy_move
ai_thoughts += (
f"Top two policy moves have cumulative weight > {overridetwo:.1%}, so overriding other strategies."
)
elif ai_mode == AI_WEIGHTED:
aimove, ai_thoughts = policy_weighted_move(
policy_moves, ai_settings["lower_bound"],
ai_settings["weaken_fac"])
elif ai_mode in AI_STRATEGIES_PICK:
if ai_mode != AI_RANK:
n_moves = max(
1,
int(ai_settings["pick_frac"] *
len(legal_policy_moves) + ai_settings["pick_n"]))
else:
orig_calib_avemodrank = 0.063015 + 0.7624 * board_squares / (
10**(-0.05737 * ai_settings["kyu_rank"] + 1.9482))
norm_leg_moves = len(legal_policy_moves) / board_squares
modified_calib_avemodrank = (
0.3931 + 0.6559 * norm_leg_moves * math.exp(
-1 *
(3.002 * norm_leg_moves * norm_leg_moves -
norm_leg_moves - 0.034889 *
ai_settings["kyu_rank"] - 0.5097)**2) - 0.01093 *
ai_settings["kyu_rank"]) * orig_calib_avemodrank
n_moves = board_squares * norm_leg_moves / (
1.31165 * (modified_calib_avemodrank + 1) - 0.082653)
n_moves = max(1, round(n_moves))
if ai_mode in [
AI_INFLUENCE, AI_TERRITORY, AI_LOCAL, AI_TENUKI
]:
if cn.depth > ai_settings["endgame"] * board_squares:
weighted_coords = [(pol, 1, *mv.coords)
for pol, mv in legal_policy_moves]
x_ai_thoughts = (
f"Generated equal weights as move number >= {ai_settings['endgame'] * size[0] * size[1]}. "
)
n_moves = int(
max(n_moves,
len(legal_policy_moves) // 2))
elif ai_mode in [AI_INFLUENCE, AI_TERRITORY]:
weighted_coords, x_ai_thoughts = generate_influence_territory_weights(
ai_mode, ai_settings, policy_grid, size)
else: # ai_mode in [AI_LOCAL, AI_TENUKI]
weighted_coords, x_ai_thoughts = generate_local_tenuki_weights(
ai_mode, ai_settings, policy_grid, cn, size)
ai_thoughts += x_ai_thoughts
else: # ai_mode in [AI_PICK, AI_RANK]:
weighted_coords = [(policy_grid[y][x], 1, x, y)
for x in range(size[0])
for y in range(size[1])
if policy_grid[y][x] > 0]
pick_moves = weighted_selection_without_replacement(
weighted_coords, n_moves)
ai_thoughts += f"Picked {min(n_moves,len(weighted_coords))} random moves according to weights. "
if pick_moves:
new_top = [
(p, Move((x, y), player=cn.next_player))
for p, wt, x, y in heapq.nlargest(5, pick_moves)
]
aimove = new_top[0][1]
ai_thoughts += f"Top 5 among these were {fmt_moves(new_top)} and picked top {aimove.gtp()}. "
if new_top[0][0] < pass_policy:
ai_thoughts += f"But found pass ({pass_policy:.2%} to be higher rated than {aimove.gtp()} ({new_top[0][0]:.2%}) so will play top policy move instead."
aimove = top_policy_move
else:
aimove = top_policy_move
ai_thoughts += f"Pick policy strategy {ai_mode} failed to find legal moves, so is playing top policy move {aimove.gtp()}."
else:
raise ValueError(f"Unknown Policy-based AI mode {ai_mode}")
else: # Engine based move
candidate_ai_moves = cn.candidate_moves
if ai_mode == AI_HANDICAP:
candidate_ai_moves = handicap_analysis["moveInfos"]
top_cand = Move.from_gtp(candidate_ai_moves[0]["move"],
player=cn.next_player)
if top_cand.is_pass and ai_mode not in [
AI_DEFAULT,
AI_HANDICAP,
]: # don't play suicidal to balance score
aimove = top_cand
ai_thoughts += f"Top move is pass, so passing regardless of strategy. "
else:
if ai_mode == AI_JIGO:
sign = cn.player_sign(cn.next_player)
jigo_move = min(
candidate_ai_moves,
key=lambda move: abs(sign * move["scoreLead"] -
ai_settings["target_score"]))
aimove = Move.from_gtp(jigo_move["move"],
player=cn.next_player)
ai_thoughts += f"Jigo strategy found {len(candidate_ai_moves)} candidate moves (best {top_cand.gtp()}) and chose {aimove.gtp()} as closest to 0.5 point win"
elif ai_mode == AI_SCORELOSS:
c = ai_settings["strength"]
moves = [(
d["pointsLost"],
math.exp(min(200, -c * max(0, d["pointsLost"]))),
Move.from_gtp(d["move"], player=cn.next_player),
) for d in candidate_ai_moves]
topmove = weighted_selection_without_replacement(moves, 1)[0]
aimove = topmove[2]
ai_thoughts += f"ScoreLoss strategy found {len(candidate_ai_moves)} candidate moves (best {top_cand.gtp()}) and chose {aimove.gtp()} (weight {topmove[1]:.3f}, point loss {topmove[0]:.1f}) based on score weights."
elif ai_mode in [AI_SIMPLE_OWNERSHIP, AI_SETTLE_STONES]:
stones_with_player = {(*s.coords, s.player)
for s in game.stones}
next_player_sign = cn.player_sign(cn.next_player)
if ai_mode == AI_SIMPLE_OWNERSHIP:
def settledness(d, player_sign, player):
return sum([
abs(o) for o in d["ownership"]
if player_sign * o > 0
])
else:
board_size_x, board_size_y = game.board_size
def settledness(d, player_sign, player):
ownership_grid = var_to_grid(
d["ownership"], (board_size_x, board_size_y))
return sum([
abs(ownership_grid[s.coords[0]][s.coords[1]])
for s in game.stones if s.player == player
])
def is_attachment(move):
if move.is_pass:
return False
attach_opponent_stones = sum(
(move.coords[0] + dx, move.coords[1] + dy,
cn.player) in stones_with_player for dx in [-1, 0, 1]
for dy in [-1, 0, 1] if abs(dx) + abs(dy) == 1)
nearby_own_stones = sum(
(move.coords[0] + dx, move.coords[1] + dy,
cn.next_player) in stones_with_player
for dx in [-2, 0, 1, 2] for dy in [-2 - 1, 0, 1, 2]
if abs(dx) + abs(dy) <= 2 # allows clamps/jumps
)
return attach_opponent_stones >= 1 and nearby_own_stones == 0
def is_tenuki(d):
return not d.is_pass and not any(
not node or not node.move or node.move.is_pass or max(
abs(last_c - cand_c) for last_c, cand_c in zip(
node.move.coords, d.coords)) < 5
for node in [cn, cn.parent])
moves_with_settledness = sorted(
[(
move,
settledness(d, next_player_sign, cn.next_player),
settledness(d, -next_player_sign, cn.player),
is_attachment(move),
is_tenuki(move),
d,
) for d in candidate_ai_moves
if d["pointsLost"] < ai_settings["max_points_lost"]
and "ownership" in d and (
d["order"] <= 1
or d["visits"] >= ai_settings.get("min_visits", 1))
for move in
[Move.from_gtp(d["move"], player=cn.next_player)]
if not (move.is_pass and d["pointsLost"] > 0.75)],
key=lambda t: t[5]["pointsLost"] + ai_settings[
"attach_penalty"] * t[3] + ai_settings[
"tenuki_penalty"] * t[4] - ai_settings[
"settled_weight"] *
(t[1] + ai_settings["opponent_fac"] * t[2]),
)
if moves_with_settledness:
cands = [
f"{move.gtp()} ({d['pointsLost']:.1f} pt lost, {d['visits']} visits, {settled:.1f} settledness, {oppsettled:.1f} opponent settledness{', attachment' if isattach else ''}{', tenuki' if istenuki else ''})"
for move, settled, oppsettled, isattach, istenuki, d in
moves_with_settledness[:5]
]
ai_thoughts += f"{ai_mode} strategy. Top 5 Candidates {', '.join(cands)} "
aimove = moves_with_settledness[0][0]
else:
raise (Exception(
"No moves found - are you using an older KataGo with no per-move ownership info?"
))
else:
if ai_mode not in [AI_DEFAULT, AI_HANDICAP]:
game.katrain.log(
f"Unknown AI mode {ai_mode} or policy missing, using default.",
OUTPUT_INFO)
ai_thoughts += f"Strategy {ai_mode} not found or unexpected fallback."
aimove = top_cand
if ai_mode == AI_HANDICAP:
ai_thoughts += f"Handicap strategy found {len(candidate_ai_moves)} moves returned from the engine and chose {aimove.gtp()} as top move. PDA based score {cn.format_score(handicap_analysis['rootInfo']['scoreLead'])} and win rate {cn.format_winrate(handicap_analysis['rootInfo']['winrate'])}"
else:
ai_thoughts += f"Default strategy found {len(candidate_ai_moves)} moves returned from the engine and chose {aimove.gtp()} as top move"
game.katrain.log(f"AI thoughts: {ai_thoughts}", OUTPUT_DEBUG)
played_node = game.play(aimove)
played_node.ai_thoughts = ai_thoughts
return aimove, played_node
def analyze_extra(self, mode, **kwargs):
stones = {s.coords for s in self.stones}
cn = self.current_node
engine = self.engines[cn.next_player]
Clock.schedule_once(self.katrain.analysis_controls.hints.activate, 0)
if mode == "extra":
if kwargs.get("continuous", False):
visits = max(engine.config["max_visits"],
math.ceil(cn.analysis_visits_requested * 1.25))
else:
visits = cn.analysis_visits_requested + engine.config[
"max_visits"]
self.katrain.controls.set_status(
i18n._("extra analysis").format(visits=visits),
STATUS_ANALYSIS)
cn.analyze(engine,
visits=visits,
priority=-1_000,
time_limit=False)
return
if mode == "game":
nodes = self.root.nodes_in_tree
if "visits" in kwargs:
visits = kwargs["visits"]
else:
min_visits = min(node.analysis_visits_requested
for node in nodes)
visits = min_visits + engine.config["max_visits"]
for node in nodes:
node.analyze(engine,
visits=visits,
priority=-1_000_000,
time_limit=False)
self.katrain.controls.set_status(
i18n._("game re-analysis").format(visits=visits),
STATUS_ANALYSIS)
return
elif mode == "sweep":
board_size_x, board_size_y = self.board_size
if cn.analysis_ready:
policy_grid = (var_to_grid(self.current_node.policy,
size=(board_size_x, board_size_y))
if self.current_node.policy else None)
analyze_moves = sorted(
[
Move(coords=(x, y), player=cn.next_player)
for x in range(board_size_x)
for y in range(board_size_y)
if (policy_grid is None and
(x, y) not in stones) or policy_grid[y][x] >= 0
],
key=lambda mv: -policy_grid[mv.coords[1]][mv.coords[0]],
)
else:
analyze_moves = [
Move(coords=(x, y), player=cn.next_player)
for x in range(board_size_x) for y in range(board_size_y)
if (x, y) not in stones
]
visits = engine.config["fast_visits"]
self.katrain.controls.set_status(
i18n._("sweep analysis").format(visits=visits),
STATUS_ANALYSIS)
priority = -1_000_000_000
elif mode == "pass":
board_size_x, board_size_y = self.board_size
analyze_moves = [Move(coords=None, player=cn.next_player)]
visits = 4 * engine.config["fast_visits"]
self.katrain.controls.set_status(
i18n._("pass analysis").format(visits=visits), STATUS_ANALYSIS)
priority = -1_000_000_000
elif mode == "equalize":
if not cn.analysis_ready:
self.katrain.controls.set_status(
i18n._("wait-before-equalize"), STATUS_INFO,
self.current_node)
return
analyze_moves = [
Move.from_gtp(gtp, player=cn.next_player)
for gtp, _ in cn.analysis["moves"].items()
]
visits = max(d["visits"] for d in cn.analysis["moves"].values())
self.katrain.controls.set_status(
i18n._("equalizing analysis").format(visits=visits),
STATUS_ANALYSIS)
priority = -1_000
else:
raise ValueError("Invalid analysis mode")
for move in analyze_moves:
cn.analyze(engine,
priority,
visits=visits,
refine_move=move,
time_limit=False
) # explicitly requested so take as long as you need
def draw_board_contents(self, *_args):
if not (self.katrain and self.katrain.game):
return
katrain = self.katrain
board_size_x, board_size_y = katrain.game.board_size
if len(self.gridpos_x) < board_size_x or len(self.gridpos_y) < board_size_y:
return # race condition
show_n_eval = self.trainer_config["eval_off_show_last"]
with self.canvas:
self.canvas.clear()
# stones
current_node = katrain.game.current_node
game_ended = katrain.game.end_result
full_eval_on = katrain.analysis_controls.eval.active
has_stone = {}
drawn_stone = {}
for m in katrain.game.stones:
has_stone[m.coords] = m.player
show_dots_for = {
p: self.trainer_config["eval_show_ai"] or katrain.players_info[p].human for p in Move.PLAYERS
}
show_dots_for_class = self.trainer_config["show_dots"]
nodes = katrain.game.current_node.nodes_from_root
realized_points_lost = None
for i, node in enumerate(nodes[::-1]): # reverse order!
points_lost = node.points_lost
evalscale = 1
if points_lost and realized_points_lost:
if points_lost <= 0.5 and realized_points_lost <= 1.5:
evalscale = 0
else:
evalscale = min(1, max(0, realized_points_lost / points_lost))
placements = node.placements
for m in node.moves + placements:
if has_stone.get(m.coords) and not drawn_stone.get(m.coords): # skip captures, last only for
move_eval_on = show_dots_for.get(m.player) and (i < show_n_eval or full_eval_on)
if move_eval_on and points_lost is not None:
evalcol = self.eval_color(points_lost, show_dots_for_class)
else:
evalcol = None
inner = STONE_COLORS[m.opponent] if i == 0 and not m in placements else None
drawn_stone[m.coords] = m.player
self.draw_stone(
x=m.coords[0],
y=m.coords[1],
player=m.player,
innercol=inner,
evalcol=evalcol,
evalscale=evalscale,
)
realized_points_lost = node.parent_realized_points_lost
if katrain.game.current_node.is_root and katrain.debug_level >= 3: # secret ;)
for y in range(0, board_size_y):
evalcol = self.eval_color(16 * y / board_size_y)
self.draw_stone(0, y, "B", evalcol=evalcol, evalscale=y / (board_size_y - 1))
self.draw_stone(1, y, "B", innercol=STONE_COLORS["W"], evalcol=evalcol)
self.draw_stone(2, y, "W", evalcol=evalcol, evalscale=y / (board_size_y - 1))
self.draw_stone(3, y, "W", innercol=STONE_COLORS["B"], evalcol=evalcol)
# ownership - allow one move out of date for smooth animation
ownership = current_node.ownership or (current_node.parent and current_node.parent.ownership)
if katrain.analysis_controls.ownership.active and ownership:
rsz = self.grid_size * 0.2
if (
current_node.children
and katrain.controls.status_state[1] == STATUS_TEACHING
and current_node.children[-1].auto_undo
and current_node.children[-1].ownership
): # loss
loss_grid = var_to_grid(
[a - b for a, b in zip(current_node.children[-1].ownership, ownership)],
(board_size_x, board_size_y),
)
for y in range(board_size_y - 1, -1, -1):
for x in range(board_size_x):
loss = max(0, (-1 if current_node.children[-1].move.player == "B" else 1) * loss_grid[y][x])
if loss > 0:
Color(*EVAL_COLORS[self.trainer_config["theme"]][1][:3], loss)
Rectangle(
pos=(self.gridpos_x[x] - rsz / 2, self.gridpos_y[y] - rsz / 2), size=(rsz, rsz)
)
else:
ownership_grid = var_to_grid(ownership, (board_size_x, board_size_y))
for y in range(board_size_y - 1, -1, -1):
for x in range(board_size_x):
ix_owner = "B" if ownership_grid[y][x] > 0 else "W"
if ix_owner != (has_stone.get((x, y), -1)):
Color(*STONE_COLORS[ix_owner][:3], abs(ownership_grid[y][x]))
Rectangle(
pos=(self.gridpos_x[x] - rsz / 2, self.gridpos_y[y] - rsz / 2), size=(rsz, rsz)
)
policy = current_node.policy
if (
not policy
and current_node.parent
and current_node.parent.policy
and katrain.last_player_info.ai
and katrain.next_player_info.ai
):
policy = (
current_node.parent.policy
) # in the case of AI self-play we allow the policy to be one step out of date
pass_btn = katrain.board_controls.pass_btn
pass_btn.canvas.after.clear()
if katrain.analysis_controls.policy.active and policy:
policy_grid = var_to_grid(policy, (board_size_x, board_size_y))
best_move_policy = max(*policy)
for y in range(board_size_y - 1, -1, -1):
for x in range(board_size_x):
if policy_grid[y][x] > 0:
polsize = 1.1 * math.sqrt(policy_grid[y][x])
policy_circle_color = (
*POLICY_COLOR,
POLICY_ALPHA + TOP_POLICY_ALPHA * (policy_grid[y][x] == best_move_policy),
)
draw_circle(
(self.gridpos_x[x], self.gridpos_y[y]), polsize * self.stone_size, policy_circle_color
)
polsize = math.sqrt(policy[-1])
with pass_btn.canvas.after:
draw_circle(
(pass_btn.pos[0] + pass_btn.width / 2, pass_btn.pos[1] + pass_btn.height / 2),
polsize * pass_btn.height / 2,
POLICY_COLOR,
)
# pass circle
passed = len(nodes) > 1 and current_node.is_pass
if passed or game_ended:
if game_ended:
text = game_ended
katrain.controls.timer.paused = True
else:
text = i18n._("board-pass")
Color(0.45, 0.05, 0.45, 0.7)
center = (self.gridpos_x[int(board_size_x / 2)], self.gridpos_y[int(board_size_y / 2)])
size = min(self.width, self.height) * 0.227
Ellipse(pos=(center[0] - size / 2, center[1] - size / 2), size=(size, size))
Color(0.85, 0.85, 0.85)
draw_text(
pos=center, text=text, font_size=size * 0.25, halign="center", outline_color=[0.95, 0.95, 0.95]
)
self.draw_hover_contents()
