"""Classes for playing."""
from abc import ABC, abstractmethod
from dataclasses import dataclass
from typing import Optional, Callable, Tuple, Dict, Iterable
import chess
import torch
from torch.distributions import Categorical
from itertools import tee
from lczerolens.model import LczeroModel
from lczerolens.board import LczeroBoard
[docs]
class Sampler(ABC):
@abstractmethod
[docs]
def get_utilities(
self, boards: Iterable[LczeroBoard], **kwargs
) -> Iterable[Tuple[torch.Tensor, torch.Tensor, Dict[str, float]]]:
"""Get the utility of the board.
Parameters
----------
boards : Iterable[LczeroBoard]
The boards to evaluate.
Returns
-------
Iterable[Tuple[torch.Tensor, torch.Tensor, Dict[str, float]]]
The iterable over utilities, legal indices, and log dictionaries.
"""
pass
[docs]
def choose_move(self, board: LczeroBoard, utility: torch.Tensor, legal_indices: torch.Tensor) -> chess.Move:
"""Choose the next moves.
Parameters
----------
board : LczeroBoard
The board.
utility : torch.Tensor
The utility of the board.
legal_indices : torch.Tensor
The legal indices.
Returns
-------
Iterable[chess.Move]
The iterable over the moves.
"""
m = Categorical(logits=utility)
idx = m.sample()
return board.decode_move(legal_indices[idx])
[docs]
def get_next_moves(self, boards: Iterable[LczeroBoard], **kwargs) -> Iterable[Tuple[chess.Move, Dict[str, float]]]:
"""Get the next move.
Parameters
----------
boards : Iterable[LczeroBoard]
The boards to evaluate.
Returns
-------
Iterable[Tuple[chess.Move, Dict[str, float]]]
The iterable over the moves and log dictionaries.
"""
util_boards, move_boards = tee(boards)
for board, (utility, legal_indices, to_log) in zip(move_boards, self.get_utilities(util_boards, **kwargs)):
predicted_move = self.choose_move(board, utility, legal_indices)
yield predicted_move, to_log
[docs]
class RandomSampler(Sampler):
[docs]
def get_utilities(
self, boards: Iterable[LczeroBoard], **kwargs
) -> Iterable[Tuple[torch.Tensor, torch.Tensor, Dict[str, float]]]:
for board in boards:
legal_indices = board.get_legal_indices()
utilities = torch.ones_like(legal_indices, dtype=torch.float32)
yield utilities, legal_indices, {}
@dataclass
[docs]
class ModelSampler(Sampler):
[docs]
use_argmax: bool = True
[docs]
draw_score: float = 0.0
[docs]
m_slope: float = 0.0027
[docs]
q_threshold: float = 0.8
[docs]
def choose_move(self, board: LczeroBoard, utility: torch.Tensor, legal_indices: torch.Tensor) -> chess.Move:
if self.use_argmax:
idx = utility.argmax()
return board.decode_move(legal_indices[idx])
return super().choose_move(board, utility, legal_indices)
@torch.no_grad()
[docs]
def get_utilities(
self, boards: Iterable[LczeroBoard], **kwargs
) -> Iterable[Tuple[torch.Tensor, torch.Tensor, Dict[str, float]]]:
batch_size = kwargs.pop("batch_size", -1)
callback = kwargs.pop("callback", None)
for legal_indices, batch_stats in self._get_batched_stats(boards, batch_size, **kwargs):
to_log = {}
utility = 0
q_values = self._get_q_values(batch_stats, to_log)
utility += self.alpha * q_values
utility += self.beta * self._get_m_values(batch_stats, q_values, to_log)
utility += self.gamma * self._get_p_values(batch_stats, legal_indices, to_log)
to_log["max_utility"] = utility.max().item()
self._use_callback(callback, batch_stats, to_log)
yield utility, legal_indices, to_log
[docs]
def _get_batched_stats(self, boards, batch_size, use_next_boards=True, **kwargs):
next_batch = []
next_legal_indices = []
def generator(next_batch, next_legal_indices):
all_stats = self.model(*next_batch, **kwargs)
offset = 0
for legal_indices in next_legal_indices:
n_boards = legal_indices.shape[0] + 1 if use_next_boards else 1
batch_stats = all_stats[offset : offset + n_boards]
offset += n_boards
yield legal_indices, batch_stats
for board in boards:
legal_indices = board.get_legal_indices()
next_boards = list(board.get_next_legal_boards()) if use_next_boards else []
if len(next_batch) + len(next_boards) + 1 > batch_size and batch_size != -1:
yield from generator(next_batch, next_legal_indices)
next_batch = []
next_legal_indices = []
next_batch.extend([board] + next_boards)
next_legal_indices.append(legal_indices)
if next_batch:
yield from generator(next_batch, next_legal_indices)
[docs]
def _get_q_values(self, batch_stats, to_log):
if "value" in batch_stats.keys():
to_log["value"] = batch_stats["value"][0].item()
return batch_stats["value"][1:, 0]
elif "wdl" in batch_stats.keys():
to_log["wdl_w"] = batch_stats["wdl"][0][0].item()
to_log["wdl_d"] = batch_stats["wdl"][0][1].item()
to_log["wdl_l"] = batch_stats["wdl"][0][2].item()
scores = torch.tensor([1, self.draw_score, -1])
return batch_stats["wdl"][1:] @ scores
return torch.zeros(batch_stats.batch_size[0] - 1)
[docs]
def _get_m_values(self, batch_stats, q_values, to_log):
if "mlh" in batch_stats.keys():
to_log["mlh"] = batch_stats["mlh"][0].item()
delta_m_values = self.m_slope * (batch_stats["mlh"][1:, 0] - batch_stats["mlh"][0, 0])
delta_m_values.clamp_(-self.m_max, self.m_max)
scaled_q_values = torch.relu(q_values.abs() - self.q_threshold) / (1 - self.q_threshold)
poly_q_values = self.k_0 + self.k_1 * scaled_q_values + self.k_2 * scaled_q_values**2
return -q_values.sign() * delta_m_values * poly_q_values
return torch.zeros(batch_stats.batch_size[0] - 1)
[docs]
def _get_p_values(
self,
batch_stats,
legal_indices,
to_log,
):
if "policy" in batch_stats.keys():
legal_policy = batch_stats["policy"][0].gather(0, legal_indices)
to_log["max_legal_policy"] = legal_policy.max().item()
return legal_policy
return torch.zeros_like(legal_indices)
[docs]
def _use_callback(self, callback, batch_stats, to_log):
if callback is not None:
to_log_update = callback(batch_stats, to_log)
if not isinstance(to_log_update, dict):
raise ValueError("Callback must return a dictionary.")
to_log |= to_log_update
@dataclass
[docs]
class PolicySampler(ModelSampler):
[docs]
use_suboptimal: bool = False
@torch.no_grad()
[docs]
def get_utilities(
self, boards: Iterable[LczeroBoard], **kwargs
) -> Iterable[Tuple[torch.Tensor, torch.Tensor, Dict[str, float]]]:
batch_size = kwargs.pop("batch_size", -1)
callback = kwargs.pop("callback", None)
to_log = {}
for legal_indices, batch_stats in self._get_batched_stats(boards, batch_size, use_next_boards=False, **kwargs):
legal_policy = batch_stats["policy"][0].gather(0, legal_indices.to(batch_stats["policy"].device))
if self.use_suboptimal:
idx = legal_policy.argmax()
legal_policy[idx] = torch.tensor(-1e3)
self._use_callback(callback, batch_stats, to_log)
yield legal_policy, legal_indices, to_log
@dataclass
[docs]
class SelfPlay:
"""A class for generating games."""
[docs]
def play(
self,
board: Optional[LczeroBoard] = None,
max_moves: int = 100,
to_play: chess.Color = chess.WHITE,
report_fn: Optional[Callable[[dict, chess.Color], None]] = None,
white_kwargs: Optional[Dict] = None,
black_kwargs: Optional[Dict] = None,
):
"""
Plays a game.
"""
if board is None:
board = LczeroBoard()
if white_kwargs is None:
white_kwargs = {}
if black_kwargs is None:
black_kwargs = {}
game = []
if to_play == chess.BLACK:
move, _ = next(iter(self.black.get_next_moves([board], **black_kwargs)))
board.push(move)
game.append(move)
for _ in range(max_moves):
if board.is_game_over() or len(game) >= max_moves:
break
move, to_log = next(iter(self.white.get_next_moves([board], **white_kwargs)))
if report_fn is not None:
report_fn(to_log, board.turn)
board.push(move)
game.append(move)
if board.is_game_over() or len(game) >= max_moves:
break
move, to_log = next(iter(self.black.get_next_moves([board], **black_kwargs)))
if report_fn is not None:
report_fn(to_log, board.turn)
board.push(move)
game.append(move)
if board.is_game_over() or len(game) >= max_moves:
break
return game, board
