chess-clmrie / model.py
clmrie's picture
Chess Challenge submission by clmrie
8deeee2 verified
from __future__ import annotations
import json
import math
import os
from dataclasses import dataclass
from typing import Dict, List, Optional, Tuple
import torch
import torch.nn as nn
import torch.nn.functional as F
from transformers import PretrainedConfig, PreTrainedModel
from transformers.modeling_outputs import CausalLMOutputWithPast
from transformers.utils.hub import cached_file
def _is_square(tok: str) -> bool:
return len(tok) == 2 and tok[0] in "abcdefgh" and tok[1] in "12345678"
def _resolve_file(name_or_path: str, filename: str) -> str:
if isinstance(name_or_path, str) and os.path.isdir(name_or_path):
p = os.path.join(name_or_path, filename)
if os.path.exists(p):
return p
return cached_file(name_or_path, filename)
def _load_vocab(name_or_path: str) -> Tuple[Dict[str, int], Dict[int, str]]:
vocab_path = _resolve_file(name_or_path, "vocab.json")
with open(vocab_path, "r", encoding="utf-8") as f:
tok2id = json.load(f)
id2tok = {int(i): t for t, i in tok2id.items()}
return tok2id, id2tok
@dataclass
class TokenScheme:
W: str
B: str
pieces: Dict[str, str]
sep: Optional[str]
suffix: Dict[str, str]
prom: Dict[str, str]
pad_id: int
bos_id: int
eos_id: int
unk_id: int
def _detect_scheme(tok2id: Dict[str, int], config) -> TokenScheme:
W = "W" if "W" in tok2id else None
B = "B" if "B" in tok2id else None
if W is None or B is None:
raise ValueError("Cannot find W/B tokens in vocab")
pieces = {}
for p in ["P", "N", "B", "R", "Q", "K"]:
if p in tok2id:
pieces[p] = p
else:
raise ValueError(f"Cannot find piece token {p} in vocab")
sep = " " if " " in tok2id else None
suffix = {}
for k, v in [
("cap", "(x)"),
("cap_check", "(x*)"),
("cap_mate", "(x+*)"),
("check", "(+)"),
("mate", "(+*)"),
("o", "(o)"),
("O", "(O)"),
]:
if v in tok2id:
suffix[k] = v
prom = {}
for p, v in [("Q", "(Q)"), ("R", "(R)"), ("B", "(B)"), ("N", "(N)")]:
if v in tok2id:
prom[p] = v
pad_id = int(getattr(config, "pad_token_id", 0))
bos_id = int(getattr(config, "bos_token_id", 1))
eos_id = int(getattr(config, "eos_token_id", 2))
unk_id = int(getattr(config, "unk_token_id", 3))
return TokenScheme(W=W, B=B, pieces=pieces, sep=sep, suffix=suffix, prom=prom,
pad_id=pad_id, bos_id=bos_id, eos_id=eos_id, unk_id=unk_id)
class ChessConfig(PretrainedConfig):
model_type = "chess_transformer"
def __init__(
self,
vocab_size: int = 85,
n_embd: int = 128,
n_layer: int = 5,
n_head: int = 4,
n_ctx: int = 256,
n_inner: Optional[int] = None,
dropout: float = 0.1,
layer_norm_epsilon: float = 1e-5,
tie_weights: bool = False,
pad_token_id: int = 0,
bos_token_id: int = 1,
eos_token_id: int = 2,
unk_token_id: int = 3,
**kwargs,
):
self.vocab_size = int(vocab_size)
self.n_embd = int(n_embd)
self.n_layer = int(n_layer)
self.n_head = int(n_head)
self.n_ctx = int(n_ctx)
self.n_inner = int(n_inner) if n_inner is not None else 3 * int(n_embd)
self.dropout = float(dropout)
self.layer_norm_epsilon = float(layer_norm_epsilon)
self.tie_weights = bool(tie_weights)
kwargs["pad_token_id"] = pad_token_id
kwargs["bos_token_id"] = bos_token_id
kwargs["eos_token_id"] = eos_token_id
kwargs["unk_token_id"] = unk_token_id
super().__init__(**kwargs)
class MLP(nn.Module):
def __init__(self, config: ChessConfig):
super().__init__()
self.c_fc = nn.Linear(config.n_embd, config.n_inner)
self.c_proj = nn.Linear(config.n_inner, config.n_embd)
self.dropout = nn.Dropout(config.dropout)
def forward(self, x: torch.Tensor) -> torch.Tensor:
x = self.c_fc(x)
x = F.gelu(x)
x = self.c_proj(x)
x = self.dropout(x)
return x
class MultiHeadAttention(nn.Module):
def __init__(self, config: ChessConfig):
super().__init__()
assert config.n_embd % config.n_head == 0
self.n_head = config.n_head
self.head_dim = config.n_embd // config.n_head
self.c_attn = nn.Linear(config.n_embd, 3 * config.n_embd)
self.c_proj = nn.Linear(config.n_embd, config.n_embd)
self.dropout = nn.Dropout(config.dropout)
bias = torch.tril(torch.ones(config.n_ctx, config.n_ctx)).view(1, 1, config.n_ctx, config.n_ctx)
self.register_buffer("bias", bias, persistent=False)
def forward(self, x: torch.Tensor, attention_mask: Optional[torch.Tensor] = None) -> torch.Tensor:
B, T, C = x.size()
qkv = self.c_attn(x)
q, k, v = qkv.split(C, dim=2)
q = q.view(B, T, self.n_head, self.head_dim).transpose(1, 2)
k = k.view(B, T, self.n_head, self.head_dim).transpose(1, 2)
v = v.view(B, T, self.n_head, self.head_dim).transpose(1, 2)
att = (q @ k.transpose(-2, -1)) / math.sqrt(self.head_dim)
att = att.masked_fill(self.bias[:, :, :T, :T] == 0, float("-inf"))
if attention_mask is not None:
att = att.masked_fill(attention_mask.view(B, 1, 1, T) == 0, float("-inf"))
att = F.softmax(att, dim=-1)
att = self.dropout(att)
y = att @ v
y = y.transpose(1, 2).contiguous().view(B, T, C)
y = self.c_proj(y)
y = self.dropout(y)
return y
class Block(nn.Module):
def __init__(self, config: ChessConfig):
super().__init__()
self.ln_1 = nn.LayerNorm(config.n_embd, eps=config.layer_norm_epsilon)
self.attn = MultiHeadAttention(config)
self.ln_2 = nn.LayerNorm(config.n_embd, eps=config.layer_norm_epsilon)
self.mlp = MLP(config)
def forward(self, x: torch.Tensor, attention_mask: Optional[torch.Tensor] = None) -> torch.Tensor:
x = x + self.attn(self.ln_1(x), attention_mask=attention_mask)
x = x + self.mlp(self.ln_2(x))
return x
class ChessForCausalLM(PreTrainedModel):
config_class = ChessConfig
base_model_prefix = ""
def __init__(self, config: ChessConfig):
super().__init__(config)
self.wte = nn.Embedding(config.vocab_size, config.n_embd)
self.wpe = nn.Embedding(config.n_ctx, config.n_embd)
self.drop = nn.Dropout(config.dropout)
self.h = nn.ModuleList([Block(config) for _ in range(config.n_layer)])
self.ln_f = nn.LayerNorm(config.n_embd, eps=config.layer_norm_epsilon)
self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=False)
if getattr(config, "tie_weights", False):
self.lm_head.weight = self.wte.weight
self.post_init()
self._tok2id = None
self._id2tok = None
self._scheme = None
def _ensure_vocab(self):
if self._tok2id is None or self._id2tok is None:
name_or_path = getattr(self.config, "_name_or_path", None) or getattr(self, "name_or_path", None)
if not name_or_path:
raise ValueError("Cannot resolve model path to load vocab.json")
self._tok2id, self._id2tok = _load_vocab(name_or_path)
def _get_scheme(self) -> TokenScheme:
if self._scheme is None:
self._ensure_vocab()
self._scheme = _detect_scheme(self._tok2id, self.config)
return self._scheme
def forward(self, input_ids, attention_mask=None, labels=None, return_dict=True, **kwargs):
B, T = input_ids.shape
if T > self.config.n_ctx:
input_ids = input_ids[:, -self.config.n_ctx :]
if attention_mask is not None:
attention_mask = attention_mask[:, -self.config.n_ctx :]
if labels is not None:
labels = labels[:, -self.config.n_ctx :]
B, T = input_ids.shape
pos = torch.arange(0, T, device=input_ids.device).unsqueeze(0)
x = self.wte(input_ids) + self.wpe(pos)
x = self.drop(x)
for block in self.h:
x = block(x, attention_mask=attention_mask)
x = self.ln_f(x)
logits = self.lm_head(x)
loss = None
if labels is not None:
shift_logits = logits[:, :-1].contiguous()
shift_labels = labels[:, 1:].contiguous()
loss = F.cross_entropy(
shift_logits.view(-1, shift_logits.size(-1)),
shift_labels.view(-1),
ignore_index=-100,
)
if not return_dict:
return (logits, loss)
return CausalLMOutputWithPast(logits=logits, loss=loss)
def _ids_to_tokens(self, ids: List[int]) -> List[str]:
self._ensure_vocab()
return [self._id2tok.get(int(i), "[UNK]") for i in ids]
def _parse_history_to_board(self, input_ids_1d: List[int]):
import chess
scheme = self._get_scheme()
toks = self._ids_to_tokens(input_ids_1d)
specials = {"[PAD]", "[BOS]", "[EOS]", "[UNK]"}
toks = [t for t in toks if t not in specials]
b = chess.Board()
i = 0
while i < len(toks):
while i < len(toks) and toks[i] not in (scheme.W, scheme.B):
i += 1
if i >= len(toks):
break
i += 1
while i < len(toks) and scheme.sep is not None and toks[i] == scheme.sep:
i += 1
if i >= len(toks) or toks[i] not in scheme.pieces.values():
break
i += 1
while i < len(toks) and scheme.sep is not None and toks[i] == scheme.sep:
i += 1
if i >= len(toks) or not _is_square(toks[i]):
break
src = toks[i]
i += 1
while i < len(toks) and scheme.sep is not None and toks[i] == scheme.sep:
i += 1
if i >= len(toks) or not _is_square(toks[i]):
break
dst = toks[i]
i += 1
suffixes = []
while i < len(toks) and toks[i] not in (scheme.W, scheme.B):
if scheme.sep is not None and toks[i] == scheme.sep:
i += 1
continue
suffixes.append(toks[i])
i += 1
uci = f"{src}{dst}"
promo = None
for p, ptok in scheme.prom.items():
if ptok in suffixes:
promo = p.lower()
break
if promo is not None:
uci += promo
try:
mv = chess.Move.from_uci(uci)
if mv in b.legal_moves:
b.push(mv)
else:
break
except Exception:
break
return b
def _move_to_ids(self, board, move_uci: str) -> List[int]:
import chess
scheme = self._get_scheme()
self._ensure_vocab()
tok2id = self._tok2id
mv = chess.Move.from_uci(move_uci)
color_tok = scheme.W if board.turn == chess.WHITE else scheme.B
piece = board.piece_at(mv.from_square)
pl = piece.symbol().upper() if piece is not None else "P"
if pl not in scheme.pieces:
pl = "P"
src = chess.square_name(mv.from_square)
dst = chess.square_name(mv.to_square)
toks = [color_tok, pl]
if scheme.sep is not None:
toks += [scheme.sep, src, scheme.sep, dst]
else:
toks += [src, dst]
is_capture = board.is_capture(mv)
board.push(mv)
is_mate = board.is_checkmate()
is_check = board.is_check()
board.pop()
suffix_tok = None
if is_capture and is_mate:
suffix_tok = scheme.suffix.get("cap_mate")
elif is_capture and is_check:
suffix_tok = scheme.suffix.get("cap_check")
elif is_capture:
suffix_tok = scheme.suffix.get("cap")
elif is_mate:
suffix_tok = scheme.suffix.get("mate")
elif is_check:
suffix_tok = scheme.suffix.get("check")
if suffix_tok is not None:
toks.append(suffix_tok)
if mv.promotion is not None:
prom = chess.piece_symbol(mv.promotion).upper()
if prom in scheme.prom:
toks.append(scheme.prom[prom])
if scheme.sep is not None:
toks.append(scheme.sep)
return [tok2id.get(t, scheme.unk_id) for t in toks]
@torch.no_grad()
def _score_candidates(self, prefix_ids, cand_ids_list, attention_mask, temperature, batch_size=64):
device = prefix_ids.device
T0 = prefix_ids.size(1)
scores = torch.empty(len(cand_ids_list), device=device, dtype=torch.float32)
pad_id = int(self.config.pad_token_id)
for start in range(0, len(cand_ids_list), batch_size):
batch = cand_ids_list[start : start + batch_size]
max_c = max(len(c) for c in batch)
input_ids_list = []
attn_list = []
for c in batch:
c_ids = torch.tensor(c, device=device, dtype=torch.long).unsqueeze(0)
seq = torch.cat([prefix_ids, c_ids], dim=1)
pad_len = (T0 + max_c) - seq.size(1)
if pad_len > 0:
pad = torch.full((1, pad_len), pad_id, device=device, dtype=torch.long)
seq = torch.cat([seq, pad], dim=1)
input_ids_list.append(seq)
if attention_mask is None:
a = torch.ones((1, seq.size(1)), device=device, dtype=torch.long)
else:
a = attention_mask
if a.size(1) != T0:
a = a[:, -T0:]
ones = torch.ones((1, len(c)), device=device, dtype=torch.long)
zeros = torch.zeros((1, max_c - len(c)), device=device, dtype=torch.long)
a = torch.cat([a, ones, zeros], dim=1)
attn_list.append(a)
input_ids = torch.cat(input_ids_list, dim=0)
attn_mask = torch.cat(attn_list, dim=0)
out = self.forward(input_ids=input_ids, attention_mask=attn_mask, return_dict=True)
logits = out.logits / float(max(1e-6, temperature))
logp = torch.log_softmax(logits, dim=-1)
for bi, c in enumerate(batch):
lp = 0.0
for j in range(len(c)):
pos = T0 + j - 1
if pos < 0:
continue
tok_id = int(c[j])
lp += float(logp[bi, pos, tok_id].item())
scores[start + bi] = lp
return scores
def generate(self, input_ids=None, attention_mask=None, max_new_tokens=16, temperature=1.0, do_sample=False, **kwargs):
import chess
if input_ids is None:
raise ValueError("generate() requires input_ids")
if input_ids.dim() == 1:
input_ids = input_ids.unsqueeze(0)
if input_ids.size(0) != 1:
return super().generate(
input_ids=input_ids,
attention_mask=attention_mask,
max_new_tokens=max_new_tokens,
temperature=temperature,
do_sample=do_sample,
**kwargs,
)
try:
board = self._parse_history_to_board(input_ids[0].tolist())
except Exception:
board = None
if board is None or board.is_game_over():
return super().generate(
input_ids=input_ids,
attention_mask=attention_mask,
max_new_tokens=max_new_tokens,
temperature=temperature,
do_sample=do_sample,
**kwargs,
)
legal = list(board.legal_moves)
if not legal:
return input_ids
cand_ids_list = [self._move_to_ids(board, mv.uci()) for mv in legal]
scores = self._score_candidates(
prefix_ids=input_ids,
cand_ids_list=cand_ids_list,
attention_mask=attention_mask,
temperature=float(temperature),
batch_size=64,
)
best = int(torch.argmax(scores).item())
best_ids = torch.tensor(cand_ids_list[best], device=input_ids.device, dtype=torch.long).unsqueeze(0)
if best_ids.size(1) > int(max_new_tokens):
best_ids = best_ids[:, : int(max_new_tokens)]
return torch.cat([input_ids, best_ids], dim=1)