File size: 16,733 Bytes
fee0e43 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 | """Strict ``bpe.json`` I/O for both tokenization schemes.
The serialization layer owns only the artifact schema. Set-BPE spatial and temporal
merge semantics stay in :mod:`actioncodec2.setbpe` and plain-BPE merge
semantics in :mod:`actioncodec2.tokenization.bpe_vocabulary`, so training and deployment
cannot drift through two independent implementations of the same algorithm.
``config.set_bpe.scheme`` selects which serializer an artifact carries:
``"set_bpe"`` (the production default) or ``"bpe"``, the FAST-style baseline
that flattens the quantized cell matrix row-major before merging.
"""
from __future__ import annotations
import json
from collections.abc import Mapping, Sequence
from numbers import Integral
from pathlib import Path
import numpy as np
from ..common._validation import active_dimensions
from .bpe_vocabulary import BPEVocabulary
from ..physical.configuration import ActionCodec2Config, ConfigError
from .setbpe_spatial import SpatialVocabulary
from .setbpe_temporal import TemporalVocabulary
BPE_NAME = "bpe.json"
def _integer(value: object, path: str, *, positive: bool = False) -> int:
if isinstance(value, bool) or not isinstance(value, int):
raise ConfigError(f"{path} must be an integer")
minimum = 1 if positive else 0
if value < minimum:
qualifier = "positive" if positive else "non-negative"
raise ConfigError(f"{path} must be {qualifier}")
return int(value)
def _rules(payload: object, width: int, path: str) -> tuple[tuple[int, ...], ...]:
"""Parse an ordered merge list without accepting lossy coercions."""
if not isinstance(payload, Sequence) or isinstance(
payload, (str, bytes, bytearray)
):
raise ConfigError(f"{path} must be a list")
output = []
for rank, raw_rule in enumerate(payload):
rule_path = f"{path}[{rank}]"
if not isinstance(raw_rule, Sequence) or isinstance(
raw_rule, (str, bytes, bytearray)
):
raise ConfigError(f"{rule_path} must be a list of {width} integers")
if len(raw_rule) != width:
raise ConfigError(f"{rule_path} must contain exactly {width} integers")
output.append(
tuple(
_integer(value, f"{rule_path}[{index}]")
for index, value in enumerate(raw_rule)
)
)
return tuple(output)
def _unique_json_object(pairs: list[tuple[str, object]]) -> dict[str, object]:
output: dict[str, object] = {}
for key, value in pairs:
if key in output:
raise ConfigError(f"{BPE_NAME} contains duplicate field {key!r}")
output[key] = value
return output
def _validate_rule_references(
*,
atom_count: int,
lookahead: int,
spatial_merges: tuple[tuple[int, int], ...],
temporal_merges: tuple[tuple[int, int, int], ...],
) -> None:
"""Give artifact-level errors before constructing vocabulary metadata."""
for rank, (left, right) in enumerate(spatial_merges):
child = atom_count + rank
if left >= child or right >= child:
raise ConfigError(
f"spatial_merges[{rank}] must reference already-created tokens"
)
spatial_size = atom_count + len(spatial_merges)
for rank, (left, right, offset) in enumerate(temporal_merges):
child = spatial_size + rank
if left >= child or right >= child:
raise ConfigError(
f"temporal_merges[{rank}] must reference already-created tokens"
)
if not 1 <= offset < lookahead:
raise ConfigError(
f"temporal_merges[{rank}] offset must be positive and smaller than lookahead"
)
def _read_payload(directory: str | Path) -> dict[str, object]:
"""Parse ``bpe.json`` from an artifact directory, rejecting duplicate keys."""
path = Path(directory) / BPE_NAME
try:
return json.loads(
path.read_text(encoding="utf-8"), object_pairs_hook=_unique_json_object
)
except OSError:
raise
except json.JSONDecodeError as error:
raise ConfigError(f"cannot parse {path}: {error}") from error
def _dense_cells(
cells: np.ndarray,
bins: Sequence[int],
active_dims: Sequence[int] | None,
) -> tuple[np.ndarray, tuple[int, ...] | None]:
"""Validate one trajectory and expand it to dense ``[T,D_max]`` cells.
Compact ``[T,D_active]`` input may provide explicit global ``active_dims``.
Without it, compact input uses prefix dimensions. Dense ``[T,D_max]`` input
may mark absent slots with ``-1``.
"""
raw = np.asarray(cells)
if raw.dtype.kind not in {"i", "u"}:
raise ValueError("cells must contain integer indices")
values = np.asarray(raw, dtype=np.int64)
if values.ndim != 2 or len(values) < 1 or values.shape[1] < 1:
raise ValueError("cells must have non-empty shape [T, D_active]")
if active_dims is None:
if values.shape[1] > len(bins):
raise ValueError(f"cells width exceeds codebook dimension {len(bins)}")
if values.shape[1] < len(bins):
dimensions = tuple(range(values.shape[1]))
dense = np.full((len(values), len(bins)), -1, dtype=np.int64)
dense[:, : values.shape[1]] = values
values = dense
else:
dimensions = None
else:
dimensions = active_dimensions(active_dims, len(bins))
if len(dimensions) != values.shape[1]:
raise ValueError(
"active_dims must have one entry per compact cell coordinate"
)
if np.any(values < 0):
raise ValueError("compact cells must not contain -1")
dense = np.full((len(values), len(bins)), -1, dtype=np.int64)
dense[:, dimensions] = values
values = dense
counts = np.asarray(tuple(int(value) for value in bins), dtype=np.int64)
if np.any(values < -1) or np.any(values >= counts[None, :]):
raise ValueError("cells contain an invalid active or absent index")
return values, dimensions
class SetBPESerializer:
"""Lossless mapping between ``[T,D]`` physical cells and token IDs.
Merge lists are interpreted in creation order. Token IDs are therefore
stable across physical save/load cycles, and decoding can infer any positive
trajectory length directly from stream coverage.
"""
scheme = "set_bpe"
def __init__(
self,
*,
bins: Sequence[int],
lookahead: int,
spatial_merges: Sequence[Sequence[int]],
temporal_merges: Sequence[Sequence[int]],
) -> None:
self.lookahead = _integer(lookahead, "lookahead", positive=True)
self.bins = tuple(
_integer(value, f"bins[{index}]", positive=True)
for index, value in enumerate(bins)
)
if not self.bins:
raise ConfigError("bins must contain at least one cell slot")
self.spatial_merges = _rules(spatial_merges, 2, "spatial_merges")
self.temporal_merges = _rules(temporal_merges, 3, "temporal_merges")
_validate_rule_references(
atom_count=sum(self.bins),
lookahead=self.lookahead,
spatial_merges=self.spatial_merges,
temporal_merges=self.temporal_merges,
)
try:
self.spatial = SpatialVocabulary(self.bins, self.spatial_merges)
self.vocabulary = TemporalVocabulary(
self.spatial, self.lookahead, self.temporal_merges
)
except ValueError as error:
raise ConfigError(str(error)) from error
self.patterns = self.vocabulary.patterns
self.root = self.vocabulary.root
self.vocab_size = self.vocabulary.vocab_size
self.token_cell_counts = np.asarray(
[
sum(
int(self.spatial.support[spatial_token]).bit_count()
for _, spatial_token in self.patterns[token]
)
for token in range(self.vocab_size)
],
dtype=np.int64,
)
def encode(
self,
cells: np.ndarray,
*,
active_dims: Sequence[int] | None = None,
) -> list[int]:
"""Serialize one non-empty trajectory without emitting absent slots.
Compact ``[T,D_active]`` input may provide explicit global
``active_dims``. Without it, compact input uses prefix dimensions.
Dense ``[T,D_max]`` input may mark absent slots with ``-1``.
"""
values, _ = _dense_cells(cells, self.bins, active_dims)
return self.vocabulary.encode(values)
def decode(
self,
tokens: Sequence[int],
*,
active_dims: Sequence[int] | None = None,
) -> np.ndarray:
"""Recover dense cells, writing ``-1`` in semantically absent slots."""
if len(tokens) < 1:
raise ValueError("token stream must not be empty")
checked = []
for index, token in enumerate(tokens):
if isinstance(token, bool) or not isinstance(token, Integral):
raise ValueError(f"tokens[{index}] must be an integer")
value = int(token)
if not 0 <= value < self.vocab_size:
raise ValueError(f"tokens[{index}] is outside the vocabulary")
checked.append(value)
return self.vocabulary.decode(checked, active_dims=active_dims)
def to_dict(self) -> dict[str, object]:
"""Return the intentionally minimal, path-independent BPE artifact."""
return {
"spatial_merges": [list(rule) for rule in self.spatial_merges],
"temporal_merges": [list(rule) for rule in self.temporal_merges],
}
@classmethod
def from_dict(
cls,
payload: object,
*,
bins: Sequence[int],
lookahead: int,
) -> SetBPESerializer:
if not isinstance(payload, Mapping) or not all(
isinstance(key, str) for key in payload
):
raise ConfigError(f"{BPE_NAME} must contain one JSON object")
allowed = {"spatial_merges", "temporal_merges"}
unknown = sorted(set(payload) - allowed)
if unknown:
raise ConfigError(
f"{BPE_NAME} contains unknown field(s): {', '.join(unknown)}"
)
missing = sorted(allowed - set(payload))
if missing:
raise ConfigError(f"{BPE_NAME} is missing field(s): {', '.join(missing)}")
return cls(
bins=bins,
lookahead=lookahead,
spatial_merges=payload["spatial_merges"],
temporal_merges=payload["temporal_merges"],
)
@classmethod
def from_pretrained(
cls,
directory: str | Path,
*,
bins: Sequence[int],
lookahead: int,
) -> SetBPESerializer:
return cls.from_dict(_read_payload(directory), bins=bins, lookahead=lookahead)
def save_pretrained(self, directory: str | Path) -> Path:
"""Write ``bpe.json`` and return its path."""
path = Path(directory) / BPE_NAME
path.parent.mkdir(parents=True, exist_ok=True)
path.write_text(json.dumps(self.to_dict(), indent=2) + "\n", encoding="utf-8")
return path
class BPESerializer:
"""Lossless mapping between ``[T,D]`` physical cells and plain-BPE ids.
The FAST-style baseline: cells are serialized row-major over the active
support and merged as an ordinary symbol sequence. It shares the primitive
atoms, the artifact layout, and the total vocabulary budget with
:class:`SetBPESerializer`, so the two can be fitted on one corpus and
compared directly.
``lookahead`` bounds a token at ``lookahead * len(bins)`` primitive cells,
the plain-BPE analogue of the Set-BPE rule that a temporal token spans at
most ``lookahead`` control steps.
"""
scheme = "bpe"
def __init__(
self,
*,
bins: Sequence[int],
lookahead: int,
merges: Sequence[Sequence[int]],
) -> None:
self.lookahead = _integer(lookahead, "lookahead", positive=True)
self.bins = tuple(
_integer(value, f"bins[{index}]", positive=True)
for index, value in enumerate(bins)
)
if not self.bins:
raise ConfigError("bins must contain at least one cell slot")
self.merges = _rules(merges, 2, "merges")
atom_count = sum(self.bins)
for rank, (left, right) in enumerate(self.merges):
child = atom_count + rank
if left >= child or right >= child:
raise ConfigError(
f"merges[{rank}] must reference already-created tokens"
)
try:
self.vocabulary = BPEVocabulary(self.bins, self.merges)
except ValueError as error:
raise ConfigError(str(error)) from error
self.max_token_cells = self.lookahead * len(self.bins)
self.token_cell_counts = self.vocabulary.token_cells
oversized = int(np.count_nonzero(self.token_cell_counts > self.max_token_cells))
if oversized:
raise ConfigError(
f"{oversized} merge rule(s) cover more than the {self.max_token_cells} "
"primitive cells allowed by lookahead"
)
self.vocab_size = self.vocabulary.vocab_size
def encode(
self,
cells: np.ndarray,
*,
active_dims: Sequence[int] | None = None,
) -> list[int]:
"""Serialize one non-empty trajectory without emitting absent slots.
Input conventions match :meth:`SetBPESerializer.encode`: compact
``[T,D_active]`` may carry explicit global ``active_dims`` and
otherwise uses prefix dimensions, while dense ``[T,D_max]`` marks
absent slots with ``-1``.
"""
values, _ = _dense_cells(cells, self.bins, active_dims)
return self.vocabulary.encode(values)
def decode(
self,
tokens: Sequence[int],
*,
active_dims: Sequence[int] | None = None,
) -> np.ndarray:
"""Recover dense cells, writing ``-1`` in semantically absent slots."""
if len(tokens) < 1:
raise ValueError("token stream must not be empty")
checked = []
for index, token in enumerate(tokens):
if isinstance(token, bool) or not isinstance(token, Integral):
raise ValueError(f"tokens[{index}] must be an integer")
value = int(token)
if not 0 <= value < self.vocab_size:
raise ValueError(f"tokens[{index}] is outside the vocabulary")
checked.append(value)
return self.vocabulary.decode(checked, active_dims=active_dims)
def to_dict(self) -> dict[str, object]:
"""Return the intentionally minimal, path-independent BPE artifact."""
return {"merges": [list(rule) for rule in self.merges]}
@classmethod
def from_dict(
cls,
payload: object,
*,
bins: Sequence[int],
lookahead: int,
) -> BPESerializer:
if not isinstance(payload, Mapping) or not all(
isinstance(key, str) for key in payload
):
raise ConfigError(f"{BPE_NAME} must contain one JSON object")
unknown = sorted(set(payload) - {"merges"})
if unknown:
raise ConfigError(
f"{BPE_NAME} contains unknown field(s): {', '.join(unknown)}"
)
if "merges" not in payload:
raise ConfigError(f"{BPE_NAME} is missing field(s): merges")
return cls(bins=bins, lookahead=lookahead, merges=payload["merges"])
@classmethod
def from_pretrained(
cls,
directory: str | Path,
*,
bins: Sequence[int],
lookahead: int,
) -> BPESerializer:
return cls.from_dict(_read_payload(directory), bins=bins, lookahead=lookahead)
def save_pretrained(self, directory: str | Path) -> Path:
"""Write ``bpe.json`` and return its path."""
path = Path(directory) / BPE_NAME
path.parent.mkdir(parents=True, exist_ok=True)
path.write_text(json.dumps(self.to_dict(), indent=2) + "\n", encoding="utf-8")
return path
SERIALIZERS: dict[str, type[SetBPESerializer] | type[BPESerializer]] = {
SetBPESerializer.scheme: SetBPESerializer,
BPESerializer.scheme: BPESerializer,
}
def serializer_class(
config: ActionCodec2Config,
) -> type[SetBPESerializer] | type[BPESerializer]:
"""Return the serializer the configured tokenization scheme requires."""
return SERIALIZERS[config.set_bpe.scheme]
|