File size: 12,157 Bytes
431ffa3 | 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 | # -*- coding: utf-8 -*-
"""Multimodal tokenizer processor."""
from __future__ import annotations
from dataclasses import dataclass, field
from typing import Any
from Tokenizer.unified.encoded import EncodedToken
from .image_placeholders import (
image_patch_count,
image_placeholder_tokens,
)
from .video_placeholders import (
video_patch_count,
video_placeholder_tokens,
)
from .tokens import IMAGE_PLACEHOLDER, VIDEO_PLACEHOLDER
@dataclass
class MultimodalEncoding:
input_ids: list[int]
attention_mask: list[int]
tokens: list[EncodedToken]
image_token_spans: list[tuple[int, int]]
video_token_spans: list[tuple[int, int]] = field(default_factory=list)
images: list[Any] = field(default_factory=list)
videos: list[Any] = field(default_factory=list)
pixel_values: Any = None
@dataclass(frozen=True)
class _PlaceholderRecord:
start: int
end: int
patch_count: int
class MultimodalProcessor:
def __init__(
self,
tokenizer,
image_processor=None,
patch_size: int = 14,
merge_size: int = 2,
temporal_patch_size: int = 2,
):
self.tokenizer = tokenizer
self.image_processor = image_processor
self.patch_size = patch_size
self.merge_size = merge_size
self.temporal_patch_size = temporal_patch_size
def __call__(
self,
text: str,
images: list[Any] | None = None,
image_sizes: list[Any] | None = None,
videos: list[Any] | None = None,
video_sizes: list[Any] | None = None,
add_bos: bool = False,
add_eos: bool = False,
) -> MultimodalEncoding:
image_list = list(images or [])
sizes = list(image_sizes or [])
video_list = list(videos or [])
vsizes = list(video_sizes or [])
n_img_holes = text.count(IMAGE_PLACEHOLDER)
n_vid_holes = text.count(VIDEO_PLACEHOLDER)
if len(image_list) != n_img_holes:
raise ValueError(
f"<image> placeholder count ({n_img_holes}) does not match "
f"images length ({len(image_list)})"
)
if len(video_list) != n_vid_holes:
raise ValueError(
f"<video> placeholder count ({n_vid_holes}) does not match "
f"videos length ({len(video_list)})"
)
if sizes and len(sizes) != n_img_holes:
raise ValueError(
f"<image> placeholder count ({n_img_holes}) does not match "
f"image_sizes length ({len(sizes)})"
)
if vsizes and len(vsizes) != n_vid_holes:
raise ValueError(
f"<video> placeholder count ({n_vid_holes}) does not match "
f"video_sizes length ({len(vsizes)})"
)
# Default sizes for missing entries: assume single-patch image/video.
if not sizes and n_img_holes:
sizes = [None] * n_img_holes
if not vsizes and n_vid_holes:
vsizes = [None] * n_vid_holes
expanded, image_records, video_records, boundary_map = self._expand_placeholders(
text, sizes, vsizes
)
result = self.tokenizer.encode_with_spans(
expanded, add_bos=add_bos, add_eos=add_eos
)
tokens = self._annotate_patches(
result.tokens, image_records, video_records, boundary_map
)
img_spans = self._collect_spans(
tokens, "<image_start>", "<image_patch>", "<image_end>", image_records
)
vid_spans = self._collect_spans(
tokens, "<video_start>", "<video_patch>", "<video_end>", video_records
)
# Invariant: attention_mask aligns with input_ids
assert len(result.attention_mask) == len(result.input_ids)
processed_images = image_list
if self.image_processor is not None and image_list:
processed_images = self.image_processor(image_list)
return MultimodalEncoding(
input_ids=result.input_ids,
attention_mask=result.attention_mask,
tokens=tokens,
image_token_spans=img_spans,
video_token_spans=vid_spans,
images=processed_images,
videos=video_list,
)
# ---- expansion ----
def _expand_placeholders(
self, text: str, image_sizes: list[Any], video_sizes: list[Any]
) -> tuple[str, list[_PlaceholderRecord], list[_PlaceholderRecord], list[int]]:
out: list[str] = []
boundary_map = [0]
image_records: list[_PlaceholderRecord] = []
video_records: list[_PlaceholderRecord] = []
img_idx = 0
vid_idx = 0
i = 0
while i < len(text):
if text.startswith(IMAGE_PLACEHOLDER, i):
n = self._patches_image(image_sizes[img_idx])
end = i + len(IMAGE_PLACEHOLDER)
image_records.append(_PlaceholderRecord(i, end, n))
expanded = "".join(image_placeholder_tokens(n))
out.append(expanded)
boundary_map.extend([end] * len(expanded))
img_idx += 1
i = end
continue
if text.startswith(VIDEO_PLACEHOLDER, i):
n = self._patches_video(video_sizes[vid_idx])
end = i + len(VIDEO_PLACEHOLDER)
video_records.append(_PlaceholderRecord(i, end, n))
expanded = "".join(video_placeholder_tokens(n))
out.append(expanded)
boundary_map.extend([end] * len(expanded))
vid_idx += 1
i = end
continue
out.append(text[i])
boundary_map.append(i + 1)
i += 1
return "".join(out), image_records, video_records, boundary_map
def _expand_images(self, text: str, sizes: list[Any]) -> str:
parts = text.split(IMAGE_PLACEHOLDER)
if len(parts) == 1:
return text
out = [parts[0]]
for i, size in enumerate(sizes):
n = self._patches_image(size)
out.append("".join(image_placeholder_tokens(n)))
out.append(parts[i + 1])
return "".join(out)
def _expand_videos(self, text: str, sizes: list[Any]) -> str:
parts = text.split(VIDEO_PLACEHOLDER)
if len(parts) == 1:
return text
out = [parts[0]]
for i, size in enumerate(sizes):
n = self._patches_video(size)
out.append("".join(video_placeholder_tokens(n)))
out.append(parts[i + 1])
return "".join(out)
def _patches_image(self, size: Any) -> int:
if size is None:
return 1
if isinstance(size, dict):
w = int(size.get("width") or size.get("w"))
h = int(size.get("height") or size.get("h"))
else:
w, h = int(size[0]), int(size[1])
return image_patch_count(w, h, self.patch_size, self.merge_size)
def _patches_video(self, size: Any) -> int:
if size is None:
return 1
if isinstance(size, dict):
f = int(size.get("frames") or size.get("num_frames"))
w = int(size.get("width") or size.get("w"))
h = int(size.get("height") or size.get("h"))
else:
f, w, h = int(size[0]), int(size[1]), int(size[2])
return video_patch_count(
f, w, h, self.patch_size, self.temporal_patch_size, self.merge_size
)
# ---- annotation & spans ----
def _annotate_patches(
self,
tokens: list[EncodedToken],
image_records: list[_PlaceholderRecord],
video_records: list[_PlaceholderRecord],
boundary_map: list[int],
) -> list[EncodedToken]:
"""Attach image_index/video_index metadata to patch tokens."""
out: list[EncodedToken] = []
img_idx = -1
vid_idx = -1
in_img = False
in_vid = False
for t in tokens:
meta = t.metadata
start, end = self._remap_offsets(t.start, t.end, boundary_map)
if t.token == "<image_start>":
img_idx += 1
in_img = True
meta = self._placeholder_metadata(meta, image_records, img_idx, "image_index")
elif t.token == "<image_end>":
meta = self._placeholder_metadata(meta, image_records, img_idx, "image_index")
in_img = False
elif t.token == "<image_patch>" and in_img:
meta = self._placeholder_metadata(meta, image_records, img_idx, "image_index")
elif t.token == "<video_start>":
vid_idx += 1
in_vid = True
meta = self._placeholder_metadata(meta, video_records, vid_idx, "video_index")
elif t.token == "<video_end>":
meta = self._placeholder_metadata(meta, video_records, vid_idx, "video_index")
in_vid = False
elif t.token == "<video_patch>" and in_vid:
meta = self._placeholder_metadata(meta, video_records, vid_idx, "video_index")
if meta is not t.metadata:
source_span = meta.get("source_span") if meta else None
if source_span is not None:
start, end = int(source_span[0]), int(source_span[1])
out.append(
EncodedToken(
t.id, t.token, t.track, start, end, t.surface, meta
)
)
elif start != t.start or end != t.end:
out.append(
EncodedToken(
t.id, t.token, t.track, start, end, t.surface, t.metadata
)
)
else:
out.append(t)
return out
def _remap_offsets(
self, start: int, end: int, boundary_map: list[int]
) -> tuple[int, int]:
if start == -1 and end == -1:
return start, end
if not boundary_map:
return start, end
start_idx = min(max(start, 0), len(boundary_map) - 1)
end_idx = min(max(end, 0), len(boundary_map) - 1)
return boundary_map[start_idx], boundary_map[end_idx]
def _placeholder_metadata(
self,
meta: dict | None,
records: list[_PlaceholderRecord],
index: int,
index_key: str,
) -> dict:
merged = {**(meta or {}), index_key: index}
if 0 <= index < len(records):
record = records[index]
merged["source_span"] = [record.start, record.end]
return merged
def _collect_spans(
self,
tokens: list[EncodedToken],
start_tok: str,
patch_tok: str,
end_tok: str,
records: list[_PlaceholderRecord],
) -> list[tuple[int, int]]:
spans: list[tuple[int, int]] = []
start: int | None = None
for idx, t in enumerate(tokens):
if t.token == start_tok:
start = idx
elif t.token == end_tok and start is not None:
spans.append((start, idx + 1))
start = None
if len(spans) != len(records):
raise ValueError(
f"{start_tok}/{end_tok} span count ({len(spans)}) does not match "
f"placeholder count ({len(records)})"
)
for i, (span_start, span_end) in enumerate(spans):
patch_count = sum(1 for token in tokens[span_start:span_end] if token.token == patch_tok)
if patch_count != records[i].patch_count:
raise ValueError(
f"{patch_tok} count in span {i} is {patch_count}, "
f"expected {records[i].patch_count}"
)
if span_end - span_start != records[i].patch_count + 2:
raise ValueError(
f"{start_tok}/{end_tok} span length for placeholder {i} is "
f"{span_end - span_start}, expected {records[i].patch_count + 2}"
)
return spans
|