Upload utils/special_tokens.py with huggingface_hub
Browse files- utils/special_tokens.py +106 -0
utils/special_tokens.py
ADDED
|
@@ -0,0 +1,106 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""
|
| 2 |
+
Special tokens for visual primitives.
|
| 3 |
+
|
| 4 |
+
Paper format:
|
| 5 |
+
Box: <|ref|>TARGET<|/ref|><|box|>[[x1,y1,x2,y2],...]<|/box|>
|
| 6 |
+
Point: <|point|>[[x1,y1],[x2,y2],...]<|/point|>
|
| 7 |
+
|
| 8 |
+
Coordinates are normalized to discrete integers in [0, 999].
|
| 9 |
+
"""
|
| 10 |
+
|
| 11 |
+
from typing import List, Tuple
|
| 12 |
+
|
| 13 |
+
# Special token strings
|
| 14 |
+
REF_START = "<|ref|>"
|
| 15 |
+
REF_END = "<|/ref|>"
|
| 16 |
+
BOX_START = "<|box|>"
|
| 17 |
+
BOX_END = "<|/box|>"
|
| 18 |
+
POINT_START = "<|point|>"
|
| 19 |
+
POINT_END = "<|/point|>"
|
| 20 |
+
|
| 21 |
+
SPECIAL_TOKENS = [
|
| 22 |
+
REF_START, REF_END,
|
| 23 |
+
BOX_START, BOX_END,
|
| 24 |
+
POINT_START, POINT_END,
|
| 25 |
+
]
|
| 26 |
+
|
| 27 |
+
# For parsing
|
| 28 |
+
import re
|
| 29 |
+
|
| 30 |
+
BOX_PATTERN = re.compile(
|
| 31 |
+
r"<\|box\|>(.*?)<\|/box\|>", re.DOTALL
|
| 32 |
+
)
|
| 33 |
+
POINT_PATTERN = re.compile(
|
| 34 |
+
r"<\|point\|>(.*?)<\|/point\|>", re.DOTALL
|
| 35 |
+
)
|
| 36 |
+
REF_PATTERN = re.compile(
|
| 37 |
+
r"<\|ref\|>(.*?)<\|/ref\|>", re.DOTALL
|
| 38 |
+
)
|
| 39 |
+
|
| 40 |
+
|
| 41 |
+
def normalize_coordinate(val: float, image_size: int) -> int:
|
| 42 |
+
"""Map a pixel coordinate to [0, 999]."""
|
| 43 |
+
return int(round(val / image_size * 999))
|
| 44 |
+
|
| 45 |
+
|
| 46 |
+
def denormalize_coordinate(val: int, image_size: int) -> int:
|
| 47 |
+
"""Map a normalized coordinate [0, 999] back to pixel space."""
|
| 48 |
+
return int(round(val / 999 * image_size))
|
| 49 |
+
|
| 50 |
+
|
| 51 |
+
def format_box_token(boxes: List[Tuple[int, int, int, int]]) -> str:
|
| 52 |
+
"""
|
| 53 |
+
boxes: list of (x1, y1, x2, y2) in normalized [0, 999] ints.
|
| 54 |
+
Returns: <|box|>[[x1,y1,x2,y2],...]<|/box|>
|
| 55 |
+
"""
|
| 56 |
+
if not boxes:
|
| 57 |
+
return f"{BOX_START}[]{BOX_END}"
|
| 58 |
+
inner = ",".join(f"[{x1},{y1},{x2},{y2}]" for x1, y1, x2, y2 in boxes)
|
| 59 |
+
return f"{BOX_START}[{inner}]{BOX_END}"
|
| 60 |
+
|
| 61 |
+
|
| 62 |
+
def format_point_token(points: List[Tuple[int, int]]) -> str:
|
| 63 |
+
"""
|
| 64 |
+
points: list of (x, y) in normalized [0, 999] ints.
|
| 65 |
+
Returns: <|point|>[[x1,y1],[x2,y2],...]<|/point|>
|
| 66 |
+
"""
|
| 67 |
+
if not points:
|
| 68 |
+
return f"{POINT_START}[]{POINT_END}"
|
| 69 |
+
inner = ",".join(f"[{x},{y}]" for x, y in points)
|
| 70 |
+
return f"{POINT_START}[{inner}]{POINT_END}"
|
| 71 |
+
|
| 72 |
+
|
| 73 |
+
def parse_box_token(text: str) -> List[Tuple[int, int, int, int]]:
|
| 74 |
+
"""Parse all box groups from text. Returns list of (x1,y1,x2,y2)."""
|
| 75 |
+
boxes = []
|
| 76 |
+
for m in BOX_PATTERN.finditer(text):
|
| 77 |
+
content = m.group(1).strip()
|
| 78 |
+
# Match [[x1,y1,x2,y2],[x3,y3,x4,y4],...]
|
| 79 |
+
for bm in re.finditer(r"\[(\d+),(\d+),(\d+),(\d+)\]", content):
|
| 80 |
+
boxes.append(tuple(int(bm.group(i)) for i in range(1, 5)))
|
| 81 |
+
return boxes
|
| 82 |
+
|
| 83 |
+
|
| 84 |
+
def parse_point_token(text: str) -> List[Tuple[int, int]]:
|
| 85 |
+
"""Parse all point groups from text. Returns list of (x,y)."""
|
| 86 |
+
points = []
|
| 87 |
+
for m in POINT_PATTERN.finditer(text):
|
| 88 |
+
content = m.group(1).strip()
|
| 89 |
+
for pm in re.finditer(r"\[(\d+),(\d+)\]", content):
|
| 90 |
+
points.append((int(pm.group(1)), int(pm.group(2))))
|
| 91 |
+
return points
|
| 92 |
+
|
| 93 |
+
|
| 94 |
+
def parse_ref_text(text: str) -> List[str]:
|
| 95 |
+
"""Parse all reference texts from text."""
|
| 96 |
+
refs = []
|
| 97 |
+
for m in REF_PATTERN.finditer(text):
|
| 98 |
+
refs.append(m.group(1).strip())
|
| 99 |
+
return refs
|
| 100 |
+
|
| 101 |
+
|
| 102 |
+
def add_special_tokens(tokenizer):
|
| 103 |
+
"""Add visual primitive special tokens to a tokenizer."""
|
| 104 |
+
special = {"additional_special_tokens": SPECIAL_TOKENS}
|
| 105 |
+
tokenizer.add_special_tokens(special)
|
| 106 |
+
return tokenizer
|