600 lines
21 KiB
Python
600 lines
21 KiB
Python
from __future__ import annotations
|
|
|
|
import argparse
|
|
import re
|
|
import time
|
|
import typing
|
|
from collections import defaultdict
|
|
from collections.abc import Callable, Sequence
|
|
from concurrent.futures import ThreadPoolExecutor
|
|
from pathlib import Path
|
|
|
|
from google import genai
|
|
from google.genai import types
|
|
from pydantic import BaseModel, Field
|
|
|
|
from copienator import configuration as config
|
|
from copienator import (
|
|
CliError,
|
|
EvaluationWorkspace,
|
|
ExitCode,
|
|
atomic_write_json,
|
|
execute,
|
|
read_json,
|
|
target_parser,
|
|
workspace_from_target,
|
|
)
|
|
from copienator.utils import natural_key, read_all_labels
|
|
|
|
MODEL_ID = config.MODEL_FOR_LABEL_ID
|
|
api_key = config.API_KEY
|
|
|
|
my_prompt = """I'm giving you an image of the left columns of a written exam.
|
|
Students answer several exercises, which can have several questions.
|
|
|
|
The image consists of several columns, separated by vertical black
|
|
lines. The image should be read top to bottom and then left to right,
|
|
meaning first column, then second column, etc.
|
|
|
|
In their sheet, students delimit exercises and questions using
|
|
delimiters such as `Ex 1`, or `Exercice 1`, and `1)` or `a)`. You need
|
|
to give me the bounding boxes of each delimiter.
|
|
|
|
When giving the bounding box of the first question of an exercise, the
|
|
box should be large enough to contain both the exercice label
|
|
(`Exercice i`) and the question label (`1)`) parts. If they are
|
|
horizontally far apart (example : if the `1)` is to the left and the
|
|
`Exercice i` is either to the right, or in the middle) then give only
|
|
the bounding box of the question label `1)` part. You should still
|
|
label it as `Exercice i : 1)` though.
|
|
|
|
You also need to give me the student name. It should appear on the top
|
|
left of the image. Disregard any mention of `MPSI 3`, it is their
|
|
class. A list of possible student names will be given below.
|
|
|
|
You will answer with a JSON object, containing a `name` field with the
|
|
name, and a `list` field, with the list of the bounding boxes and
|
|
their labels. The box_2d should be [ymin, xmin, ymax, xmax] normalized
|
|
to 0-1000.
|
|
|
|
Here is an example :
|
|
{\"name\" : \"John Doe\", \"list\" : [{\"box_2d\": (10, 20, 30, 40), \"label\" : \"Ex 1 : 1)\"}]}
|
|
|
|
Do not provide a box_2d for the name. Only for the labels. Order the
|
|
box_2d by their position in the page, column by column : first column
|
|
(top to bottom), then second column, etc.
|
|
|
|
You may find the same label present several times, as a student either
|
|
recall the current label on a new page, or adds content to its answer
|
|
later on. Give the position of each instance of each label.
|
|
|
|
For this exam you should look for the labels given below, separated by
|
|
newlines. A student need not have answered every question, so some may
|
|
be missing.
|
|
|
|
##labels##
|
|
|
|
##wrong_labels##
|
|
|
|
##wrong_label_text_context##
|
|
|
|
Here's a list of the names of the students, pick the one that matches
|
|
the best or `\"Unknown\"` if you cannot read the name
|
|
|
|
##names##"""
|
|
my_prompt2 = """I'm giving you an image of the left columns of a written exam.
|
|
Students answer several exercises, which can have several questions.
|
|
|
|
The image consists of several columns, separated by vertical black
|
|
lines. The image should be read top to bottom and then left to right,
|
|
meaning first column, then second column, etc.
|
|
|
|
In their sheet, students delimit exercises and questions using
|
|
delimiters such as `Ex 1`, or `Exercice 1`, and `1)` or `a)`. You need
|
|
to give me the bounding boxes of each delimiter.
|
|
|
|
When giving the bounding box of the first question of an exercise, the
|
|
box should be large enough to contain both the exercice label
|
|
(`Exercice i`) and the question label (`1)`) parts.
|
|
|
|
You also need to give me the student name. It should appear on the top
|
|
left of the image. Disregard any mention of `MPSI 3`, it is their
|
|
class. A list of possible student names will be given below.
|
|
|
|
You will answer with a JSON object, containing a `name` field with the
|
|
name, and a `list` field, with the list of the bounding boxes and
|
|
their labels. The box_2d should be [ymin, xmin, ymax, xmax] normalized
|
|
to 0-1000.
|
|
|
|
Here is an example :
|
|
{\"name\" : \"John Doe\", \"list\" : [{\"box_2d\": (10, 20, 30, 40), \"label\" : \"Ex 1 : 1)\"}]}
|
|
|
|
Do not provide a box_2d for the name. Only for the labels.
|
|
|
|
You may find the same label present several times, as a student either
|
|
recall the current label on a new page, or adds content to its answer
|
|
later on. Give the position of each instance of each label.
|
|
|
|
This image is one part of a sequence (e.g., part 2 of 3) for a single
|
|
student. Here is the list of labels found in the *previous* parts of
|
|
this copy:
|
|
|
|
[
|
|
##prev_context##
|
|
]
|
|
|
|
If the first column starts with a number like =3)= or =c)=, look at
|
|
the labels in the list above. If the last relevant label was =Ex 4 :
|
|
2)=, you should label the new box =Ex 4 : 3)=.
|
|
|
|
For this exam you should look for the labels given below, separated by
|
|
newlines. A student need not have answered every question, so some may
|
|
be missing.
|
|
|
|
##labels##
|
|
|
|
##wrong_labels##
|
|
|
|
##wrong_label_text_context##
|
|
|
|
Since this copy isn't the first part of a sequence, simply set the
|
|
name to `\"Continued\"`."""
|
|
|
|
class BoxItem(BaseModel):
|
|
box_2d: list[int] = Field(description="Bounding box coordinates (e.g., [ymin, xmin, ymax, xmax])")
|
|
label: str = Field(description="The label associated with the specific box")
|
|
|
|
class AnnotationData(BaseModel):
|
|
name: str = Field(description="The name identifier")
|
|
list: typing.List[BoxItem] = Field( # noqa: UP006 - field name shadows list
|
|
description="List of bounding box items"
|
|
)
|
|
|
|
|
|
TEXT_CONTEXT_MAX_CHARS = 4000
|
|
|
|
|
|
def _label_filename(path: Path) -> str:
|
|
return path.stem if path.suffix.casefold() in {".tex", ".txt"} else path.name
|
|
|
|
|
|
def _common_prefix_length(left: str, right: str) -> int:
|
|
left_folded = left.casefold()
|
|
right_folded = right.casefold()
|
|
limit = min(len(left_folded), len(right_folded))
|
|
for index in range(limit):
|
|
if left_folded[index] != right_folded[index]:
|
|
return index
|
|
return limit
|
|
|
|
|
|
def closest_text_context(
|
|
workspace: EvaluationWorkspace, wrong_labels: list[str]
|
|
) -> tuple[Path | None, str]:
|
|
"""Return a bounded excerpt from the Text file closest to an invalid label."""
|
|
text_dir = workspace.root / "Text"
|
|
if not wrong_labels or not text_dir.is_dir():
|
|
return None, ""
|
|
|
|
ranked: list[tuple[int, str, Path]] = []
|
|
for path in text_dir.iterdir():
|
|
if not path.is_file() or path.suffix.casefold() == ".pdf":
|
|
continue
|
|
filename = _label_filename(path)
|
|
prefix_length = max(
|
|
_common_prefix_length(filename, wrong_label)
|
|
for wrong_label in wrong_labels
|
|
)
|
|
if prefix_length:
|
|
ranked.append((prefix_length, filename.casefold(), path))
|
|
|
|
for _prefix_length, _filename, path in sorted(
|
|
ranked, key=lambda item: (-item[0], item[1])
|
|
):
|
|
try:
|
|
content = path.read_text(encoding="utf-8")
|
|
except (OSError, UnicodeError):
|
|
continue
|
|
if len(content) > TEXT_CONTEXT_MAX_CHARS:
|
|
content = content[:TEXT_CONTEXT_MAX_CHARS] + "\n[excerpt truncated]"
|
|
return path, content
|
|
return None, ""
|
|
|
|
|
|
def generate_request(
|
|
file,
|
|
labels,
|
|
names,
|
|
context_labels,
|
|
wrong_labels,
|
|
wrong_label_text_context="",
|
|
wrong_label_text_file: Path | None = None,
|
|
seed: int = 0,
|
|
):
|
|
"""Generates request for Gemini with context."""
|
|
|
|
image_path = Path(file)
|
|
|
|
# Format context list as a string
|
|
context_str = ", ".join([f'"{l}"' for l in context_labels]) if context_labels else "No previous context"
|
|
|
|
if context_labels == []:
|
|
text = my_prompt.replace("##labels##", labels)\
|
|
.replace("##names##", names)
|
|
else:
|
|
text = my_prompt2.replace("##labels##", labels)\
|
|
.replace("##prev_context##", context_str)
|
|
if wrong_labels:
|
|
formatted_wrong_labels = "\n".join(f'- "{label}"' for label in wrong_labels)
|
|
text = text.replace(
|
|
"##wrong_labels##",
|
|
"On the previous request for this image, you answered with these "
|
|
"invalid labels:\n"
|
|
f"{formatted_wrong_labels}\n"
|
|
"They are wrong because they do not exactly match any label in the "
|
|
"valid list above.\n\n"
|
|
"CRITICAL RETRY CONSTRAINT: NEVER return any of the invalid labels "
|
|
"listed above again. Your answer must use only exact labels copied "
|
|
"verbatim from the valid list. If the handwriting resembles an "
|
|
"invalid label, choose the closest exact valid label instead.",
|
|
)
|
|
else:
|
|
text = text.replace("##wrong_labels##", "")
|
|
|
|
if wrong_label_text_context and wrong_label_text_file:
|
|
text = text.replace(
|
|
"##wrong_label_text_context##",
|
|
"Here is an excerpt from the exam text file whose name has the "
|
|
"longest prefix in common with the invalid label(s), "
|
|
f"`{wrong_label_text_file.name}`:\n\n"
|
|
"<exam_text_excerpt>\n"
|
|
f"{wrong_label_text_context}\n"
|
|
"</exam_text_excerpt>\n\n"
|
|
"Use this excerpt as extra context for identifying the handwritten "
|
|
"label, but return only an exact label from the valid list above.",
|
|
)
|
|
else:
|
|
text = text.replace("##wrong_label_text_context##", "")
|
|
|
|
|
|
contents = [
|
|
types.Content(
|
|
role="user",
|
|
parts=[
|
|
types.Part.from_bytes(
|
|
data=image_path.read_bytes(),
|
|
mime_type="image/jpeg"
|
|
),
|
|
types.Part.from_text(text=text),
|
|
],
|
|
)
|
|
]
|
|
|
|
generate_content_config = types.GenerateContentConfig(
|
|
automatic_function_calling=types.AutomaticFunctionCallingConfig(disable=True),
|
|
temperature=1.0,
|
|
top_p=0.95,
|
|
seed=seed,
|
|
max_output_tokens=65535,
|
|
response_mime_type= "application/json",
|
|
response_json_schema= AnnotationData.model_json_schema(),
|
|
)
|
|
return (contents, generate_content_config)
|
|
|
|
TARGET_INTERVAL = 3.5
|
|
Sleep = Callable[[float], None]
|
|
|
|
|
|
def selected_images(
|
|
workspace: EvaluationWorkspace,
|
|
targets: list[Path],
|
|
) -> tuple[list[Path], list[str]]:
|
|
"""Resolve evaluation, copy-PDF, or Cutleft-image targets."""
|
|
workspace.require_directories("Copies", "Cutleft")
|
|
images: list[Path] = []
|
|
warnings: list[str] = []
|
|
for target in targets:
|
|
if target.is_dir():
|
|
copy_pdfs = sorted(
|
|
workspace.copies_dir.glob("Copie*.pdf"), key=natural_key
|
|
)
|
|
if not copy_pdfs:
|
|
warnings.append(f"No Copie*.pdf files found in {workspace.copies_dir}")
|
|
stems = [path.stem for path in copy_pdfs]
|
|
elif target.suffix.casefold() in {".jpg", ".jpeg"}:
|
|
if target.parent != workspace.cutleft_dir:
|
|
raise CliError(
|
|
f"Image target is not in {workspace.cutleft_dir}: {target}",
|
|
ExitCode.INVALID_ARGUMENTS,
|
|
)
|
|
images.append(target)
|
|
continue
|
|
elif target.suffix.casefold() == ".pdf":
|
|
stems = [target.stem]
|
|
else:
|
|
raise CliError(
|
|
f"Unsupported target for label detection: {target}",
|
|
ExitCode.INVALID_ARGUMENTS,
|
|
)
|
|
for stem in stems:
|
|
found = sorted(
|
|
workspace.cutleft_dir.glob(f"{stem}_*.jpg"), key=natural_key
|
|
)
|
|
if found:
|
|
images.extend(found)
|
|
else:
|
|
warnings.append(
|
|
f"No Cutleft image variants found for {stem} in "
|
|
f"{workspace.cutleft_dir}"
|
|
)
|
|
return list(dict.fromkeys(images)), warnings
|
|
|
|
|
|
def group_images(image_files: list[Path]) -> dict[str, list[Path]]:
|
|
groups: defaultdict[str, list[Path]] = defaultdict(list)
|
|
for image in image_files:
|
|
match = re.match(r"(.+)_(\d+)$", image.stem)
|
|
groups[match.group(1) if match else image.stem].append(image)
|
|
for files in groups.values():
|
|
files.sort(key=natural_key)
|
|
return dict(groups)
|
|
|
|
|
|
def sort_boxes_for_image(
|
|
workspace: EvaluationWorkspace,
|
|
image_file: Path,
|
|
boxes: list[BoxItem],
|
|
) -> list[BoxItem]:
|
|
"""Sort boxes in the image's page-column reading order when schema exists."""
|
|
match = re.match(r"(.+)_(\d+)$", image_file.stem)
|
|
if not match:
|
|
return boxes
|
|
schema_path = workspace.cutleft_dir / f"{match.group(1)}_schema.json"
|
|
try:
|
|
schema = read_json(schema_path)
|
|
columns_per_file = schema["columns_per_file"]
|
|
column_count = int(columns_per_file[int(match.group(2)) - 1])
|
|
if column_count < 1:
|
|
return boxes
|
|
except (OSError, KeyError, IndexError, TypeError, ValueError):
|
|
return boxes
|
|
|
|
def position(item: BoxItem) -> tuple[int, int, int]:
|
|
ymin, xmin, _ymax, xmax = item.box_2d
|
|
center_x = (xmin + xmax) // 2
|
|
column = min(column_count - 1, max(0, center_x * column_count // 1000))
|
|
return column, ymin, xmin
|
|
|
|
return sorted(boxes, key=position)
|
|
|
|
|
|
def _existing_context(output_json: Path) -> list[str]:
|
|
try:
|
|
loaded = read_json(output_json)
|
|
if not isinstance(loaded, dict):
|
|
return []
|
|
return [
|
|
str(item["label"])
|
|
for item in loaded.get("list", [])
|
|
if isinstance(item, dict) and "label" in item
|
|
]
|
|
except (OSError, TypeError, ValueError):
|
|
return []
|
|
|
|
|
|
def process_copy_group(
|
|
workspace: EvaluationWorkspace,
|
|
group_key: str,
|
|
files: list[Path],
|
|
*,
|
|
client,
|
|
labels_text: str,
|
|
names_text: str,
|
|
valid_labels: set[str],
|
|
valid_names: set[str],
|
|
overwrite: bool,
|
|
sleep: Sleep = time.sleep,
|
|
target_interval: float = TARGET_INTERVAL,
|
|
) -> int:
|
|
"""Process one student's image parts sequentially to preserve context."""
|
|
accumulated_labels: list[str] = []
|
|
generated = 0
|
|
for image_file in files:
|
|
started = time.monotonic()
|
|
output_json = workspace.copies_dir / f"{image_file.stem}.json"
|
|
if output_json.exists() and not overwrite:
|
|
print(f"[{group_key}] Skipping {image_file.name}, output exists.")
|
|
accumulated_labels.extend(_existing_context(output_json))
|
|
continue
|
|
|
|
print(
|
|
f"[{group_key}] Processing {image_file.name} with "
|
|
f"{len(accumulated_labels)} accumulated labels..."
|
|
)
|
|
attempt = 0
|
|
label_retry_count = 0
|
|
wrong_labels: list[str] = []
|
|
while True:
|
|
if attempt > 0:
|
|
sleep(10 * attempt)
|
|
try:
|
|
text_context_file, text_context = closest_text_context(
|
|
workspace, wrong_labels
|
|
)
|
|
if text_context_file:
|
|
print(
|
|
f"[{group_key}] Retry context for {image_file.name}: "
|
|
f"{text_context_file.relative_to(workspace.root)}"
|
|
)
|
|
request_seed = max(0, label_retry_count - 1)
|
|
contents, request_config = generate_request(
|
|
image_file,
|
|
labels_text,
|
|
names_text,
|
|
accumulated_labels,
|
|
wrong_labels,
|
|
text_context,
|
|
text_context_file,
|
|
seed=request_seed,
|
|
)
|
|
response = client.models.generate_content(
|
|
model=MODEL_ID,
|
|
contents=contents,
|
|
config=request_config,
|
|
)
|
|
annotation = AnnotationData.model_validate_json(response.text)
|
|
unknown = [
|
|
item.label
|
|
for item in annotation.list
|
|
if item.label not in valid_labels
|
|
]
|
|
if unknown:
|
|
print(
|
|
f"Error: {image_file.name} contained unknown labels: "
|
|
f"{unknown}"
|
|
)
|
|
unique_unknown = list(dict.fromkeys(unknown))
|
|
if (
|
|
label_retry_count >= 2
|
|
and set(unique_unknown) == set(wrong_labels)
|
|
):
|
|
for item in annotation.list:
|
|
if item.label in unique_unknown:
|
|
item.label = f"??{item.label}"
|
|
print(
|
|
f"Warning: {image_file.name} repeated the same unknown "
|
|
"label(s) on the third try; keeping them with a ?? prefix."
|
|
)
|
|
else:
|
|
wrong_labels = unique_unknown
|
|
label_retry_count += 1
|
|
attempt += 1
|
|
continue
|
|
if annotation.name not in valid_names:
|
|
print(
|
|
f"Error: {image_file.name} returned unknown name: "
|
|
f"{annotation.name}"
|
|
)
|
|
if attempt == 0:
|
|
attempt += 1
|
|
continue
|
|
annotation.name = "Unknown"
|
|
|
|
annotation.list = sort_boxes_for_image(
|
|
workspace, image_file, annotation.list
|
|
)
|
|
atomic_write_json(output_json, annotation.model_dump())
|
|
accumulated_labels.extend(box.label for box in annotation.list)
|
|
generated += 1
|
|
break
|
|
except KeyboardInterrupt:
|
|
raise
|
|
except Exception as exc: # noqa: BLE001 - remote API retry boundary
|
|
print(
|
|
f"Error processing {image_file.name}: {exc}\n"
|
|
"\tIt will be retried."
|
|
)
|
|
attempt += 1
|
|
sleep(max(0.0, target_interval - (time.monotonic() - started)))
|
|
return generated
|
|
|
|
|
|
def run(
|
|
workspace: EvaluationWorkspace,
|
|
targets: list[Path],
|
|
*,
|
|
overwrite: bool = False,
|
|
client=None,
|
|
sleep: Sleep = time.sleep,
|
|
max_workers: int = 12,
|
|
) -> ExitCode:
|
|
workspace.require_files("labels")
|
|
images, warnings = selected_images(workspace, targets)
|
|
for warning in warnings:
|
|
print(f"Warning: {warning}")
|
|
if not images:
|
|
return ExitCode.PARTIAL
|
|
|
|
all_labels = read_all_labels(workspace.root)
|
|
labels_text = "\n".join(all_labels) + "\n"
|
|
names_path = workspace.names_file()
|
|
if not names_path.is_file():
|
|
raise CliError(f"Names file not found: {names_path}", ExitCode.INVALID_WORKSPACE)
|
|
names_text = names_path.read_text(encoding="utf-8")
|
|
valid_names = {
|
|
line.strip() for line in names_text.splitlines() if line.strip()
|
|
} | {"Unknown", "Continued"}
|
|
if client is None:
|
|
client = genai.Client(api_key=api_key)
|
|
|
|
groups = group_images(images)
|
|
with ThreadPoolExecutor(max_workers=max_workers) as executor:
|
|
futures = [
|
|
executor.submit(
|
|
process_copy_group,
|
|
workspace,
|
|
group_key,
|
|
files,
|
|
client=client,
|
|
labels_text=labels_text,
|
|
names_text=names_text,
|
|
valid_labels=set(all_labels),
|
|
valid_names=valid_names,
|
|
overwrite=overwrite,
|
|
sleep=sleep,
|
|
)
|
|
for group_key, files in groups.items()
|
|
]
|
|
for future in futures:
|
|
future.result()
|
|
return ExitCode.PARTIAL if warnings else ExitCode.SUCCESS
|
|
|
|
|
|
def build_parser() -> argparse.ArgumentParser:
|
|
parser = target_parser("Detect handwritten question labels with Gemini")
|
|
parser.add_argument(
|
|
"additional_targets",
|
|
nargs="*",
|
|
type=Path,
|
|
help="Additional copy PDFs or Cutleft images from the same evaluation",
|
|
)
|
|
parser.add_argument(
|
|
"--overwrite",
|
|
action="store_true",
|
|
help="Regenerate JSON outputs that already exist",
|
|
)
|
|
return parser
|
|
|
|
|
|
def main(argv: Sequence[str] | None = None) -> int:
|
|
parser = build_parser()
|
|
|
|
def handle(args: argparse.Namespace) -> ExitCode:
|
|
workspace, target = workspace_from_target(
|
|
args, repository=Path(__file__).resolve().parents[2]
|
|
)
|
|
targets = [target]
|
|
for additional in args.additional_targets:
|
|
resolved = additional.expanduser().resolve()
|
|
if not resolved.exists():
|
|
raise CliError(
|
|
f"Target does not exist: {resolved}",
|
|
ExitCode.INVALID_WORKSPACE,
|
|
)
|
|
additional_workspace = EvaluationWorkspace.discover(
|
|
resolved, repository=workspace.repository
|
|
)
|
|
if additional_workspace.root != workspace.root:
|
|
raise CliError(
|
|
"All targets must belong to the same evaluation",
|
|
ExitCode.INVALID_ARGUMENTS,
|
|
)
|
|
targets.append(resolved)
|
|
return run(workspace, targets, overwrite=args.overwrite)
|
|
|
|
return execute(parser, argv, handle)
|
|
|
|
|
|
if __name__ == "__main__":
|
|
raise SystemExit(main())
|