evaluation_kit/evaluation/reports.py

"""Build per-box, per-region and filtered-text reports for a page."""

from __future__ import annotations

import sys
from pathlib import Path

_BOX_GROUPING = str(Path(__file__).resolve().parent.parent / "box_grouping")
if _BOX_GROUPING not in sys.path:
    sys.path.insert(0, _BOX_GROUPING)

from typing import Any

from models import AnnotationBox, annotation_box_type, annotation_box_metadata
from spatial import (
    normalize_whitespace,
    rounded_box,
    flatten_text_units,
    group_items_left_to_right_top_to_bottom,
)
from text_metrics import (
    safe_mean,
    safe_error_rate,
    build_cer_bucket_summary,
    compute_text_metrics,
)
from prediction import build_region_predicted_text
from models import gt_box_report


def filtered_box_report(annotation_box: AnnotationBox) -> dict[str, Any]:
    """Build a report dict for a box that was excluded by a label-based filter."""
    return {
        **annotation_box_metadata(annotation_box),
        "gt_box": rounded_box(annotation_box.bounds),
        "gt_normalized_text": normalize_whitespace(annotation_box.text),
        "has_transcription": annotation_box.has_transcription,
        "excluded_as_non_armenian_text": annotation_box.excluded_as_non_armenian_text,
    }


def build_region_summary(region_reports: list[dict[str, Any]]) -> dict[str, Any]:
    """Aggregate CER/bucket stats from a list of already-built region reports.

    Shared by build_ocr_region_reports (per-page) and aggregate_reports
    (across pages) so the two levels can't drift apart on which fields get
    computed how.
    """
    total_gt_chars = 0
    total_char_distance = 0
    total_char_distance_lower = 0
    region_cers: list[float] = []
    normal_single_box_region_cers: list[float] = []
    normal_single_box_region_gt_chars = 0
    normal_single_box_region_char_distance = 0
    multibox_region_count = 0

    for region in region_reports:
        text_metrics = region["text_metrics"]
        total_gt_chars += text_metrics["gt_char_count"]
        total_char_distance += text_metrics["char_edit_distance"]
        total_char_distance_lower += text_metrics["char_edit_distance_lowercase"]
        region_cers.append(text_metrics["cer"])
        if len(region["box_ids"]) > 1:
            multibox_region_count += 1
        if region.get("normal_single_box_region"):
            normal_single_box_region_cers.append(text_metrics["cer"])
            normal_single_box_region_gt_chars += text_metrics["gt_char_count"]
            normal_single_box_region_char_distance += text_metrics[
                "char_edit_distance"
            ]

    return {
        "ocr_region_count": len(region_reports),
        "multibox_region_count": multibox_region_count,
        "mean_cer": safe_mean(region_cers),
        "gt_char_count": total_gt_chars,
        "char_edit_distance": total_char_distance,
        "char_edit_distance_lowercase": total_char_distance_lower,
        "cer": safe_error_rate(total_char_distance, total_gt_chars),
        "cer_lowercase": safe_error_rate(total_char_distance_lower, total_gt_chars),
        "cer_buckets": build_cer_bucket_summary(region_cers),
        "normal_single_box_region": {
            "count": len(normal_single_box_region_cers),
            "mean_cer": safe_mean(normal_single_box_region_cers),
            "cer": safe_error_rate(
                normal_single_box_region_char_distance,
                normal_single_box_region_gt_chars,
            ),
        },
    }


def build_ocr_region_reports(
    details: list[dict[str, Any]],
    annotation_boxes: list[AnnotationBox],
    predicted_rows_by_id: dict[str, Any],
    should_exclude_box: Any = None,
) -> tuple[list[dict[str, Any]], dict[str, Any]]:
    """Group boxes into reading-order regions and compute per-region text metrics."""
    annotation_box_by_id = {
        box.box_id: box for box in annotation_boxes if box.has_transcription
    }
    excluded_annotation_boxes = (
        [box for box in annotation_box_by_id.values() if should_exclude_box(box)]
        if should_exclude_box is not None
        else None
    )
    row_by_id = {row["row_id"]: row for row in details}
    adjacency: dict[str, set[str]] = {
        box_id: set() for box_id in annotation_box_by_id
    }

    for row in details:
        touched_text_box_ids = [
            box_id
            for box_id in row.get("touched_box_ids", [])
            if box_id in annotation_box_by_id
        ]
        if len(touched_text_box_ids) < 2:
            continue
        for left_position, left_box_id in enumerate(touched_text_box_ids):
            for right_box_id in touched_text_box_ids[left_position + 1 :]:
                adjacency[left_box_id].add(right_box_id)
                adjacency[right_box_id].add(left_box_id)

    visited: set[str] = set()
    region_reports: list[dict[str, Any]] = []

    for start_box_id in sorted(annotation_box_by_id):
        if start_box_id in visited:
            continue

        stack = [start_box_id]
        component_box_ids: list[str] = []
        while stack:
            current_box_id = stack.pop()
            if current_box_id in visited:
                continue
            visited.add(current_box_id)
            component_box_ids.append(current_box_id)
            stack.extend(sorted(adjacency.get(current_box_id, ())))

        ordered_box_columns = group_items_left_to_right_top_to_bottom(
            [annotation_box_by_id[box_id] for box_id in component_box_ids],
            lambda box: box.bounds,
        )
        ordered_boxes = [
            box for column_boxes in ordered_box_columns for box in column_boxes
        ]
        if should_exclude_box is not None:
            ordered_boxes = [
                box for box in ordered_boxes if not should_exclude_box(box)
            ]
            if not ordered_boxes:
                continue
        ordered_box_ids = [box.box_id for box in ordered_boxes]
        gt_text = flatten_text_units([box.text for box in ordered_boxes])
        region_box_ids = set(ordered_box_ids)
        region_rows = [
            row
            for row in details
            if any(
                box_id in region_box_ids
                for box_id in row.get("touched_box_ids", [])
            )
        ]
        (
            predicted_text,
            predicted_line_count,
            assigned_row_ids,
            predicted_box,
            join_word_indices,
        ) = build_region_predicted_text(
            region_rows,
            predicted_rows_by_id,
            ordered_boxes,
            excluded_annotation_boxes=excluded_annotation_boxes,
        )
        text_metrics = compute_text_metrics(
            gt_text,
            predicted_text,
            predicted_hyphen_join_word_indices=join_word_indices,
        )
        normal_single_box_region = (
            len(ordered_boxes) == 1
            and bool(assigned_row_ids)
            and all(
                row_by_id[row_id]["status"] == "exactly_one_box"
                for row_id in assigned_row_ids
            )
        )

        region_reports.append(
            {
                "region_id": ordered_box_ids[0],
                "box_ids": ordered_box_ids,
                "box_types": [
                    annotation_box_type(box) for box in ordered_boxes
                ],
                "gt_boxes": [rounded_box(box.bounds) for box in ordered_boxes],
                "gt_box_details": [gt_box_report(box) for box in ordered_boxes],
                "predicted_box": (
                    rounded_box(predicted_box) if predicted_box is not None else None
                ),
                "predicted_line_count": predicted_line_count,
                "normal_single_box_region": normal_single_box_region,
                "text_metrics": text_metrics,
            }
        )

    region_summary = build_region_summary(region_reports)
    return region_reports, region_summary