eval_utils.py

"""
eval_utils.py — shared utilities for ChestX-ray8 grounding evaluation.

Coordinate conventions
──────────────────────
GT   : BBox_List_2017.csv stores (x, y, w, h) absolute pixels on a 1024×1024 image.
       Loaded → converted to (x1, y1, x2, y2) absolute pixels for IoU.

Pred : Lingshu / Qwen2.5-VL natively outputs [x1, y1, x2, y2] in range [0, 1000].
       We ask for this format in the prompt (unambiguous, no heuristic needed).
       parse_xyxy_1000()  →  scale by (image_size / 1000)  →  absolute pixels.

       If a future model outputs a different format, use the explicit helpers
       parse_xywh_abs() or parse_norm_xywh() instead.

Fallback : parse_bbox_from_text() tries xyxy-1000 first, then heuristic fallbacks,
           and returns (x1,y1,x2,y2) absolute pixels or None on failure.

Missing prediction : IoU = 0.0  (未检出 treated as worst case).
"""

import re
import json
import numpy as np
import pandas as pd
from pathlib import Path
from typing import Optional


# ─────────────────────────────────────────────────────────────────────────────
# 1. BBox conversion helpers
# ─────────────────────────────────────────────────────────────────────────────

def xywh_to_xyxy(x, y, w, h):
    """(x_topleft, y_topleft, width, height)  →  (x1, y1, x2, y2) absolute."""
    return float(x), float(y), float(x) + float(w), float(y) + float(h)


def xyxy_to_xywh(x1, y1, x2, y2):
    return x1, y1, x2 - x1, y2 - y1


def scale_xyxy(box, src_size: int, dst_size: int):
    """Rescale an (x1,y1,x2,y2) box from src_size to dst_size coordinate space."""
    ratio = dst_size / src_size
    return tuple(v * ratio for v in box)


# ─────────────────────────────────────────────────────────────────────────────
# 2. IoU / IoBB
# ─────────────────────────────────────────────────────────────────────────────

def compute_iou(box_a, box_b) -> float:
    """Standard IoU between two (x1,y1,x2,y2) boxes."""
    if box_a is None or box_b is None:
        return 0.0
    ix1 = max(box_a[0], box_b[0])
    iy1 = max(box_a[1], box_b[1])
    ix2 = min(box_a[2], box_b[2])
    iy2 = min(box_a[3], box_b[3])
    inter = max(0.0, ix2 - ix1) * max(0.0, iy2 - iy1)
    area_a = (box_a[2] - box_a[0]) * (box_a[3] - box_a[1])
    area_b = (box_b[2] - box_b[0]) * (box_b[3] - box_b[1])
    union = area_a + area_b - inter
    return inter / union if union > 0 else 0.0


def compute_iobb(pred_box, gt_box) -> float:
    """Intersection over predicted-Box area — used in original ChestX-ray8 paper."""
    if pred_box is None or gt_box is None:
        return 0.0
    ix1 = max(pred_box[0], gt_box[0])
    iy1 = max(pred_box[1], gt_box[1])
    ix2 = min(pred_box[2], gt_box[2])
    iy2 = min(pred_box[3], gt_box[3])
    inter = max(0.0, ix2 - ix1) * max(0.0, iy2 - iy1)
    area_pred = (pred_box[2] - pred_box[0]) * (pred_box[3] - pred_box[1])
    return inter / area_pred if area_pred > 0 else 0.0


# ─────────────────────────────────────────────────────────────────────────────
# 3. Dataset loading
# ─────────────────────────────────────────────────────────────────────────────

def load_bbox_csv(bbox_csv: str) -> pd.DataFrame:
    """
    Load BBox_List_2017.csv.

    Real column names in the file (after pandas CSV parse):
        'Image Index', 'Finding Label', 'Bbox [x', 'y', 'w', 'h]'
    (The bracket is split across columns by the comma separator.)

    Returns a clean DataFrame with columns:
        image_index, finding_label, x, y, w, h  (all float for x/y/w/h)

    Note: 984 rows, 880 unique images, 8 disease classes.
    """
    df = pd.read_csv(bbox_csv)
    df.columns = [c.strip() for c in df.columns]

    rename = {}
    for c in df.columns:
        lc = c.lower()
        if "image" in lc and "index" in lc:
            rename[c] = "image_index"
        elif "finding" in lc and "label" in lc:
            rename[c] = "finding_label"
        # CSV splits "Bbox [x" / "y" / "w" / "h]" — match each fragment
        elif re.search(r"bbox.*\[?x\]?$|^bbox \[x$", lc):
            rename[c] = "x"
        elif re.search(r"^y$", lc):
            rename[c] = "y"
        elif re.search(r"^w$", lc):
            rename[c] = "w"
        elif re.search(r"^h\]?$", lc):
            rename[c] = "h"

    df = df.rename(columns=rename)
    for col in ["x", "y", "w", "h"]:
        df[col] = df[col].astype(float)
    # drop stray unnamed columns
    df = df[[c for c in df.columns if not c.startswith("Unnamed")]]
    return df


# ─────────────────────────────────────────────────────────────────────────────
# 4. Output parsing
# ─────────────────────────────────────────────────────────────────────────────

_RE_FOUR_NUMS = re.compile(
    r"[\[\(]?\s*(\d+\.?\d*)\s*,\s*(\d+\.?\d*)\s*,\s*(\d+\.?\d*)\s*,\s*(\d+\.?\d*)\s*[\]\)]?"
)


def _clip_xyxy(x1, y1, x2, y2, image_size: int):
    x1 = max(0.0, min(float(x1), image_size))
    y1 = max(0.0, min(float(y1), image_size))
    x2 = max(0.0, min(float(x2), image_size))
    y2 = max(0.0, min(float(y2), image_size))
    return (x1, y1, x2, y2) if x2 > x1 and y2 > y1 else None


def parse_xyxy_1000(text: str, image_size: int = 1024) -> Optional[tuple]:
    """
    PRIMARY PARSER — for Qwen2.5-VL / Lingshu.

    Model is prompted to output [x1, y1, x2, y2] with values in [0, 1000]
    where (0,0) is top-left and (1000,1000) is bottom-right.

    Converts to absolute pixels: multiply by (image_size / 1000).

    Use this when the prompt explicitly requests the 0-1000 format.
    """
    if not text:
        return None
    m = _RE_FOUR_NUMS.search(text)
    if not m:
        return None
    a, b, c, d = [float(v) for v in m.groups()]
    scale = image_size / 1000.0
    return _clip_xyxy(a * scale, b * scale, c * scale, d * scale, image_size)


def parse_xywh_abs(text: str, image_size: int = 1024) -> Optional[tuple]:
    """
    Parser for models that output [x, y, w, h] in absolute pixels.
    (x,y) = top-left corner; w,h = width, height.
    """
    if not text:
        return None
    m = _RE_FOUR_NUMS.search(text)
    if not m:
        return None
    x, y, w, h = [float(v) for v in m.groups()]
    return _clip_xyxy(x, y, x + w, y + h, image_size)


def parse_norm_xywh(text: str, image_size: int = 1024) -> Optional[tuple]:
    """
    Parser for models that output [x, y, w, h] normalised to [0, 1].
    """
    if not text:
        return None
    m = _RE_FOUR_NUMS.search(text)
    if not m:
        return None
    x, y, w, h = [float(v) for v in m.groups()]
    s = image_size
    return _clip_xyxy(x * s, y * s, (x + w) * s, (y + h) * s, image_size)


def parse_bbox_from_text(text: str, image_size: int = 1024,
                         coord_format: str = "xyxy_1000") -> Optional[tuple]:
    """
    High-level dispatcher.  coord_format options:
        "xyxy_1000"  (default) — Qwen2.5-VL / Lingshu native
        "xywh_abs"             — absolute pixel xywh
        "norm_xywh"            — normalised 0-1 xywh
        "xyxy_abs"             — absolute pixel xyxy

    All return (x1, y1, x2, y2) in absolute pixels, or None.

    NOTE on missing predictions (未检出):
        If the model produces no parseable box, this returns None.
        The caller should treat None as IoU = 0.0.
    """
    if coord_format == "xyxy_1000":
        return parse_xyxy_1000(text, image_size)
    elif coord_format == "xywh_abs":
        return parse_xywh_abs(text, image_size)
    elif coord_format == "norm_xywh":
        return parse_norm_xywh(text, image_size)
    elif coord_format == "xyxy_abs":
        if not text:
            return None
        m = _RE_FOUR_NUMS.search(text)
        if not m:
            return None
        a, b, c, d = [float(v) for v in m.groups()]
        return _clip_xyxy(a, b, c, d, image_size)
    else:
        raise ValueError(f"Unknown coord_format: {coord_format!r}")


# ─────────────────────────────────────────────────────────────────────────────
# 5. Evaluation aggregation
# ─────────────────────────────────────────────────────────────────────────────

def evaluate_predictions(records: list, iou_thresholds=(0.1, 0.25, 0.5)) -> dict:
    """
    records: list of dicts with keys:
        image_index    : str
        finding_label  : str
        gt_xyxy        : (x1,y1,x2,y2) absolute pixels
        pred_xyxy      : (x1,y1,x2,y2) absolute pixels, or None  → IoU = 0

    Returns metrics dict.
    """
    from collections import defaultdict
    thresholds = list(iou_thresholds)

    per_class = defaultdict(lambda: {"ious": [], **{t: [] for t in thresholds}})
    all_ious = []
    parse_fail = 0

    for r in records:
        pred = r.get("pred_xyxy")
        if pred is None:
            parse_fail += 1
            iou = 0.0
        else:
            iou = compute_iou(r["gt_xyxy"], pred)

        label = r["finding_label"]
        all_ious.append(iou)
        per_class[label]["ious"].append(iou)
        for t in thresholds:
            per_class[label][t].append(float(iou >= t))

    overall = {}
    for t in thresholds:
        hits = [float(iou >= t) for iou in all_ious]
        overall[f"hit@{t}"] = float(np.mean(hits)) if hits else 0.0

    per_class_summary = {}
    for label, d in per_class.items():
        per_class_summary[label] = {
            "mean_iou": float(np.mean(d["ious"])) if d["ious"] else 0.0,
            **{f"hit@{t}": float(np.mean(d[t])) if d[t] else 0.0
               for t in thresholds},
        }

    return {
        "n_instances": len(records),
        "parse_fail": parse_fail,
        "parse_fail_rate": parse_fail / max(len(records), 1),
        "mean_iou": float(np.mean(all_ious)) if all_ious else 0.0,
        "overall": overall,
        "per_class": per_class_summary,
    }


# ─────────────────────────────────────────────────────────────────────────────
# 6. I/O helpers
# ─────────────────────────────────────────────────────────────────────────────

def print_results(metrics: dict, model_name: str = "Model"):
    print(f"\n{'='*62}")
    print(f"  {model_name}")
    print(f"  ChestX-ray8 Grounding  (984 instances, 880 images, 8 classes)")
    print(f"{'='*62}")
    print(f"  Total instances : {metrics['n_instances']}")
    print(f"  Parse failures  : {metrics['parse_fail']} "
          f"({metrics['parse_fail_rate']:.1%})")
    print(f"  Mean IoU        : {metrics['mean_iou']:.4f}")
    print()
    print("  Overall:")
    for k, v in metrics["overall"].items():
        print(f"    {k:12s} : {v:.4f}")
    print()
    print(f"  {'Class':<22} {'Hit@0.1':>8} {'Hit@0.25':>9} {'Hit@0.5':>8} {'mIoU':>8}")
    print(f"  {'-'*22} {'-'*8} {'-'*9} {'-'*8} {'-'*8}")
    for label, d in sorted(metrics["per_class"].items()):
        print(f"  {label:<22} "
              f"{d.get('hit@0.1',0):.4f}   "
              f"{d.get('hit@0.25',0):.4f}    "
              f"{d.get('hit@0.5',0):.4f}   "
              f"{d.get('mean_iou',0):.4f}")
    print(f"{'='*62}\n")


def save_results(metrics: dict, records: list,
                 out_dir: str, prefix: str):
    out_path = Path(out_dir)
    out_path.mkdir(parents=True, exist_ok=True)

    with open(out_path / f"{prefix}_metrics.json", "w") as f:
        json.dump(metrics, f, indent=2)

    rows = []
    for r in records:
        pred = r.get("pred_xyxy")
        rows.append({
            "image_index":   r["image_index"],
            "finding_label": r["finding_label"],
            "gt_x1": r["gt_xyxy"][0], "gt_y1": r["gt_xyxy"][1],
            "gt_x2": r["gt_xyxy"][2], "gt_y2": r["gt_xyxy"][3],
            "pred_x1": pred[0] if pred else None,
            "pred_y1": pred[1] if pred else None,
            "pred_x2": pred[2] if pred else None,
            "pred_y2": pred[3] if pred else None,
            "iou":        r.get("iou", 0.0),
            "raw_output": r.get("raw_output", ""),
        })
    pd.DataFrame(rows).to_csv(out_path / f"{prefix}_instances.csv", index=False)
    print(f"[saved] {out_path}/{prefix}_metrics.json  +  _instances.csv")