Blame for scripts/evaluate.py

#!/usr/bin/env python3
"""Run CADGUI-Bench against any OpenAI-compatible multimodal API."""
import argparse
import json
import os
import sys
import time
from pathlib import Path

REPO_ROOT = Path(__file__).resolve().parents[1]
sys.path.insert(0, str(REPO_ROOT / "evaluation"))
import benchmark_evaluator as be  # noqa: E402
from openai import OpenAI  # noqa: E402


TASKS = {
    "1.1": ("localization", "任务1.1.json", "task1"),
    "1.2": ("localization", "任务1.2.json", "task1"),
    "1.3": ("localization", "任务1.3.json", "task1"),
    "2.1": ("localization", "任务2.1.json", "task2"),
    "2.2": ("localization", "任务2.2.json", "task2"),
    "2.3": ("localization", "任务2.3.json", "task2"),
    "3.1": ("localization", "任务3.1基础任务.json", "task3.1"),
    "3.2": ("localization", "任务3.2方位任务.json", "task3.2"),
    "3.3": ("localization", "任务3.3属性任务.json", "task3.3"),
    "3.4": ("localization", "任务3.4几何任务.json", "task3.4"),
    "3.5": ("localization", "任务3.5三维定位.json", "task3.5"),
    "4.1": ("action", "任务4.1平移缩放旋转.json", "task4.1"),
    "4.2": ("action", "任务4.2特殊操作.json", "task4.2"),
    "4.3": ("action", "任务4.3框选拖拽.json", "task4.3"),
    "5.1": ("multistep", "任务5.1_非画图操作.json", "task5.1"),
    "5.2": ("multistep", "任务5.2_画图操作.json", "task5.2"),
}


def load_json(path):
    with path.open("r", encoding="utf-8-sig") as stream:
        return json.load(stream)


def write_result(path, result):
    path.parent.mkdir(parents=True, exist_ok=True)
    path.write_text(json.dumps(result, ensure_ascii=False, indent=2), encoding="utf-8")


def run_localization(client, model, task_id, benchmark_path, image_dir, mode, tries):
    results = be.run_localization(
        client, model, str(benchmark_path), str(image_dir), mode=mode, tries=tries
    )
    total_tries = sum(row["tries"] for row in results)
    summary = {
        "task": task_id,
        "model": model,
        "mode": mode,
        "n_items": len(results),
        "hits": sum(row["hits"] for row in results),
        "tries": total_tries,
        "hit_rate": round(100 * sum(row["hits"] for row in results) / total_tries, 2) if total_tries else 0,
        "parse_fail": sum(row.get("parse_fail", 0) for row in results),
        "results": results,
    }
    if mode == "bbox":
        ious = [row["avg_iou"] for row in results if "avg_iou" in row]
        if ious:
            summary["mean_iou"] = round(sum(ious) / len(ious), 4)
            summary["acc50"] = round(100 * sum(row["acc50"] for row in results if "acc50" in row) / len(ious), 2)
    return summary


def run_action(client, model, task_id, benchmark_path, image_dir, tries):
    data = load_json(benchmark_path)
    rows = []
    for item in data:
        image = image_dir / item["image_path"]
        width, height = item["image_resolution"]
        prompt = be.PROMPT_ACTION.format(instruction=item["instruction"])
        action_hits = 0
        full_hits = 0
        parse_fail = 0
        for _ in range(tries):
            reply = be.call_vlm(client, model, str(image), prompt, max_tokens=300)
            if reply.startswith("API_ERROR"):
                parse_fail += 1
                continue
            score = be.eval_step(reply, item["ground_truth"], width, height)
            action_hits += int(score["action_ok"])
            full_hits += int(score["step_ok"])
        rows.append({
            "id": item["id"], "action_hits": action_hits, "full_hits": full_hits,
            "tries": tries, "parse_fail": parse_fail,
        })
    total = len(rows) * tries
    return {
        "task": task_id, "model": model, "n_items": len(rows), "tries_per_item": tries,
        "action_acc": round(100 * sum(row["action_hits"] for row in rows) / total, 2) if total else 0,
        "full_acc": round(100 * sum(row["full_hits"] for row in rows) / total, 2) if total else 0,
        "parse_fail": sum(row["parse_fail"] for row in rows), "results": rows,
    }


def run_multistep(client, model, task_id, benchmark_path, image_dir):
    data = load_json(benchmark_path)
    task_results = []
    for task in data:
        task_id_in_data = task["task_id"]
        instruction = task["high_level_instruction"]
        history = []
        steps = []
        first_error = None
        for step in task["steps"]:
            relative = step["image_path"].replace("\\", "/")
            relative = relative.replace("/raw_images/", "/")
            image = image_dir / relative
            if not image.exists():
                image = image_dir / f"record_{task_id_in_data:03d}" / "raw_images" / Path(relative).name
            if not image.exists():
                raise FileNotFoundError(f"Missing benchmark image: {image}")
            width, height = step["image_resolution"]
            history_text = "\n".join(history) if history else "(none)"
            prompt = (
                "You are controlling SOLIDWORKS from the current screenshot.\n"
                f"Task: {instruction}\n"
                f"Previously completed actions (in order):\n{history_text}\n"
                f"Choose the next single action for the current screenshot.\n"
                f"{be.ACTION_SPACE}\n\n{be.COORD_SYSTEM}\n"
                "Output exactly:\nAction: <action_name>\n"
                "Params: <[x,y] or [x1,y1] [x2,y2] or direction or text or NONE>"
            )
            reply = be.call_vlm(client, model, str(image), prompt, max_tokens=400)
            if reply.startswith("API_ERROR"):
                score = {"action_ok": False, "param_ok": False, "step_ok": False, "action": None}
            else:
                score = be.eval_step(reply, step["ground_truth"], width, height)
            steps.append({"step": step["step"], **score})
            if not score["step_ok"] and first_error is None:
                first_error = step["step"]
            history.append(f"Step {step['step']}: {step['ground_truth']['action_type']}")
        correct = sum(row["step_ok"] for row in steps)
        task_results.append({
            "task_id": task_id_in_data, "total_steps": len(steps), "correct_steps": correct,
            "step_accuracy": round(100 * correct / len(steps), 2) if steps else 0,
            "task_success": correct == len(steps), "first_error_step": first_error, "steps": steps,
        })
    item_count = len(task_results)
    total_steps = sum(row["total_steps"] for row in task_results)
    correct_steps = sum(row["correct_steps"] for row in task_results)
    return {
        "task": task_id, "model": model, "n_items": item_count, "total_steps": total_steps,
        "correct_steps": correct_steps,
        "step_accuracy": round(100 * correct_steps / total_steps, 2) if total_steps else 0,
        "task_success_rate": round(100 * sum(row["task_success"] for row in task_results) / item_count, 2) if item_count else 0,
        "results": task_results,
    }


def main():
    parser = argparse.ArgumentParser(description=__doc__)
    parser.add_argument("--task", choices=[*TASKS, "all"], default="all")
    parser.add_argument("--model", required=True, help="Model name sent to the OpenAI-compatible endpoint")
    parser.add_argument("--api-base-url", default=os.getenv("API_BASE_URL", "http://localhost:8000/v1"))
    parser.add_argument("--api-key", default=os.getenv("API_KEY", "EMPTY"))
    parser.add_argument("--mode", choices=["point", "bbox", "both"], default="point")
    parser.add_argument("--tries", type=int, default=None, help="Calls per item; defaults to 1 for tasks 1-2, 10 otherwise")
    parser.add_argument("--output-dir", type=Path, default=Path("results"))
    args = parser.parse_args()
    if args.tries is not None and args.tries < 1:
        parser.error("--tries must be at least 1")

    client = OpenAI(api_key=args.api_key, base_url=args.api_base_url, timeout=120, max_retries=2)
    selected_tasks = list(TASKS) if args.task == "all" else [args.task]
    for task_id in selected_tasks:
        kind, filename, image_folder = TASKS[task_id]
        benchmark_path = REPO_ROOT / "Benchmark" / filename
        image_dir = REPO_ROOT / "images" / image_folder
        supports_bbox = kind == "localization" and task_id.startswith(("1.", "2."))
        if supports_bbox and args.mode == "both":
            modes = ["point", "bbox"]
        elif supports_bbox:
            modes = [args.mode]
        elif kind == "localization":
            modes = ["point"]
        else:
            modes = [None]
        for mode in modes:
            started = time.perf_counter()
            tries = args.tries or (1 if task_id.startswith(("1.", "2.")) else 10)
            if kind == "localization":
                result = run_localization(client, args.model, task_id, benchmark_path, image_dir, mode, tries)
            elif kind == "action":
                result = run_action(client, args.model, task_id, benchmark_path, image_dir, tries)
            else:
                result = run_multistep(client, args.model, task_id, benchmark_path, image_dir)
            result["elapsed_seconds"] = round(time.perf_counter() - started, 2)
            suffix = f"_{mode}" if mode else ""
            output_path = args.output_dir / args.model.replace("/", "_") / f"task_{task_id}{suffix}.json"
            write_result(output_path, result)
            print(json.dumps({key: value for key, value in result.items() if key != "results"}, ensure_ascii=False))
            print(f"Detailed results: {output_path}")


if __name__ == "__main__":
    main()