""" CAD-GUI Benchmark 统一评测引擎 (benchmark_evaluator.py) ======================================================== 整合了项目中所有任务的评测逻辑，服务器上直接用这一个文件即可。 包含： 1. 几何判定引擎（点/线[圆滑]/多边形/椭圆/避让区/拖拽起终点） 2. 模型输出解析（point / bbox / action / 坐标 / 方向 / 输入内容） 3. 坐标转换（0-1000 归一化 <-> 像素） 4. 各任务评测：定位(Point/BBox) / 动作类型+参数 / 多步规划(Teacher Forcing) 5. LLM-as-Judge 意图一致性判定 依赖: pip install shapely numpy scipy openai """ import json import math import re import numpy as np from shapely.geometry import Point, Polygon, LineString from scipy.interpolate import splprep, splev TOLERANCE = 5.0 # 像素容差（基于真实CAD鼠标点击物理容差） # ============================================================ # 1. 几何判定引擎 # ============================================================ def smooth_trajectory(kps): """对 >=3 个点的轨迹做 B-spline 圆滑插值，返回 LineString。 注意：标注点顺序必须沿曲线走向，否则会过冲/交叉。""" if len(kps) < 3: return LineString(kps) try: pts = np.array(kps) k = min(3, len(kps) - 1) tck, u = splprep([pts[:, 0], pts[:, 1]], s=0, k=k) u_new = np.linspace(0, 1, max(len(kps) * 10, 100)) x_new, y_new = splev(u_new, tck) return LineString(list(zip(x_new.tolist(), y_new.tolist()))) except Exception: return LineString(kps) def is_point_in_ellipse(px, py, el): """判断点是否在旋转椭圆内""" cx, cy, rx, ry = el['cx'], el['cy'], el['rx'], el['ry'] angle = math.radians(el.get('rotation', 0)) cos_a, sin_a = math.cos(angle), math.sin(angle) dx, dy = px - cx, py - cy rx2 = cos_a * dx + sin_a * dy ry2 = -sin_a * dx + cos_a * dy if rx == 0 or ry == 0: return False return (rx2**2 / rx**2) + (ry2**2 / ry**2) <= 1 def hit_region(px, py, region, tol=TOLERANCE): """判断像素点 (px,py) 是否命中一个区域 (含 polygons/keypoints/ellipses/arc_trajectory)""" p = Point(px, py) for poly in region.get("polygons", []): if len(poly) >= 3 and Polygon(poly).buffer(tol).contains(p): return True for el in region.get("ellipses", []): if is_point_in_ellipse(px, py, el): return True kps = region.get("keypoints", []) if len(kps) == 1: if p.distance(Point(kps[0][0], kps[0][1])) <= tol: return True elif len(kps) >= 2: traj = smooth_trajectory(kps) if len(kps) >= 3 else LineString(kps) if traj.distance(p) <= tol: return True arc = region.get("arc_trajectory", []) if len(arc) >= 2: traj = smooth_trajectory(arc) if len(arc) >= 3 else LineString(arc) if traj.distance(p) <= tol: return True return False def evaluate_prediction(px, py, ground_truth, tol=TOLERANCE): """核心判定：像素坐标是否命中目标且避开禁区。 先判 exclude（命中即False），再判 include（命中即True）。""" exclude = ground_truth.get("exclude", {}) if hit_region(px, py, exclude, tol): return False include = ground_truth.get("include", {}) return hit_region(px, py, include, tol) def compute_iou(box1, box2): """两个 bbox [xmin,ymin,xmax,ymax] 的 IoU""" x1, y1 = max(box1[0], box2[0]), max(box1[1], box2[1]) x2, y2 = min(box1[2], box2[2]), min(box1[3], box2[3]) inter = max(0, x2 - x1) * max(0, y2 - y1) a1 = (box1[2] - box1[0]) * (box1[3] - box1[1]) a2 = (box2[2] - box2[0]) * (box2[3] - box2[1]) union = a1 + a2 - inter return inter / union if union > 0 else 0 # ============================================================ # 2. 坐标转换 # ============================================================ def norm_to_pixel(nx, ny, w, h): """0-1000 归一化 -> 像素""" return nx / 1000.0 * w, ny / 1000.0 * h def pixel_bbox_to_norm(bbox_pixels, resolution): """像素 bbox -> 0-1000 归一化（用于 IoU 评测时把 GT 转到归一化空间）""" w, h = resolution return [bbox_pixels[0]/w*1000, bbox_pixels[1]/h*1000, bbox_pixels[2]/w*1000, bbox_pixels[3]/h*1000] def smart_resize(h, w, factor=28, min_pixels=3136, max_pixels=12845056): """Qwen2-VL/2.5-VL 的 smart_resize：返回 (resized_h, resized_w)。 像素坐标模型(UI-Venus/Holo2/InfiGUI)输出的是 resize 后空间的像素， 需按各模型的 min/max_pixels 反算回原图。""" import math h_bar = round(h / factor) * factor w_bar = round(w / factor) * factor if h_bar * w_bar > max_pixels: beta = math.sqrt((h * w) / max_pixels) h_bar = math.floor(h / beta / factor) * factor w_bar = math.floor(w / beta / factor) * factor elif h_bar * w_bar < min_pixels: beta = math.sqrt(min_pixels / (h * w)) h_bar = math.ceil(h * beta / factor) * factor w_bar = math.ceil(w * beta / factor) * factor return h_bar, w_bar # 各像素坐标模型的 smart_resize 参数（min_pixels, max_pixels） SMART_RESIZE_PARAMS = { "ui-venus": (2000000, 4800000), # 官方 README "holo2": (3136, 12845056), # Qwen2.5-VL 默认（Holo2 用默认 smart_resize） "infigui": (3136, 12845056), # 实测接近原图，用默认 "ui-tars": (3136, 12845056), # UI-TARS-1.5: Qwen2.5-VL 默认 smart_resize，输出绝对像素 "opencua": (3136, 12845056), # OpenCUA: Qwen2.5-VL 默认 smart_resize，输出绝对像素 } # ============================================================ # 3. 模型输出解析 # ============================================================ def parse_point(text): """解析 'Point: [x, y]' 或任意 [x,y]，返回 0-1000 范围的 [x,y] 或 None""" for m in re.findall(r'\[(\d+\.?\d*),\s*(\d+\.?\d*)\]', text): x, y = float(m[0]), float(m[1]) if 0 <= x <= 1000 and 0 <= y <= 1000: return [x, y] return None def parse_bbox(text): """解析 'Box: [xmin,ymin,xmax,ymax]'，返回 0-1000 的 4 元组或 None""" for m in re.findall(r'\[(\d+\.?\d*),\s*(\d+\.?\d*),\s*(\d+\.?\d*),\s*(\d+\.?\d*)\]', text): v = [float(x) for x in m] if all(0 <= x <= 1000 for x in v): return v return None def parse_action(text): """解析 'Action: xxx'，返回小写动作名或 None""" m = re.search(r'Action:\s*([a-z_]+)', text, re.IGNORECASE) return m.group(1).strip().lower() if m else None def parse_coords_list(text): """解析 Params 行里的所有 [x,y]，返回列表（用于拖拽/框选的起终点）""" pm = re.search(r'Params:\s*(.*)', text, re.IGNORECASE | re.DOTALL) seg = pm.group(1) if pm else text return [[float(a), float(b)] for a, b in re.findall(r'\[(\d+\.?\d*)\s*,\s*(\d+\.?\d*)\]', seg)] def parse_direction(text): """解析方向，归一化为英文 (up/down/left/right/left_up/...)""" pm = re.search(r'Params:\s*(.*)', text, re.IGNORECASE) seg = (pm.group(1).lower() if pm else text.lower()) for key in ['left_up', 'left_down', 'right_up', 'right_down', '向左上', '向右上', '向左下', '向右下', '上', '下', '左', '右', 'up', 'down', 'left', 'right']: if key in seg: return {'上': 'up', '下': 'down', '左': 'left', '右': 'right', '向左上': 'left_up', '向右上': 'right_up', '向左下': 'left_down', '向右下': 'right_down'}.get(key, key) return None def parse_thought(text): """解析 'Thought: ...'（到 Action 之前）""" m = re.search(r'Thought:\s*(.*?)(?=Action:|$)', text, re.IGNORECASE | re.DOTALL) return m.group(1).strip() if m else "" def parse_type_value(text): """解析 type_enter 的输入内容""" pm = re.search(r'Params:\s*(.*)', text, re.IGNORECASE) return pm.group(1).strip().strip('[]"\'') if pm else None # ---- 专用模型坐标解析器（各模型原生输出格式不同）---- def parse_paren_point(text): """UGround: 输出 '(x, y)' 圆括号，0-1000。取最后一个括号对。""" ms = re.findall(r'\(\s*(\d+\.?\d*)\s*,\s*(\d+\.?\d*)\s*\)', text) if ms: x, y = float(ms[-1][0]), float(ms[-1][1]) return [x, y] return None def parse_uitars_point(text): """UI-TARS: 'click(start_box='(x,y)')' 或 '(x,y)'，0-1000。""" m = re.search(r"start_box\s*=\s*['\"]?\(?\s*(\d+\.?\d*)\s*,\s*(\d+\.?\d*)", text) if m: return [float(m.group(1)), float(m.group(2))] return parse_paren_point(text) def parse_osatlas_point(text): """OS-Atlas: 可能输出 '(x1,y1),(x2,y2)' 圆括号 或 '[[x1,y1,x2,y2]]' 方括号，0-1000。 出框时取中心点；出点时直接用。会先复述指令文字，正则只抓坐标。""" # 先试方括号 bbox [[x1,y1,x2,y2]] 或 [x1,y1,x2,y2] m = re.search(r'\[+\s*(\d+\.?\d*)\s*,\s*(\d+\.?\d*)\s*,\s*(\d+\.?\d*)\s*,\s*(\d+\.?\d*)\s*\]', text) if m: v = [float(m.group(i)) for i in range(1, 5)] return [(v[0]+v[2])/2, (v[1]+v[3])/2] # 再试圆括号点对 pts = re.findall(r'\(\s*(\d+\.?\d*)\s*,\s*(\d+\.?\d*)\s*\)', text) if len(pts) >= 2: return [(float(pts[0][0])+float(pts[1][0]))/2, (float(pts[0][1])+float(pts[1][1]))/2] if len(pts) == 1: return [float(pts[0][0]), float(pts[0][1])] # 最后试方括号点 [x,y] m2 = re.search(r'\[\s*(\d+\.?\d*)\s*,\s*(\d+\.?\d*)\s*\]', text) if m2: return [float(m2.group(1)), float(m2.group(2))] return None def parse_osatlas_bbox(text): """OS-Atlas 出框：支持 '[[x1,y1,x2,y2]]' 方括号 或 '(x1,y1),(x2,y2)' 圆括号，0-1000。""" m = re.search(r'\[+\s*(\d+\.?\d*)\s*,\s*(\d+\.?\d*)\s*,\s*(\d+\.?\d*)\s*,\s*(\d+\.?\d*)\s*\]', text) if m: v = [float(m.group(i)) for i in range(1, 5)] return [min(v[0], v[2]), min(v[1], v[3]), max(v[0], v[2]), max(v[1], v[3])] pts = re.findall(r'\(\s*(\d+\.?\d*)\s*,\s*(\d+\.?\d*)\s*\)', text) if len(pts) >= 2: x1, y1 = float(pts[0][0]), float(pts[0][1]) x2, y2 = float(pts[1][0]), float(pts[1][1]) return [min(x1, x2), min(y1, y2), max(x1, x2), max(y1, y2)] return None def parse_json_point(text): """InfiGUI-G1: 去掉后 JSON [{"point_2d":[x,y]}]；像素坐标。 兼容 ```json 围栏与多候选（取第一个）。""" seg = text.split("")[-1] seg = seg.replace("```json", "").replace("```", "") m = re.search(r'"point_2d"\s*:\s*\[\s*(\d+\.?\d*)\s*,\s*(\d+\.?\d*)\s*\]', seg) if m: return [float(m.group(1)), float(m.group(2))] return None def parse_click_absolute(text): """Holo2: JSON {"action":"click_absolute","x":..,"y":..}；像素坐标。""" mx = re.search(r'"x"\s*:\s*(\d+\.?\d*)', text) my = re.search(r'"y"\s*:\s*(\d+\.?\d*)', text) if mx and my: return [float(mx.group(1)), float(my.group(1))] return None def parse_uivenus_bbox(text): """UI-Venus: 输出 [x1,y1,x2,y2]（smart_resize 像素），返回原始 4 元组。""" m = re.search(r'\[\s*(\d+\.?\d*)\s*,\s*(\d+\.?\d*)\s*,\s*(\d+\.?\d*)\s*,\s*(\d+\.?\d*)\s*\]', text) if m: return [float(m.group(i)) for i in range(1, 5)] return None def parse_uivenus_point(text): """UI-Venus: [x1,y1,x2,y2] -> 中心点（smart_resize 像素）。""" bb = parse_uivenus_bbox(text) if bb: return [(bb[0] + bb[2]) / 2, (bb[1] + bb[3]) / 2] return None def parse_pyautogui_point(text): """OpenCUA: 'pyautogui.click(x=1443, y=343)'，smart_resize 后绝对像素。 兼容 click/moveTo/dragTo 等，取第一个 x=,y=。""" m = re.search(r'x\s*=\s*(\d+\.?\d*)\s*,\s*y\s*=\s*(\d+\.?\d*)', text) if m: return [float(m.group(1)), float(m.group(2))] return None def parse_point_by_style(text, style): """按 parse_style 分发到对应解析器，统一返回原始数值对 [x,y] 或 None。 范围/归一化由各模型 cfg 决定，这里只负责从文本里把数提出来。""" if style == "paren": return parse_paren_point(text) if style == "uitars": return parse_uitars_point(text) if style == "osatlas": return parse_osatlas_point(text) if style == "json_point": # InfiGUI-G1 return parse_json_point(text) if style == "click_absolute": # Holo2 return parse_click_absolute(text) if style == "uivenus": # UI-Venus: bbox 中心点 return parse_uivenus_point(text) if style == "pyautogui": # OpenCUA: pyautogui.click(x=,y=) return parse_pyautogui_point(text) return parse_point(text) # bracket 默认 def parse_bbox_by_style(text, style): """按 parse_style 分发 bbox 解析。""" if style == "osatlas": return parse_osatlas_bbox(text) if style == "uivenus": return parse_uivenus_bbox(text) return parse_bbox(text) # bracket 默认 # ============================================================ # 4. 统一 Prompt 模板 # ============================================================ COORD_SYSTEM = """COORDINATE SYSTEM: - (0, 0) = top-left corner of the image - (1000, 1000) = bottom-right corner of the image - Simply estimate the target's position as a percentage of image width/height, then multiply by 10. - Do NOT assume any screen resolution (not 1920x1080, not 2560x1440).""" PROMPT_POINT = """You are a GUI expert. Given a screenshot and an action intent about interacting with a specific UI element, you need to identify the click point (x, y) of the target element, normalized to 0-1000. Note that the X-axis runs horizontally from left (0) to right (999), and the Y-axis runs vertically from top (0) to bottom (999). """ + COORD_SYSTEM + """ Action Intent: {instruction} IMPORTANT: Output ONLY the result in the exact format below. Do NOT explain. Output format: Point: [x, y]""" PROMPT_BBOX = """You are a GUI expert. Given a screenshot and an action intent, identify the bounding box [xmin, ymin, xmax, ymax] of the target element, normalized to 0-1000. """ + COORD_SYSTEM + """ Action Intent: {instruction} IMPORTANT: Output ONLY the result. Do NOT explain. Output format: Box: [xmin, ymin, xmax, ymax]""" ACTION_SPACE = """Available actions: - left_click [x,y] - double_click [x,y] - right_click [x,y] - ctrl_click [x,y] - drag [x1,y1] [x2,y2] - box_select [x1,y1] [x2,y2] - middle_drag [direction] # rotate 3D view - ctrl_middle_drag [direction] # pan view - shift_middle_drag [direction] # zoom: up=zoom in, down=zoom out - double_middle_click # reset/fit view - type_enter [text] - press_esc - press_delete""" PROMPT_ACTION = """You are a SOLIDWORKS operation expert. Given a screenshot and a user's intent, decide the SINGLE next action. """ + ACTION_SPACE + "\n\n" + COORD_SYSTEM + """ User intent: {instruction} Output STRICTLY: Action: Params: <[x,y] or [x1,y1] [x2,y2] or direction or text or NONE>""" PROMPT_MULTISTEP = """You are a SOLIDWORKS expert doing a multi-step task. High-level task: {high_level_instruction} """ + ACTION_SPACE + "\n\n" + COORD_SYSTEM + """ History (already done correctly): {history} Based on the CURRENT screenshot, decide the NEXT step. Output: Thought: Action: Params: <[x,y] or [x1,y1] [x2,y2] or text or NONE>""" # ============================================================ # 5. 单步评测（支持 alternatives 多方案 / 各种动作类型） # ============================================================ NEED_COORD = {"left_click", "double_click", "right_click", "ctrl_click"} NEED_DRAG = {"drag", "box_select"} NO_COORD = {"press_esc", "press_delete", "double_middle_click"} def eval_one_alternative(pred_action, coords, type_text, gt, w, h): """判定模型输出是否匹配单个 GT 方案。返回 (action_ok, param_ok)""" gt_action = gt["action_type"] action_ok = (pred_action == gt_action) if not action_ok: return False, False if gt_action in NEED_COORD: if coords: px, py = norm_to_pixel(coords[0][0], coords[0][1], w, h) return True, evaluate_prediction(px, py, gt) return True, False elif gt_action in NEED_DRAG: if len(coords) >= 2: ps = norm_to_pixel(coords[0][0], coords[0][1], w, h) pe = norm_to_pixel(coords[1][0], coords[1][1], w, h) # 兼容4.4格式: drag_start/end 可能是 {"polygon":[[x,y],...]} 需转成 {"polygons":[[[x,y],...]]} ds = gt.get("drag_start") or {} de = gt.get("drag_end") or {} if "polygon" in ds and "polygons" not in ds: ds = {"polygons": [ds["polygon"]], "keypoints": [], "ellipses": []} if "polygon" in de and "polygons" not in de: de = {"polygons": [de["polygon"]], "keypoints": [], "ellipses": []} s_ok = hit_region(ps[0], ps[1], ds) e_ok = hit_region(pe[0], pe[1], de) return True, (s_ok and e_ok) return True, False elif gt_action == "type_enter": tv = gt.get("type_value", "") return True, (tv in (type_text or "")) if tv else True else: # NO_COORD 或视图方向类 if "direction" in gt: return True, True # 方向需调用方单独判 return True, True def eval_step(reply, gt, w, h): """评测一步。支持 gt 带 alternatives（多等价方案，命中任一即对）。 返回 dict: thought, action, action_ok, param_ok, step_ok""" thought = parse_thought(reply) action = parse_action(reply) coords = parse_coords_list(reply) type_text = parse_type_value(reply) direction = parse_direction(reply) alts = gt["alternatives"] if "alternatives" in gt else [gt] best = (False, False) for alt in alts: a_ok, p_ok = eval_one_alternative(action, coords, type_text, alt, w, h) # 方向类单独判 if a_ok and "direction" in alt: p_ok = (direction == alt["direction"]) if a_ok and p_ok: best = (True, True); break if a_ok and not best[0]: best = (True, False) action_ok, param_ok = best return { "thought": thought, "action": action, "action_ok": action_ok, "param_ok": param_ok, "step_ok": action_ok and param_ok, } # ============================================================ # 6. LLM-as-Judge 意图一致性 # ============================================================ def judge_thought(client, gt_thought, pred_thought, high_instr, judge_model="qwen-plus"): """用 LLM 判断模型 thought 与 GT thought 语义是否一致。返回 True/False""" prompt = f"""你是评判员。在SolidWorks多步操作任务中，判断模型的"当前步骤意图"和标准答案意图是否在语义上一致（想做的事相同即可，措辞不同没关系）。 高层任务: {high_instr} 标准意图: {gt_thought} 模型意图: {pred_thought} 只回答一个词: 一致 或 不一致""" try: r = client.chat.completions.create( model=judge_model, messages=[{"role": "user", "content": prompt}], max_tokens=10, timeout=60) reply = (r.choices[0].message.content or "").strip().strip('"\'“”‘’。.!！ \t\r\n') return reply == "一致" except Exception: return False def attribute_verdict(action_ok, param_ok, intent_ok): """归因分类""" if action_ok and param_ok: return "完全正确" if action_ok and not param_ok and intent_ok: return "懂意图懂动作_点歪了" if action_ok and not param_ok and not intent_ok: return "动作蒙对_没真懂" if not action_ok and intent_ok: return "懂意图_动作选错" return "不懂该做什么" # ============================================================ # 9. 分组统计与能力分级 # ============================================================ def grade(rate): """能力分级：掌握>70% / 部分30-70% / 缺失<30%""" if rate >= 70: return "掌握" if rate >= 30: return "部分" return "缺失" def summarize_groups(results, groups): """按分组汇总定位结果。 results: run_localization 返回的逐题列表（每条含 id/hits/tries，可选 avg_iou/acc50） groups: [(组名, set(id集合)), ...] 返回每组的命中率、best-of-N、能力分级、（若有）平均IoU/Acc@0.5 """ by_id = {r["id"]: r for r in results} table = [] for gname, ids in groups: grp = [by_id[i] for i in ids if i in by_id] if not grp: continue total_hits = sum(r["hits"] for r in grp) total_tries = sum(r["tries"] for r in grp) best_of_n = sum(1 for r in grp if r["hits"] > 0) hit_rate = total_hits / total_tries * 100 if total_tries else 0 best_rate = best_of_n / len(grp) * 100 if grp else 0 row = { "group": gname, "n_items": len(grp), "hit_rate": round(hit_rate, 1), "best_of_n_rate": round(best_rate, 1), "grade": grade(hit_rate), } ious = [r["avg_iou"] for r in grp if "avg_iou" in r] if ious: row["avg_iou"] = round(sum(ious)/len(ious), 4) acc50 = [r["acc50"] for r in grp if "acc50" in r] if acc50: row["acc50"] = round(sum(acc50)/len(acc50)*100, 1) table.append(row) return table def make_groups_every_n(results, n=10): """便捷：每 n 个 id 一组，组名 Group1/Group2...""" ids = sorted(r["id"] for r in results) groups = [] for g, start in enumerate(range(0, len(ids), n), 1): chunk = set(ids[start:start+n]) groups.append((f"Group{g}", chunk)) return groups def make_groups_by_direction(results): """便捷：按 id 末位分'从左到右'(1-5)和'从右到左'(6-0)两组（任务3.1方位）""" l2r = {r["id"] for r in results if r["id"] % 10 in (1, 2, 3, 4, 5)} r2l = {r["id"] for r in results if r["id"] % 10 in (6, 7, 8, 9, 0)} return [("从左到右", l2r), ("从右到左", r2l)] def print_group_table(table, title=""): """打印分组汇总表""" if title: print(f"\n{'='*70}\n=== {title} ===\n{'='*70}") header = f"{'组别':<18}{'题数':>5}{'命中率':>9}{'Best-of-N':>11}{'分级':>7}" has_iou = any("avg_iou" in r for r in table) if has_iou: header += f"{'Acc@0.5':>9}{'平均IoU':>9}" print(header) print("-" * len(header)) for r in table: line = f"{r['group']:<18}{r['n_items']:>5}{r['hit_rate']:>8.1f}%{r['best_of_n_rate']:>10.1f}%{r['grade']:>7}" if has_iou and "avg_iou" in r: line += f"{r.get('acc50',0):>8.1f}%{r['avg_iou']:>9.4f}" print(line) # ============================================================ # 10. 模型坐标格式适配（不同模型坐标约定不同） # ============================================================ # 用法：根据模型名选择对应的解析+转换策略 MODEL_COORD_CONFIG = { # type: specialized=专用GUI模型(原生模板,不加CoT) / general=通用大模型(GUI expert prompt) # prompt: 用哪个模板 parse: 用哪个解析器 norm/order: 坐标转换 # point_only: 只能出点(任务1/2的IoU列留空) "default": {"type": "general", "prompt": "point", "parse": "bracket", "norm": 1000, "order": "xy"}, # ---- 专用 GUI 模型：各自原生模板 + 各自解析 ---- # UI-TARS-1.5: 输出 smart_resize 后绝对像素坐标(不是0-1000!)，需反算回原图 "ui-tars": {"type": "specialized", "prompt": "uitars", "parse": "uitars", "norm": "smart_resize", "resize": "ui-tars", "order": "xy", "point_only": True}, # OS-Atlas: 输出 [[x1,y1,x2,y2]]方括号 或 (x1,y1),(x2,y2)圆括号 0-1000，会复述指令文字需大token "os-atlas": {"type": "specialized", "prompt": "osatlas", "parse": "osatlas", "norm": 1000, "order": "xy", "max_tokens": 640}, "uground": {"type": "specialized", "prompt": "uground", "parse": "paren", "norm": 1000, "order": "xy", "point_only": True}, # OpenCUA: 官方system + pyautogui.click(x=,y=)，smart_resize绝对像素(默认参数)，point_only "opencua": {"type": "specialized", "prompt": "opencua", "parse": "pyautogui", "norm": "smart_resize", "resize": "opencua", "order": "xy", "system": "opencua", "point_only": True, "max_tokens": 512}, # UI-Venus: 官方 prompt 出 [x1,y1,x2,y2]，坐标在 smart_resize(min2M,max4.8M)像素空间，取中心点 "ui-venus": {"type": "specialized", "prompt": "uivenus", "parse": "uivenus", "norm": "smart_resize", "resize": "ui-venus", "order": "xy"}, "gui-owl": {"type": "specialized", "prompt": "point", "parse": "bracket", "norm": 1000, "order": "xy"}, # InfiGUI-G1: 像素坐标 + system prompt 告知分辨率 + + JSON point_2d（实测≈原图像素） "infigui": {"type": "specialized", "prompt": "infigui", "parse": "json_point", "norm": "smart_resize", "resize": "infigui", "order": "xy", "system": "infigui", "point_only": True}, # Holo2: 实测输出 0-1000 归一化(非smart_resize像素)，JSON click_absolute，话痨需大max_tokens "holo2": {"type": "specialized", "prompt": "holo2", "parse": "click_absolute", "norm": 1000, "order": "xy", "point_only": True, "max_tokens": 640}, # GELab-Zero(StepGUI): 0-999 归一化 "gelab": {"type": "specialized", "prompt": "point", "parse": "bracket", "norm": 999, "order": "xy"}, # ---- 通用大模型：统一 GUI expert prompt，thinking 模型自带 CoT ---- "qwen3-vl": {"type": "general", "prompt": "point", "parse": "bracket", "norm": 1000, "order": "xy"}, "glm-4.5v": {"type": "general", "prompt": "point", "parse": "bracket", "norm": 1000, "order": "xy"}, "glm-4.6v": {"type": "general", "prompt": "point", "parse": "bracket", "norm": 1000, "order": "xy", "max_tokens": 1024}, "glm-5v": {"type": "general", "prompt": "point", "parse": "bracket", "norm": 1000, "order": "xy", "max_tokens": 1024}, "gpt-5": {"type": "general", "prompt": "point", "parse": "bracket", "norm": 1000, "order": "xy", "max_tokens": 2048}, "claude": {"type": "general", "prompt": "point", "parse": "bracket", "norm": 1000, "order": "xy"}, "o3": {"type": "general", "prompt": "point", "parse": "bracket", "norm": 1000, "order": "xy"}, # Gemini: point 输出 xy(正常)，但 bbox 是 [ymin,xmin,ymax,xmax] → bbox_order=yx "gemini-2.5": {"type": "general", "prompt": "point", "parse": "bracket", "norm": 1000, "order": "xy", "bbox_order": "yx", "max_tokens": 8192}, "gemini": {"type": "general", "prompt": "point", "parse": "bracket", "norm": 1000, "order": "xy", "bbox_order": "yx", "max_tokens": 1024}, # Qwen2.5-VL: 输出像素坐标，需告知真实分辨率，转换时 norm=None "qwen2.5-vl": {"type": "general", "prompt": "point", "parse": "bracket", "norm": None, "order": "xy"}, } def get_model_cfg(model_name): """按模型名（子串匹配）取配置，找不到用 default。""" key = model_name.lower() for name, cfg in MODEL_COORD_CONFIG.items(): if name != "default" and name in key: return cfg return MODEL_COORD_CONFIG["default"] def normalize_coord(coord, cfg, w, h): """根据模型配置把模型输出的坐标转成原图像素 [x,y]。 coord: 模型输出的原始数值对; cfg: MODEL_COORD_CONFIG[xxx] norm 取值: - 数值(1000/999): 归一化基数, 坐标/norm*原图 - "smart_resize": 输出是 resize 后像素, 按 cfg['resize'] 参数反算回原图 - "pixel"/None: 已是原图像素, 直接用""" if cfg["order"] == "yx": coord = [coord[1], coord[0]] norm = cfg.get("norm") if isinstance(norm, (int, float)) and norm: return coord[0]/norm*w, coord[1]/norm*h if norm == "smart_resize": params = SMART_RESIZE_PARAMS.get(cfg.get("resize"), (3136, 12845056)) rh, rw = smart_resize(h, w, min_pixels=params[0], max_pixels=params[1]) return coord[0]/rw*w, coord[1]/rh*h return coord[0], coord[1] # 已是原图像素（norm == "pixel" / None） # ============================================================ # 12. 专用模型原生模板 + CoT 模板 # ============================================================ # UGround-V1：输出单个 "(x, y)" 0-1000（官方模板，末尾 Answer:） PROMPT_UGROUND = """Your task is to help the user identify the precise coordinates (x, y) of a specific area/element/object on the screen based on a description. - Your response should aim to point to the center or a representative point within the described area/element/object as accurately as possible. - If the description is unclear or ambiguous, infer the most relevant area or element based on its likely context or purpose. - Your answer should be a single string (x, y) corresponding to the point of the interest. Description: {instruction} Answer:""" # OS-Atlas：输出 "<|box_start|>(x1,y1),(x2,y2)<|box_end|>" 0-1000，可出框/点 PROMPT_OSATLAS_POINT = """In the screenshot of this page, please give me the coordinates of the element I want to click on according to my instructions(with point). "{instruction}\"""" # UI-TARS-1.5：grounding 模式，输出 click(start_box='(x,y)') 0-1000 PROMPT_UITARS = """Output only the coordinate of one point in your response. What element matches the following instruction: {instruction}""" # InfiGUI-G1：像素坐标 + system prompt 告知分辨率 + + JSON # system prompt 与 user prompt 在 call 时分别传入 INFIGUI_SYSTEM = ('You FIRST think about the reasoning process as an internal monologue ' 'and then provide the final answer.\n' 'The reasoning process MUST BE enclosed within tags.') PROMPT_INFIGUI = ("The screen's resolution is {w}x{h}.\n" 'Locate the UI element(s) for "{instruction}", ' 'output the coordinates using JSON format: [{{"point_2d": [x, y]}}, ...]') # Holo2：像素坐标(smart_resize) + JSON click_absolute PROMPT_HOLO2 = ('Localize an element on the GUI image according to the provided target ' 'and output a click position. You must output a valid JSON following the format: ' '{{"action": "click_absolute", "x": , "y": }}\n' 'Your target is: {instruction}') # UI-Venus：官方 grounding prompt，输出 [x1,y1,x2,y2]（smart_resize 像素空间） PROMPT_UIVENUS = ('Outline the position corresponding to the instruction: {instruction}. ' 'The output should be only [x1,y1,x2,y2].') # OpenCUA：官方 system + 直接给 instruction，输出 pyautogui.click(x=,y=)，smart_resize 绝对像素 OPENCUA_SYSTEM = ('You are a GUI agent. You are given a task and a screenshot of the screen. ' 'You need to perform a series of pyautogui actions to complete the task.') PROMPT_OPENCUA = "{instruction}" # 任务1.3/2.3 CoT 对照（通用大模型） PROMPT_POINT_COT = """You are a GUI expert. Analyze step by step. """ + COORD_SYSTEM + """ Action Intent: {instruction} Output: Observation: Reasoning: Point: [x, y]""" # ============================================================ # 11. 模型调用（OpenAI 兼容 API；本地部署模型用 vLLM/sglang 起 OpenAI server 即可） # ============================================================ import base64 import os import time def encode_image(path): with open(path, "rb") as f: return base64.b64encode(f.read()).decode("utf-8") def call_vlm(client, model, image_path, prompt, max_tokens=300, retries=3, temperature=0.0, system=None): """调用多模态模型。client 为 openai.OpenAI 实例。 temperature 默认 0.0：任务1/2 每题仅 1 次，必须确定性输出才可复现 （UGround 等官方亦强调 temperature=0）。 system: 可选 system prompt（InfiGUI-G1 等需要）。 """ b64 = encode_image(image_path) messages = [] if system: messages.append({"role": "system", "content": system}) messages.append({"role": "user", "content": [ {"type": "image_url", "image_url": {"url": f"data:image/png;base64,{b64}"}}, {"type": "text", "text": prompt} ]}) # GPT-5 / o3 等推理模型：用 max_completion_tokens，且只支持默认 temperature is_reasoning = any(k in model.lower() for k in ("gpt-5", "o3", "o4")) for a in range(retries): try: kwargs = dict(model=model, messages=messages, timeout=120) if is_reasoning: kwargs["max_completion_tokens"] = max_tokens else: kwargs["max_tokens"] = max_tokens kwargs["temperature"] = temperature resp = client.chat.completions.create(**kwargs) return resp.choices[0].message.content or "" except Exception as e: if a < retries - 1: time.sleep(3) else: return f"API_ERROR: {e}" SYSTEM_PROMPTS = {"infigui": INFIGUI_SYSTEM, "opencua": OPENCUA_SYSTEM} # ============================================================ # 8. 任务级评测函数 # ============================================================ def build_localization_prompt(prompt_style, instruction, mode, cot=False, w=None, h=None): """按 prompt_style 选模板生成定位 prompt。 prompt_style: point/uground/osatlas/uitars/infigui/holo2; mode: point/bbox; cot: 意图题用 w/h: 像素模型(infigui)需要把分辨率写进 prompt""" if cot: return PROMPT_POINT_COT.format(instruction=instruction) if prompt_style == "uground": return PROMPT_UGROUND.format(instruction=instruction) if prompt_style == "osatlas": return PROMPT_OSATLAS_POINT.format(instruction=instruction) if prompt_style == "uitars": return PROMPT_UITARS.format(instruction=instruction) if prompt_style == "infigui": return PROMPT_INFIGUI.format(instruction=instruction, w=w, h=h) if prompt_style == "holo2": return PROMPT_HOLO2.format(instruction=instruction) if prompt_style == "uivenus": return PROMPT_UIVENUS.format(instruction=instruction) if prompt_style == "opencua": return PROMPT_OPENCUA.format(instruction=instruction) # 通用大模型 / 默认 if mode == "bbox": return PROMPT_BBOX.format(instruction=instruction) return PROMPT_POINT.format(instruction=instruction) def run_localization(client, model, benchmark_file, image_dir, tries=None, mode="point", cfg=None, cot=False, max_tokens=None): """定位任务（任务1/2/3/4.1）。 - cfg: MODEL_COORD_CONFIG 条目；None 时按 model 名自动取（含 prompt/parse/norm/order/point_only） - tries: 每题测试次数；None 时按数据规模自动（任务1/2 大数据集=1次，任务3/4=10次） - mode: 'point' 或 'bbox'；point_only 模型强制 point - cot: 意图题(1.3/2.3) CoT 对照实验时置 True benchmark 既支持 bbox_pixels（任务1/2）也支持 ground_truth（任务3/4）格式。""" cfg = cfg or get_model_cfg(model) data = json.load(open(benchmark_file, "r", encoding="utf-8")) # point_only 模型不能出框：bbox 模式自动降级为 point（IoU 留空） if cfg.get("point_only") and mode == "bbox": mode = "point" # tries 默认：bbox_pixels 格式(任务1/2)每题1次，ground_truth 格式(任务3/4)每题10次 if tries is None: tries = 1 if ("bbox_pixels" in data[0]) else 10 # CoT 模式需要更长输出；专用话痨模型(如Holo2)可在cfg里指定 max_tokens if max_tokens is None: max_tokens = cfg.get("max_tokens", 512 if cot else 100) prompt_style = cfg.get("prompt", "point") parse_style = cfg.get("parse", "bracket") results = [] consec_quota_err = 0 # 连续配额错误计数，达阈值则中止（避免余额耗尽后空跑） for item in data: img = os.path.join(image_dir, item["image_path"]) w, h = item["image_resolution"] if "ground_truth" in item: gt = item["ground_truth"] gt_box_norm = None else: # bbox_pixels 格式 bp = item["bbox_pixels"] gt = {"include": {"polygons": [[[bp[0],bp[1]],[bp[2],bp[1]],[bp[2],bp[3]],[bp[0],bp[3]]]], "keypoints": [], "ellipses": []}, "exclude": {"polygons": [], "keypoints": [], "ellipses": []}} gt_box_norm = pixel_bbox_to_norm(bp, item["image_resolution"]) prompt = build_localization_prompt(prompt_style, item["instruction"], mode, cot, w, h) sys_prompt = SYSTEM_PROMPTS.get(cfg.get("system")) hits = 0 ious = [] parse_fail = 0 trials = [] for _ in range(tries): started_at = time.perf_counter() reply = call_vlm(client, model, img, prompt, max_tokens, system=sys_prompt) latency_s = round(time.perf_counter() - started_at, 4) # 配额耗尽保护：连续 20 次 429/quota 错误则中止整轮，避免空跑污染结果 if isinstance(reply, str) and reply.startswith("API_ERROR") and \ ("429" in reply or "quota" in reply.lower() or "insufficient" in reply.lower()): consec_quota_err += 1 if consec_quota_err >= 20: raise RuntimeError( f"[配额耗尽中止] 连续{consec_quota_err}次429/quota错误，" f"已完成的结果已落盘，请充值后重跑(断点续跑会跳过已完成)。") else: consec_quota_err = 0 if mode == "point": pt = parse_point_by_style(reply, parse_style) if pt is None: parse_fail += 1 trials.append({ "parse_ok": False, "hit": False, "latency_s": latency_s, }) continue px, py = normalize_coord(pt, cfg, w, h) hit = evaluate_prediction(px, py, gt) if hit: hits += 1 trials.append({ "parse_ok": True, "raw_point": [round(float(pt[0]), 4), round(float(pt[1]), 4)], "point_px": [round(float(px), 4), round(float(py), 4)], "hit": bool(hit), "latency_s": latency_s, }) else: # bbox bb = parse_bbox_by_style(reply, parse_style) if bb is None: parse_fail += 1 trials.append({ "parse_ok": False, "latency_s": latency_s, }) continue if gt_box_norm: # 把模型 bbox 的两个角点经 normalize_coord 转成原图像素， # 再转 0-1000 归一化（统一处理 yx 顺序 / smart_resize / 归一化基数） # bbox 顺序可能与 point 不同(如 Gemini point=xy 但 bbox=yx)，用 bbox_order 覆盖 bcfg = dict(cfg, order=cfg.get("bbox_order", cfg["order"])) p1 = normalize_coord([bb[0], bb[1]], bcfg, w, h) p2 = normalize_coord([bb[2], bb[3]], bcfg, w, h) bb_norm = [p1[0]/w*1000, p1[1]/h*1000, p2[0]/w*1000, p2[1]/h*1000] bb_norm = [min(bb_norm[0], bb_norm[2]), min(bb_norm[1], bb_norm[3]), max(bb_norm[0], bb_norm[2]), max(bb_norm[1], bb_norm[3])] iou = compute_iou(bb_norm, gt_box_norm) ious.append(iou) trials.append({ "parse_ok": True, "raw_bbox": [round(float(value), 4) for value in bb], "bbox_norm": [round(float(value), 4) for value in bb_norm], "iou": round(float(iou), 6), "latency_s": latency_s, }) else: trials.append({ "parse_ok": True, "raw_bbox": [round(float(value), 4) for value in bb], "latency_s": latency_s, }) rec = { "id": item["id"], "hits": hits, "tries": tries, "parse_fail": parse_fail, "trials": trials, } if mode == "bbox" and ious: rec["avg_iou"] = sum(ious)/len(ious) rec["acc50"] = sum(1 for v in ious if v >= 0.5)/len(ious) rec["acc75"] = sum(1 for v in ious if v >= 0.75)/len(ious) results.append(rec) return results def run_multistep(client, model, benchmark_file, image_base, judge_model="qwen-plus", judge_client=None): """多步规划评测（任务5），Teacher Forcing + 三层归因。 image_base: 多步截图目录；支持数据中保存的相对路径或 record/raw_images 布局。""" judge_client = judge_client or client data = json.load(open(benchmark_file, "r", encoding="utf-8")) task_results = [] for task in data: tid = task["task_id"] instr = task["high_level_instruction"] history = [] step_results = [] died_at = None for s in task["steps"]: rel_image = s["image_path"].replace("\\", "/") img = os.path.join(image_base, *rel_image.split("/")) if not os.path.exists(img): img = os.path.join(image_base, f"record_{tid:03d}", "raw_images", os.path.basename(rel_image)) if not os.path.exists(img): img = os.path.join(image_base, *[part for part in rel_image.split("/") if part != "raw_images"]) w, h = s["image_resolution"] gt = s["ground_truth"] hist_txt = "\n".join(history) if history else "(none, this is the first step)" prompt = PROMPT_MULTISTEP.format(high_level_instruction=instr, history=hist_txt) reply = call_vlm(client, model, img, prompt, 400) ev = eval_step(reply, gt, w, h) if ev["step_ok"]: verdict = "完全正确"; intent_ok = None elif "thought" in s: intent_ok = judge_thought(judge_client, s["thought"], ev["thought"], instr, judge_model) verdict = attribute_verdict(ev["action_ok"], ev["param_ok"], intent_ok) else: intent_ok = None verdict = "动作或参数错误" if not ev["step_ok"] and died_at is None: died_at = s["step"] step_results.append({"step": s["step"], "verdict": verdict, **ev, "intent_ok": intent_ok}) history.append(f"Step {s['step']}: [action]{gt['action_type']}") correct = sum(1 for r in step_results if r["step_ok"]) task_results.append({ "task_id": tid, "total_steps": len(task["steps"]), "correct_steps": correct, "task_success": correct == len(task["steps"]), "died_at_step": died_at if died_at else "全对", "steps": step_results, }) return task_results # ============================================================ # 使用示例 # ============================================================ if __name__ == "__main__": from openai import OpenAI # 本地部署模型：用 vLLM / sglang 起一个 OpenAI 兼容 server # 例: python -m vllm.entrypoints.openai.api_server --model --port 8000 client = OpenAI(api_key="EMPTY", base_url="http://localhost:8000/v1") MODEL = "your-model-name" # 定位任务 res = run_localization(client, MODEL, "任务3.1基础任务.json", "任务3截图", tries=10, mode="point") total_hits = sum(r["hits"] for r in res) total_tries = sum(r["tries"] for r in res) print(f"命中率: {total_hits}/{total_tries} = {total_hits/total_tries*100:.1f}%") # 多步规划 # task_res = run_multistep(client, MODEL, "任务5_多步规划.json", "dataset_records") # for t in task_res: # print(f"Task{t['task_id']}: {'成功' if t['task_success'] else '失败'} " # f"{t['correct_steps']}/{t['total_steps']} 死在第{t['died_at_step']}步")