Upload 2 files

metric/gpt_judge.py

@@ -51,12 +51,6 @@ def save_json_file(
             raise RuntimeError(f"save json failed: {e}") from e
 
 def read_json_file(file_path):
-    """
-    Reads a JSON file and returns the parsed data as a Python object.
-
-    :param file_path: The path to the JSON file
-    :return: The data parsed from the JSON file
-    """
     with open(file_path, 'r', encoding='utf-8') as f:
         data = json.load(f)
     return data
@@ -67,7 +61,6 @@ def openai_api( prompt = None):
 
     client = OpenAI(
         base_url='your_url',
-
         api_key='your_key'
     )
     response = client.chat.completions.create(
@@ -86,25 +79,6 @@ def openai_api( prompt = None):
     return response
 
 
-def extract_last_bracket_list(s: str) -> list:
-    """
-    """
-    last_open = s.rfind('[')
-    last_close = s.rfind(']')
-    content = s[last_open:last_close+1]
-
-    try:
-
-        result = json.loads(content)
-    except:
-        content = re.sub(r'[\n\t\r]', '', content)
-        content = ' '.join(content.split())
-        result = []
-        content = content.split("\",")
-        for item in content:
-            for item2 in item.split("\',"):
-                result.append(item2)
-    return result
 
 
 def ads_task_llm_judge(jsonlist):
@@ -114,6 +88,7 @@ def ads_task_llm_judge(jsonlist):
         retry_wait = 10
         tasks = read_json_file(json_file)
         for item in tqdm(tasks):
+            if item["task"]!="advertisement reasoning": continue
             if "judge" in item: continue
 
             prompt_templet = 'Please help me determine if the content in the Description contains Key Information. If it does, answer directly with "Yes"; if it does not, answer directly with "No". Please respond only with "Yes" or "No", without any additional output.'
@@ -158,6 +133,13 @@ def ads_task_llm_judge(jsonlist):
 
 if __name__ == "__main__":
     jsonlist = [
-        r"your_path/gpt4o_bench.json",
+        # r"C:\Users\11978\Desktop\poster\metric\gpt-4o-mini_bench.json",
+        # r"C:\Users\11978\Desktop\poster\metric\gpt-4.5-preview-2025-02-27.json",
+        # r"C:\Users\11978\Desktop\poster\metric\gemini-2.0-pro-exp-02-05_bench.json",
+        # r"C:\Users\11978\Desktop\poster\metric\gemini-2.5-pro-preview-03-25_bench.json",
+        # r"C:\Users\11978\Desktop\poster\metric\doubao-1-5-vision-pro-250328_bench.json",
+        # r"C:\Users\11978\Desktop\poster\metric\gemma-3-27b-it.json",
+        # r"C:\Users\11978\Desktop\poster\metric\gemma-3-12b-it.json",
+        # r"C:\Users\11978\Desktop\poster\metric\gemma-3-4b-it.json"
     ]
     ads_task_llm_judge(jsonlist)

metric/metric.py

@@ -243,11 +243,24 @@ def extract_numbers_float2(s):
     return numbers
 def group_numbers_into_fours(num_list):
     """
+    将数字列表按每4个元素分组，并验证总长度是否为4的倍数
 
+    参数:
+    num_list -- 全数字组成的列表，如 [1,2,3,4,5,6,7,8]
+
+    返回:
+    分组后的二维列表，如 [[1,2,3,4], [5,6,7,8]]
+
+    异常:
+    ValueError -- 当输入列表长度不是4的倍数时抛出
     """
     n = len(num_list)
 
+    # 验证长度是否为4的倍数
+    # if n % 4 != 0:
+    #     raise ValueError(f"数字个数 {n} 不是4的倍数，无法完整分组")
 
+    # 按步长4切割列表
     result = [num_list[i:i + 4] for i in range(0, n-3, 4)]
     return result
 
@@ -274,7 +287,7 @@ def parse_bbox_string2(bbox_str):
     """
     """
     try:
-
+        # 使用literal_eval将字符串解析为Python对象
         bbox_str = clean_string_for_box(bbox_str)
         bbox_nums = extract_numbers_float2(bbox_str)
         bboxes = group_numbers_into_fours(bbox_nums)
@@ -288,7 +301,9 @@ def calculate_iou(box1, box2):
     """
     """
 
-
+    # 解析坐标
+    # print("box 1",box1)
+    # print("box 2",box2)
     x1_1, y1_1, x2_1, y2_1 = box1
     x1_2, y1_2, x2_2, y2_2 = box2
 
@@ -304,21 +319,21 @@ def calculate_iou(box1, box2):
     if x1_2 > x2_2: return 0.0
     if y1_2 > y2_2: return 0.0
 
-
+    # 处理无交集情况
     if x_right < x_left or y_bottom < y_top:
         return 0.0
 
-
+    # 计算交集面积
     intersection_area = (x_right - x_left) * (y_bottom - y_top)
 
-
+    # 计算各自面积
     box1_area = (x2_1 - x1_1) * (y2_1 - y1_1)
     box2_area = (x2_2 - x1_2) * (y2_2 - y1_2)
 
-
+    # 计算并集面积
     union_area = box1_area + box2_area - intersection_area
 
-
+    # 计算IoU
     iou = intersection_area / union_area
     return iou
 
@@ -342,7 +357,9 @@ def calculate_area_ratio(box1, box2):
     """
     """
 
-
+    # 解析坐标
+    # print("box 1",box1)
+    # print("box 2",box2)
     x1_1, y1_1, x2_1, y2_1 = box1
     x1_2, y1_2, x2_2, y2_2 = box2
 
@@ -378,7 +395,10 @@ def norm_bbox(norm_bbox,size):
 
 def bbox_number_types(bboxes):
     """
+    判断一组 bbox 列表中的每个数字是整数还是小数。
 
+    :param bboxes: List[List[float]]，每个 bbox 为 [x1, y1, x2, y2]
+    :return: List[List[str]]，与 bboxes 结构相同，每个位置返回 "int" 或 "float"
     """
     result = []
     for box in bboxes:
@@ -431,11 +451,70 @@ def task_4_ocr(data):
             #     # print(response)
             #     continue
             colorful_words_ocr.append(word_level_ac(gt, response))
+    pc_wr1 = sum(pure_char_ocr)/len(pure_char_ocr)
+    cc_wr1 = sum(colorful_char_ocr)/len(colorful_char_ocr)
+    pw_wr1 = sum(pure_words_ocr) / len(pure_words_ocr)
+    cw_wr1 = sum(colorful_words_ocr) / len(colorful_words_ocr)
+    # print("pure_char_ocr      word-level accuracy: ", f"{sum(pure_char_ocr)/len(pure_char_ocr):.3f}", f" total imgs:{len(pure_char_ocr)/400}")
+    # print("colorful_char_ocr  word-level accuracy: ", f"{sum(colorful_char_ocr)/len(colorful_char_ocr):.3f}", f" total imgs:{len(colorful_char_ocr)/400}")
+    # print("pure_words_ocr     word-level accuracy: ", f"{sum(pure_words_ocr) / len(pure_words_ocr):.3f}", f" total imgs:{len(pure_words_ocr)/400}")
+    # print("colorful_words_ocr word-level accuracy: ", f"{sum(colorful_words_ocr) / len(colorful_words_ocr):.3f}", f" total imgs:{len(colorful_words_ocr)/400}")
+
+    """  character ocr and words ocr  """
+    pure_char_ocr = []
+    colorful_char_ocr = []
+    pure_words_ocr = []
+    colorful_words_ocr = []
+    for item in data:
+        if item["task"] == "pure_char ocr":
+            gt = item["gt"]
+            response = item["response"]
+            if word_level_ac(gt, response) < 0.1:
+            #     # print(response)
+                continue
+            pure_char_ocr.append(word_level_ac(gt, response))
 
-    print("pure_char_ocr      word-level accuracy: ", f"{sum(pure_char_ocr)/len(pure_char_ocr):.3f}", f" total imgs:{len(pure_char_ocr)/400}")
-    print("colorful_char_ocr  word-level accuracy: ", f"{sum(colorful_char_ocr)/len(colorful_char_ocr):.3f}", f" total imgs:{len(colorful_char_ocr)/400}")
-    print("pure_words_ocr     word-level accuracy: ", f"{sum(pure_words_ocr) / len(pure_words_ocr):.3f}", f" total imgs:{len(pure_words_ocr)/400}")
-    print("colorful_words_ocr word-level accuracy: ", f"{sum(colorful_words_ocr) / len(colorful_words_ocr):.3f}", f" total imgs:{len(colorful_words_ocr)/400}")
+        if item["task"] == "colorful_char ocr":
+            gt = item["gt"]
+            response = item["response"]
+            if word_level_ac(gt, response) < 0.1:
+            #     # print(response)
+                continue
+            colorful_char_ocr.append(word_level_ac(gt, response))
+
+        if item["task"] == "pure_words ocr":
+            gt = item["gt"]
+            response = item["response"]
+            if word_level_ac(gt, response) < 0.1:
+            #     # print(response)
+                continue
+            pure_words_ocr.append(word_level_ac(gt, response))
+        if item["task"] == "colorful_words ocr":
+            gt = item["gt"]
+            response = item["response"]
+            if word_level_ac(gt, response)<0.1:
+            #     # print(response)
+                continue
+            colorful_words_ocr.append(word_level_ac(gt, response))
+    pc_wr2 = sum(pure_char_ocr) / len(pure_char_ocr)
+    cc_wr2 = sum(colorful_char_ocr) / len(colorful_char_ocr)
+    pw_wr2 = sum(pure_words_ocr) / len(pure_words_ocr)
+    cw_wr2 = sum(colorful_words_ocr) / len(colorful_words_ocr)
+    pc_r = len(pure_char_ocr) / 400
+    cc_r = len(colorful_char_ocr) / 400
+    pw_r = len(pure_words_ocr) / 400
+    cW_r = len(colorful_words_ocr) / 400
+    # print("pure_char_ocr      word-level accuracy: ", f"{sum(pure_char_ocr) / len(pure_char_ocr):.3f}",
+    #       f" total imgs:{len(pure_char_ocr) / 400}")
+    # print("colorful_char_ocr  word-level accuracy: ", f"{sum(colorful_char_ocr) / len(colorful_char_ocr):.3f}",
+    #       f" total imgs:{len(colorful_char_ocr) / 400}")
+    # print("pure_words_ocr     word-level accuracy: ", f"{sum(pure_words_ocr) / len(pure_words_ocr):.3f}",
+    #       f" total imgs:{len(pure_words_ocr) / 400}")
+    # print("colorful_words_ocr word-level accuracy: ", f"{sum(colorful_words_ocr) / len(colorful_words_ocr):.3f}",
+    #       f" total imgs:{len(colorful_words_ocr) / 400}")
+
+
+    return pc_wr1, pc_wr2, pc_r,   cc_wr1, cc_wr2, cc_r,   pw_wr1, pw_wr2, pw_r,  cw_wr1, cw_wr2, cW_r
 
 def task_font_size(data):
     """  font size robustness  """
@@ -492,8 +571,9 @@ def task_font_size(data):
     mean_r = statistics.mean(recall_num)
 
     # print(f"Mean: {mean:.3f} Std: {std:.3f} Mean: {mean3:.3f} Std: {std3:.3f} reacall_num: {mean_r}")
-    print(f"Mean: {mean:.3f} Std: {std:.3f} Mean: {mean3:.3f} Std: {std3:.3f} reacall_num: {mean_r:.3f}")
+    # print(f"Mean: {mean:.3f} Std: {std:.3f} Mean: {mean3:.3f} Std: {std3:.3f} reacall_num: {mean_r:.3f}")
 
+    return mean, std, mean3, std3, mean_r/100
 def task_logo_cor(data):
     """  logo ocr  """
     result = []
@@ -502,7 +582,8 @@ def task_logo_cor(data):
             gt = item["gt"]
             response = item["response"]
             result.append(logo_ocr_ac(gt, response))
-    print(f"logo  ocr  accuracy: {sum(result)/len(result):.3f}  total imgs: {len(result)}")
+    # print(f"logo  ocr  accuracy: {sum(result)/len(result):.3f}  total imgs: {len(result)}")
+    return sum(result)/len(result)
 
 def task_poster_ocr(data):
     """  real poster ocr  """
@@ -517,7 +598,8 @@ def task_poster_ocr(data):
             ac = real_poster_ac(gt, response)
             if ac<0.05: continue
             result.append(ac)
-    print(f"poster ocr accuracy (entity-level): {sum(result)/len(result):.3f}  total imgs: {len(result)}")
+    # print(f"poster ocr accuracy (entity-level): {sum(result)/len(result):.3f}  total imgs: {len(result)}")
+    return sum(result)/len(result)
 
 def task_font_matching_1(data):
     """  font matching 1  """
@@ -533,7 +615,7 @@ def task_font_matching_1(data):
                 continue
             # print(response)
             result.append(font_matching_ac(gt, response))
-    print(f"font matching 1 accuracy: {sum(result) / len(result):5f}  total imgs: {len(result)}")
+    # print(f"font matching 1 accuracy: {sum(result) / len(result):5f}  total imgs: {len(result)}")
     return sum(result) / len(result)
 
 def task_font_matching_2(data):
@@ -550,7 +632,7 @@ def task_font_matching_2(data):
                 continue
             # print(response)
             result.append(font_matching_ac(gt, response))
-    print(f"font matching 2 accuracy: {sum(result) / len(result):5f}  total imgs: {len(result)}")
+    # print(f"font matching 2 accuracy: {sum(result) / len(result):5f}  total imgs: {len(result)}")
     return sum(result) / len(result)
 
 def task_font_matching(data):
@@ -584,7 +666,7 @@ def task_font_attr(data):
                 continue
             # print(response)
             result.append(font_attr_ac(gt, response))
-    print(f"font attributes accuracy: {sum(result) / len(result):5f}  total imgs: {len(result)}")
+    # print(f"font attributes accuracy: {sum(result) / len(result):5f}  total imgs: {len(result)}")
 
     font_attr_list = []
     font_attr_dic = {}
@@ -622,7 +704,7 @@ def task_font_effect(data):
             response = item["response"]
             # print(response)
             result.append(font_effect_ac(gt, response))
-    print(f"font effect accuracy: {sum(result) / len(result):.5f}  total imgs: {len(result)}")
+    # print(f"font effect accuracy: {sum(result) / len(result):.5f}  total imgs: {len(result)}")
 
     font_effect_list = []
     font_effect_dic = {}
@@ -664,8 +746,8 @@ def task_font_effect_2(data):
             if effect_ac != None:
                 result_e.append(effect_ac)
 
-    print(f"font effect 2  color accuracy: {sum(result_c) / len(result_c):5f}  total imgs: {len(result_c)}")
-    print(f"font effect 2 effect accuracy: {sum(result_e) / len(result_e):5f}  total imgs: {len(result_e)}")
+    # print(f"font effect 2  color accuracy: {sum(result_c) / len(result_c):5f}  total imgs: {len(result_c)}")
+    # print(f"font effect 2 effect accuracy: {sum(result_e) / len(result_e):5f}  total imgs: {len(result_e)}")
 
     return sum(result_c) / len(result_c), sum(result_e) / len(result_e)
 
@@ -681,7 +763,7 @@ def task_layout_comprison(data):
             response = item["response"]
             # print(response)
             result.append(font_effect_ac(gt, response))
-    print(f"layout disorder comparison accuracy: {sum(result) / len(result):5f}  total imgs: {len(result)}")
+    # print(f"layout disorder comparison accuracy: {sum(result) / len(result):5f}  total imgs: {len(result)}")
     return sum(result) / len(result)
 
 
@@ -745,10 +827,10 @@ def task_align_rotate(data):
     r_result.extend(r3_result)
 
 
-    print(f"alignment accuracy: {sum(a_result) / len(a_result):5f}  total imgs: {len(a_result)}")
-    print(f"rotation  accuracy: {sum(r1_result) / len(r1_result):5f}  total imgs: {len(r1_result)}")
-    print(f"rotation  accuracy: {sum(r2_result) / len(r2_result):5f}  total imgs: {len(r2_result)}")
-    print(f"rotation  accuracy: {sum(r3_result) / len(r3_result):5f}  total imgs: {len(r3_result)}")
+    # print(f"alignment accuracy: {sum(a_result) / len(a_result):5f}  total imgs: {len(a_result)}")
+    # print(f"rotation  accuracy: {sum(r1_result) / len(r1_result):5f}  total imgs: {len(r1_result)}")
+    # print(f"rotation  accuracy: {sum(r2_result) / len(r2_result):5f}  total imgs: {len(r2_result)}")
+    # print(f"rotation  accuracy: {sum(r3_result) / len(r3_result):5f}  total imgs: {len(r3_result)}")
 
     return sum(a_result) / len(a_result), sum(r_result) / len(r_result)
 
@@ -1083,7 +1165,7 @@ def task_ads(data):
         if item["task"]=="advertisement reasoning":
             item_point_list = []
             if "judge" in item:
-                
+
                 for content in item["judge"]:
                     if "Yes" in content:
                         points += 1
@@ -1094,37 +1176,80 @@ def task_ads(data):
             score_list.append(score)
 
         # print(f"{sum(score_list)/len(score_list):.3f} {points} ")
-        print(f"{sum(score_list) / len(score_list):.3f}")
+        # print(f"{sum(score_list) / len(score_list):.3f}")
+    return sum(score_list) / len(score_list)
 
 
 if __name__=="__main__":
-
-    # # data = read_json_file(r"Llama-3.2-11B-Vision-Instruct_bench.json")
-    # task_4_ocr(data)
-    # task_logo_cor(data)
-    # task_poster_ocr(data)
-    # task_font_matching_1(data)
-    # task_font_matching_2(data)
-    # task_font_attr(data)
-    # task_font_effect(data)
-    # task_font_effect_2(data)
-    # task_font_size(data)
-    # task_layout_comprison(data)
-    # task_poster_detection(data)
-    # task_layout_generation(data)
-    # task_align_rotate(data)
-    # task_empty_space(data)
     jsonlist = [
+        r"C:\Users\11978\Desktop\poster\metric\gpt4o_bench.json",
+        r"C:\Users\11978\Desktop\poster\metric\gemini_bench.json",
+        r"C:\Users\11978\Desktop\poster\metric\claude_bench.json",
+        r"C:\Users\11978\Desktop\poster\metric\tasks_intervl_8b.json",
+        r"C:\Users\11978\Desktop\poster\metric\Kimi-VL-A3B-Instruct_bench.json",
+        r"C:\Users\11978\Desktop\poster\metric\Phi-3.5-vision-instruct_bench.json",
+        r"C:\Users\11978\Desktop\poster\metric\Llama-3.2-11B-Vision-Instruct_bench(1).json",
+        r"C:\Users\11978\Desktop\poster\metric\qwen_32b_bench.json",
+        r"C:\Users\11978\Desktop\poster\metric\tasks_qwen_7b.json",
+        r"C:\Users\11978\Desktop\poster\metric\tasks_qwen_3b.json",
+        r"C:\Users\11978\Desktop\poster\metric\gpt-4o-mini_bench.json",
+        r"C:\Users\11978\Desktop\poster\metric\gpt-4.5-preview-2025-02-27.json",
+        r"C:\Users\11978\Desktop\poster\metric\gemini-2.0-pro-exp-02-05_bench.json",
+        r"C:\Users\11978\Desktop\poster\metric\gemini-2.5-pro-preview-03-25_bench.json",
+        r"C:\Users\11978\Desktop\poster\metric\doubao-1-5-vision-pro-250328_bench.json",
+        r"C:\Users\11978\Desktop\poster\metric\gemma-3-27b-it.json",
+        r"C:\Users\11978\Desktop\poster\metric\gemma-3-12b-it.json",
+        r"C:\Users\11978\Desktop\poster\metric\gemma-3-4b-it.json"
     ]
+
     for json_item in jsonlist:
         print(os.path.basename(json_item))
         data = read_json_file(json_item)
-        top1_iou, _ = task_poster_detection(data, max_box_num=1)
-        top3_iou, _ = task_poster_detection(data, max_box_num=3)
-        top5_iou, _ = task_poster_detection(data, max_box_num=5)
-        mean_iou, recall = task_poster_detection(data, max_box_num=30)
-        print(f"{top1_iou:.3f} & {top3_iou:.3f} & {top5_iou:.3f} & {mean_iou:.3f} & {recall:.3f}")
-
+        """ocr"""
+        # logo_ac = task_logo_cor(data)
+        # poster_ac = task_poster_ocr(data)
+        # pc_wr1, pc_wr2, pc_r,   cc_wr1, cc_wr2, cc_r,   pw_wr1, pw_wr2, pw_r,  cw_wr1, cw_wr2, cW_r = task_4_ocr(data)
+        # print(f"& {logo_ac:.3f} & {poster_ac:.3f} & {pc_wr1:.3f} & {pc_wr2:.3f} & {pc_r:.3f} & {cc_wr1:.3f} & {cc_wr2:.3f} & {cc_r:.3f} & {pw_wr1:.3f} & {pw_wr2:.3f} & {pw_r:.3f} & {cw_wr1:.3f} & {cw_wr2:.3f} & {cW_r:.3f}")
+        """font size ocr"""
+        # mean, std, mean3, std3, mean_r = task_font_size(data)
+        # print(f"& {mean:.3f} & {std:.3f} & {mean3:.3f} & {std3:.3f} & {mean_r:.3f}")
+        """font task"""
+        # fm1 = task_font_matching_1(data)
+        # fm2 = task_font_matching_2(data)
+        # fm = (fm1 + fm2) /2
+        # fm_score = k_option_norm(fm, k=9)
+        #
+        # fattr = task_font_attr(data)
+        # fattr_score = k_option_norm(fattr, k=2)
+        #
+        # fe1 = task_font_effect(data)
+        # fc,fe2 = task_font_effect_2(data)
+        # fe1_score = k_option_norm(fe1,k=9)
+        # fc_score, fe2_score = k_option_norm(fc,k=16), k_option_norm(fe2, k=48)
         #
+        # print(f"& {fm_score:.3f} & {fattr_score:.3f} & {fe1_score:.3f} & {fc_score:.3f} & {fe2_score:.3f}")
+
+        """text localization"""
+        # top1_iou, _ = task_poster_detection(data, max_box_num=1)
+        # top3_iou, _ = task_poster_detection(data, max_box_num=3)
+        # top5_iou, _ = task_poster_detection(data, max_box_num=5)
+        # mean_iou, recall = task_poster_detection(data, max_box_num=30)
+        # # print(f"{top1_iou:.3f} & {top3_iou:.3f} & {top5_iou:.3f} & {mean_iou:.3f} & {recall:.3f}")
+        # print(f"{top1_iou:.3f} & {top3_iou:.3f} & {mean_iou:.3f} & {recall:.3f}")
+        """text positioning"""
+        # a, r = task_align_rotate(data)
+        # a, r = k_option_norm(a, k=3), k_option_norm(r, k=3)
+        # print(f"{a:.3f} & {r:.3f}")
+        """empty space"""
+        # iou, match = task_empty_space(data)
+        # print(f"{iou:.3f} & {match:.3f}")
+        """layout comparison"""
+        # vs = task_layout_comprison(data)
+        # vs_score = k_option_norm(vs, k=2)
+        # print(f"& {vs_score:.3f} & ")
+        """layout generation"""
         # bias, area_rate, rate = task_layout_generation(data)
         # print(f"{bias:.3f} & {area_rate:.3f} & {rate:.3f}")
+        """advertisement understanding"""
+        points = task_ads(data)
+        print(f"& {points:.3f}")