human_corr_vqa_entity.py

import csv
import numpy as np
from scipy.stats import spearmanr, kendalltau

def add_to_csv(input_files,human_score_csv):
    output_rows = []
    for idx in [0,10, 20, 30, 40, 50, 60, 70, 80, 90, 100, 110, 120, 130, 140, 150, 160, 170, 180, 190]:
        for filename in input_files:
            with open(filename, newline='', encoding='utf-8') as f:
                reader = csv.reader(f)
                header = next(reader)  # 跳过表头
                rows = list(reader)
                if idx < len(rows):
                    for row in rows:
                        if row[0] == f"{idx+1:04d}.png":
                            target_row = row
                            break
                    # row = rows[idx]
                    
                    # 倒数第三和倒数第二个元素
                    fourth_last = target_row[-4]
                    third_last = target_row[-3] #align score
                    second_last = target_row[-2] #quality score
                    output_rows.append([fourth_last, third_last, second_last])
                        
    # 2. 读取原CSV内容
    with open(human_score_csv, newline='', encoding='utf-8') as f:
        reader = csv.reader(f)
        original_rows = list(reader)

    # 3. 检查行数是否匹配
    if len(original_rows) - 1 != len(output_rows):
        raise ValueError("原CSV数据行数与新列数据行数不一致，请检查！")

    # 4. 拼接新列，写入新文件
    with open(human_score_csv, 'w', newline='', encoding='utf-8') as f:
        writer = csv.writer(f)
        # 写表头
        writer.writerow(original_rows[0][0:10] + ['qwen_reason', 'qwen_alignment', 'qwen_quality'])
        # 写数据行
        for orig_row, new_col in zip(original_rows[1:], output_rows):
            writer.writerow(orig_row[0:10] + new_col)
            
            
def corr(combined_csv):
    avg_align_list = []
    avg_quality_list= []
    qwen_align_list = []
    qwen_quality_list = []
    
    with open(combined_csv, newline='', encoding='utf-8') as f:
        reader = csv.reader(f)
        header = next(reader)  # 跳过表头
        for row in reader:
            # 取倒数第3、4、5列
            alignment_123 = [float(row[-9]), float(row[-8]), float(row[-7])]
            # quality_123 = [float(row[-5]), float(row[-4]), float(row[-3])]
            alignment_avg = np.mean(alignment_123)
            # quality_avg = np.mean(quality_123)
            avg_align_list.append(alignment_avg)
            # avg_quality_list.append(quality_avg)
            # 取倒数第2列
            qwen_alignment = float(row[-3])
            qwen_detail = float(row[-2])
            qwen_quality = float(row[-1])
            
            qwen_alignment_qualilty = 0.7*qwen_alignment + 0.2*qwen_detail + 0.1*qwen_quality
            
            qwen_align_list.append(qwen_alignment_qualilty)
            # qwen_quality_list.append(qwen_quality)
            
            

    # 2. 计算相关性
    align_spearman_corr, align_spearman_p = spearmanr(avg_align_list, qwen_align_list)
    align_kendall_corr, align_kendall_p = kendalltau(avg_align_list, qwen_align_list)
    
    # quality_spearman_corr, quality_spearman_p = spearmanr(avg_quality_list, qwen_quality_list)
    # quality_kendall_corr, quality_kendall_p = kendalltau(avg_quality_list, qwen_quality_list)

    print(f"align Spearman correlation: {align_spearman_corr:.4f}, p-value: {align_spearman_p:.4g}")
    print(f"align Kendall correlation: {align_kendall_corr:.4f}, p-value: {align_kendall_p:.4g}")

    print(f"{align_kendall_corr:.4f} & {align_spearman_corr:.4f} &")
    # print(f"quality Spearman correlation: {quality_spearman_corr:.4f}, p-value: {quality_spearman_p:.4g}")
    # print(f"quality Kendall correlation: {quality_kendall_corr:.4f}, p-value: {quality_kendall_p:.4g}")
    
    
if __name__ == "__main__":
    

    
    # human_score_csv = "/group/xihuiliu/sky/reasoning/human_eval/human_score/entity_vqa_2.csv"
    # human_score_csv = "/group/xihuiliu/sky/reasoning/human_eval/human_score/entity_clip.csv"
    
    # input_files = [
    #     "/group/xihuiliu/sky/reasoning/csv/text_image_new/bagel.csv",
    #     "/group/xihuiliu/sky/reasoning/csv/text_image_new/GPT.csv",
    #     "/group/xihuiliu/sky/reasoning/csv/text_image_new/hidream.csv",
    #     "/group/xihuiliu/sky/reasoning/csv/text_image_new/janus_pro_7B.csv",
    #     "/group/xihuiliu/sky/reasoning/csv/text_image_new/sd30_medium.csv",
    # ]
    # human_score_csv = "/group/xihuiliu/sky/reasoning/human_eval/human_score/text_image.csv"
    
    # input_files = [
    #     "/group/xihuiliu/sky/reasoning/csv/entity/bagel.csv",
    #     "/group/xihuiliu/sky/reasoning/csv/entity/GPT.csv",
    #     "/group/xihuiliu/sky/reasoning/csv/entity/hidream.csv",
    #     "/group/xihuiliu/sky/reasoning/csv/entity/janus_pro_7B.csv",
    #     "/group/xihuiliu/sky/reasoning/csv/entity/sd30_medium.csv",
    # ]
    # human_score_csv = "/group/xihuiliu/sky/reasoning/human_eval/human_score/entity.csv"
    
    
    input_files = [
        "/group/xihuiliu/sky/reasoning/csv/physics/bagel.csv",
        "/group/xihuiliu/sky/reasoning/csv/physics/GPT.csv",
        "/group/xihuiliu/sky/reasoning/csv/physics/hidream.csv",
        "/group/xihuiliu/sky/reasoning/csv/physics/janus_pro_7B.csv",
        "/group/xihuiliu/sky/reasoning/csv/physics/sd30_medium.csv",
    ]
    human_score_csv = "/group/xihuiliu/sky/reasoning/human_eval/human_score/physics.csv"
    
    add_to_csv(input_files,human_score_csv)

    corr(human_score_csv)
    
    


# entity 0.7 0.3
# align Spearman correlation: 0.6148, p-value: 1.012e-11
# align Kendall correlation: 0.4732, p-value: 2.094e-10
# quality Spearman correlation: 0.3947, p-value: 4.824e-05
# quality Kendall correlation: 0.3141, p-value: 0.0001425

# VQA_Score
# align Spearman correlation: 0.6133, p-value: 1.175e-11
# align Kendall correlation: 0.4483, p-value: 7.196e-10

# CLIP
# align Spearman correlation: 0.3837, p-value: 8.11e-05
# align Kendall correlation: 0.2732, p-value: 0.0001741