def main()

in utils/compute_score.py [0:0]

• 60 lines of code
• 3 McCabe index (conditional complexity)

def main(model, path):
    files = os.listdir(f'{path}')

    qa_bert_list = []
    tg_bert_list, sub_task_tg_bert_list = [], []
    qa_list, tg_list, e2z_bleu_list, e2z_comet_list, z2e_bleu_list, z2e_comet_list, comet_list, acc_list, f1_list = [], [], [], [], [], [], [], [], []
    sub_task_qa_list, sub_task_tg_list, sub_task_tc_list, sub_task_re_list = [], [], [], []
    for f in files:
        if model in f:
            print('Model: %s' % model)
            file_path = f'{path}/{f}'
            # QA
            rouge_l, qa_bert = compute_finqa(file_path)
            qa_list.append(rouge_l)
            qa_bert_list.append(qa_bert)
            # TG
            rouge_l_tg, sub_task_tg, tg_bert, tg_sub_task_bert = compute_text_generation(file_path)
            tg_list.append(rouge_l_tg)
            sub_task_tg_list.append(sub_task_tg)
            tg_bert_list.append(tg_bert)
            sub_task_tg_bert_list.append(tg_sub_task_bert)
            # MT-e2zh
            bleu, comet = compute_nmt_en2zh(file_path)
            e2z_bleu_list.append(bleu)
            e2z_comet_list.append(comet)
            # MT-zh2e
            bleu, comet = compute_nmt_zh2en(file_path)
            z2e_bleu_list.append(bleu)
            z2e_comet_list.append(comet)
            # TC
            acc, sub_task_acc = compute_text_classification(file_path)
            acc_list.append(acc)
            sub_task_tc_list.append(sub_task_acc)
            # RE
            f1, sub_task_re = compute_extraction(file_path)
            f1_list.append(f1)
            sub_task_re_list.append(sub_task_re)

    # 总分类
    print('QA mean: %s' % np.mean(qa_list), '\n')
    print('QA Std: %s' % np.std(qa_list), '\n')
    print('QA bert mean: %s' % np.mean(qa_bert_list), '\n')
    print('QA bert Std: %s' % np.std(qa_bert_list), '\n')
    print('TG mean: %s' % np.mean(tg_list), '\n')
    print('TG Std: %s' % np.std(tg_list), '\n')
    print('TG bert mean: %s' % np.mean(tg_bert_list), '\n')
    print('TG bert Std: %s' % np.std(tg_bert_list), '\n')
    print('EN2CH mean: %s' % np.mean(e2z_bleu_list), '\n')
    print('EN2CH Std: %s' % np.std(e2z_bleu_list), '\n')
    print('EN2CH comet mean: %s' % np.mean(e2z_comet_list), '\n')
    print('EN2CH comet Std: %s' % np.std(e2z_comet_list), '\n')
    print('CH2EN mean: %s' % np.mean(z2e_bleu_list), '\n')
    print('CH2EN Std: %s' % np.std(z2e_bleu_list), '\n')
    print('CH2EN comet mean: %s' % np.mean(z2e_comet_list), '\n')
    print('CH2EN comet Std: %s' % np.std(z2e_comet_list), '\n')
    print('TG mean: %s' % np.mean(tg_list), '\n')
    print('TG Std: %s' % np.std(tg_list), '\n')

    ## 子任务分类
    # TG
    tg_df = pd.DataFrame(sub_task_tg_list)
    print('TG sub task mean: %s' % tg_df.mean())
    print('TG sub task std: %s' % tg_df.std())
    tg_bert_df = pd.DataFrame(sub_task_tg_bert_list)
    print('TG sub task bert mean: %s' % tg_bert_df.mean())
    print('TG sub task bert std: %s' % tg_bert_df.std())
    # TC
    tc_df = pd.DataFrame(sub_task_tc_list)
    print('TG sub task mean: %s' % tc_df.mean())
    print('TG sub task std: %s' % tc_df.std())
    # RE
    re_df = pd.DataFrame(sub_task_re_list)
    print('TG sub task mean: %s' % re_df.mean())
    print('TG sub task std: %s' % re_df.std())