From b151a9b672fcdb3da4982e6a7802262d39524903 Mon Sep 17 00:00:00 2001 From: jenniew Date: Thu, 11 Apr 2024 17:35:36 +0800 Subject: [PATCH] edit csv_to_html to combine en & zh --- python/llm/test/benchmark/perplexity/fp16.csv | 16 +-- .../benchmark/perplexity/ppl_csv_to_html.py | 124 ++++++++++++------ 2 files changed, 92 insertions(+), 48 deletions(-) diff --git a/python/llm/test/benchmark/perplexity/fp16.csv b/python/llm/test/benchmark/perplexity/fp16.csv index 2b812f84..00915746 100644 --- a/python/llm/test/benchmark/perplexity/fp16.csv +++ b/python/llm/test/benchmark/perplexity/fp16.csv @@ -1,8 +1,8 @@ -Index,Model,Precision,en,zh -0,Llama-2-7b-chat-hf,fp16,4.7019, -1,chatglm2-6b,fp16,,22.321 -2,chatglm3-6b,fp16,,30.1281 -3,Baichuan2-7B-Chat,fp16,,10.7676 -4,mpt-7b-chat,fp16,5.7882, -5,falcon-7b-instruct-with-patch,fp16,5.2532, -6,mistral-7b-v0.1,fp16,3.6597, \ No newline at end of file +Index,Model,Precision,ppl_result +0,Llama-2-7b-chat-hf,fp16,4.7019 +1,chatglm2-6b,fp16,22.321 +2,chatglm3-6b,fp16,30.1281 +3,Baichuan2-7B-Chat,fp16,10.7676 +4,mpt-7b-chat,fp16,5.7882 +5,falcon-7b-instruct-with-patch,fp16,5.2532 +6,Mistral-7B-v0.1,fp16,3.6597 diff --git a/python/llm/test/benchmark/perplexity/ppl_csv_to_html.py b/python/llm/test/benchmark/perplexity/ppl_csv_to_html.py index d1175d19..24efaead 100644 --- a/python/llm/test/benchmark/perplexity/ppl_csv_to_html.py +++ b/python/llm/test/benchmark/perplexity/ppl_csv_to_html.py @@ -37,8 +37,8 @@ def nonzero_min(lst): non_zero_lst = [num for num in lst if num > 0.0] return min(non_zero_lst) if non_zero_lst else None -def is_diffs_within_normal_range(diff_en, diff_zh, threshold=5.0): - return not any(diff < (-threshold) for diff in diff_en + diff_zh if isinstance(diff, float)) +def is_diffs_within_normal_range(diff_ppl_result, threshold=5.0): + return not any(diff < (-threshold) for diff in diff_ppl_result + diff_ppl_result if isinstance(diff, float))#diff_en + diff_zh def add_to_dict(dict, key, value): if key not in dict: @@ -58,11 +58,17 @@ def create_fp16_dict(fp16_path): fp16_dict = {} for _, row in fp16_df.iterrows(): model = row['Model'] + # print("I want to test the ppl_result row", row['ppl_result']) # Formalize the data to have 2 decimal places + # fp16_dict[model] = { + # 'en': "{:.2f}".format(row['en']), + # 'zh': "{:.2f}".format(row['zh']) + # } fp16_dict[model] = { - 'en': "{:.2f}".format(row['en']), - 'zh': "{:.2f}".format(row['zh']) + 'ppl_result': "{:.2f}".format(row['ppl_result']) + # 'zh': "{:.2f}".format(row['zh']) } + # print(fp16_dict[model]) return fp16_dict def calculate_percentage_difference(current, fp16): @@ -104,24 +110,30 @@ def main(): diffs_within_normal_range = True - # Add display of FP16 values for each model and add percentage difference column - for task in ['en', 'zh']: + # Add display of FP16 values for each model and add percentage difference column + + for task in ['ppl_result']:#['en', 'zh']: latest_csv[f'{task}_FP16'] = latest_csv['Model'].apply(lambda model: fp16_dict.get(model, {}).get(task, 'N/A')) + # print("This is the stuff I want to check",latest_csv['Model'].apply(lambda model: fp16_dict.get(model, {}).get(task, 'N/A') == 'N/A')) latest_csv[f'{task}_diff_FP16(%)'] = latest_csv.apply(lambda row: calculate_percentage_difference(row[task], row[f'{task}_FP16']), axis=1) + print(csv_files) if len(csv_files)>1: + if args.baseline_path: previous_csv = pd.read_csv(args.baseline_path, index_col=0) else: previous_csv = pd.read_csv(csv_files[1], index_col=0) - last_en=['']*len(latest_csv.index) - diff_en=['']*len(latest_csv.index) - last_zh=['']*len(latest_csv.index) - diff_zh=['']*len(latest_csv.index) + last_ppl_result=['']*len(latest_csv.index) + diff_ppl_result=['']*len(latest_csv.index) + # last_en=['']*len(latest_csv.index) + # diff_en=['']*len(latest_csv.index) + # last_zh=['']*len(latest_csv.index) + # diff_zh=['']*len(latest_csv.index) - en='en' - zh='zh' + ppl_result = 'ppl_result'#en='en' + #zh='zh' csv_dict = {} for csv_file in csv_files: @@ -129,17 +141,22 @@ def main(): for current_csv_ind,current_csv_row in current_csv.iterrows(): current_csv_model=current_csv_row['Model'].strip() current_csv_precision=current_csv_row['Precision'].strip() - current_csv_model_en=current_csv_model+'-'+current_csv_precision+'-'+'en' - current_csv_model_zh=current_csv_model+'-'+current_csv_precision+'-'+'zh' - add_to_dict(csv_dict, current_csv_model_en, current_csv_row[en]) - add_to_dict(csv_dict, current_csv_model_zh, current_csv_row[zh]) + + current_csv_model_ppl_result=current_csv_model+'-'+current_csv_precision+'-'+'ppl_result' + # current_csv_model_en=current_csv_model+'-'+current_csv_precision+'-'+'en' + # current_csv_model_zh=current_csv_model+'-'+current_csv_precision+'-'+'zh' + + add_to_dict(csv_dict, current_csv_model_ppl_result, current_csv_row[ppl_result]) + # add_to_dict(csv_dict, current_csv_model_en, current_csv_row[en]) + # add_to_dict(csv_dict, current_csv_model_zh, current_csv_row[zh]) for latest_csv_ind,latest_csv_row in latest_csv.iterrows(): latest_csv_model=latest_csv_row['Model'].strip() latest_csv_precision=latest_csv_row['Precision'].strip() - latest_en=latest_csv_row[en] - latest_zh=latest_csv_row[zh] + latest_ppl_result=latest_csv_row[ppl_result]#latest_en=latest_csv_row[en] + print("This is the latest_ppl_result",latest_ppl_result) + # latest_zh=latest_csv_row[zh] in_previous_flag=False @@ -150,40 +167,48 @@ def main(): if latest_csv_model==previous_csv_model and latest_csv_precision==previous_csv_precision: - previous_en=previous_csv_row[en] - previous_zh=previous_csv_row[zh] - if previous_en > 0.0 or previous_zh > 0.0: - last_en[latest_csv_ind]=previous_en - diff_en[latest_csv_ind]=round((latest_en-previous_en)*100/previous_en,2) - last_zh[latest_csv_ind]=previous_zh - diff_zh[latest_csv_ind]=round((latest_zh-previous_zh)*100/previous_zh,2) + previous_ppl_result=previous_csv_row[ppl_result] #previous_en=previous_csv_row[en] + print("This is the previous_ppl_result", previous_ppl_result) + # previous_zh=previous_csv_row[zh] + if previous_ppl_result > 0.0:# or previous_zh > 0.0: + last_ppl_result[latest_csv_ind]=previous_ppl_result + diff_ppl_result[latest_csv_ind]=round((latest_ppl_result-previous_ppl_result)*100/previous_ppl_result,2) + # last_zh[latest_csv_ind]=previous_zh + # diff_zh[latest_csv_ind]=round((latest_zh-previous_zh)*100/previous_zh,2) in_previous_flag=True + # last_en[latest_csv_ind]=previous_en + # diff_en[latest_csv_ind]=round((latest_en-previous_en)*100/previous_en,2) + # last_zh[latest_csv_ind]=previous_zh + # diff_zh[latest_csv_ind]=round((latest_zh-previous_zh)*100/previous_zh,2) + # in_previous_flag=True + if not in_previous_flag: - last_en[latest_csv_ind]=pd.NA - diff_en[latest_csv_ind]=pd.NA - last_zh[latest_csv_ind]=pd.NA - diff_zh[latest_csv_ind]=pd.NA + last_ppl_result[latest_csv_ind]=pd.NA + diff_ppl_result[latest_csv_ind]=pd.NA + # last_zh[latest_csv_ind]=pd.NA + # diff_zh[latest_csv_ind]=pd.NA - latest_csv.insert(loc=9,column='last_en',value=last_en) - latest_csv.insert(loc=10,column='diff_en(%)',value=diff_en) - latest_csv.insert(loc=11,column='last_zh',value=last_zh) - latest_csv.insert(loc=12,column='diff_zh(%)',value=diff_zh) + latest_csv.insert(loc=6,column='last_ppl_result',value=last_ppl_result) + latest_csv.insert(loc=7,column='last_diff_ppl_result(%)',value=diff_ppl_result) + # latest_csv.insert(loc=11,column='last_zh',value=last_zh) + # latest_csv.insert(loc=12,column='diff_zh(%)',value=diff_zh) - diffs_within_normal_range = is_diffs_within_normal_range(diff_en, diff_zh, threshold=highlight_threshold) + diffs_within_normal_range = is_diffs_within_normal_range(diff_ppl_result, threshold=highlight_threshold)#en, diff_zh, threshold=highlight_threshold) - subset1=['diff_en(%)','diff_zh(%)'] + subset1=['last_diff_ppl_result(%)']#['diff_en(%)','diff_zh(%)'] - columns={'en': '{:.2f}', 'zh': '{:.2f}', 'last_en': '{:.2f}', 'diff_en(%)': '{:.2f}', - 'last_zh': '{:.2f}', 'diff_zh(%)': '{:.2f}'} - + columns={'ppl_result': '{:.2f}', 'last_ppl_result': '{:.2f}', 'last_diff_ppl_result(%)': '{:.2f}'} + #{'en': '{:.2f}', 'zh': '{:.2f}', 'last_en': '{:.2f}', 'diff_en(%)': '{:.2f}', + #'last_zh': '{:.2f}', 'diff_zh(%)': '{:.2f}'} latest_csv.drop('Index', axis=1, inplace=True) styled_df = latest_csv.style.format(columns).applymap(lambda val: highlight_vals(val, max=3.0, is_last=True), subset=subset1) - for task in ['en', 'zh']: + for task in ['ppl_result']:#['en', 'zh']: styled_df = styled_df.applymap(lambda val: highlight_vals(val, max=highlight_threshold, is_last=False), subset=[f'{task}_diff_FP16(%)']) + # add css style to restrict width and wrap text styled_df.set_table_styles([{ 'selector': 'th, td', @@ -195,7 +220,26 @@ def main(): with open(daily_html, 'w') as f: f.write(html_output) else: - latest_csv.to_html(daily_html) + latest_csv.drop('Index', axis=1, inplace=True) + columns_sec={'ppl_result': '{:.2f}'} + #{'en': '{:.2f}', 'zh': '{:.2f}', 'last_en': '{:.2f}', 'diff_en(%)': '{:.2f}', + #'last_zh': '{:.2f}', 'diff_zh(%)': '{:.2f}'} + styled_df = latest_csv.style.format(columns_sec).applymap(lambda val: highlight_vals(val, max=3.0, is_last=True))#, subset=subset1) + for task in ['ppl_result']:#['en', 'zh']: + styled_df = styled_df.applymap(lambda val: highlight_vals(val, max=highlight_threshold, is_last=False), subset=[f'{task}_diff_FP16(%)']) + + + # add css style to restrict width and wrap text + styled_df.set_table_styles([{ + 'selector': 'th, td', + 'props': [('max-width', '88px'), ('word-wrap', 'break-word')] + }], overwrite=False) + + html_output = styled_df.set_table_attributes("border=1").to_html() + + with open(daily_html, 'w') as f: + f.write(html_output) + # latest_csv.to_html(daily_html) if args.baseline_path and not diffs_within_normal_range: print("The diffs are outside the normal range: %" + str(highlight_threshold))