From bb34c6e32520ad98c9b3415782e9ccfca505ef78 Mon Sep 17 00:00:00 2001 From: jenniew Date: Fri, 12 Apr 2024 13:26:36 +0800 Subject: [PATCH] Mark Color Modification --- .../benchmark/perplexity/ppl_csv_to_html.py | 182 ++++++++++-------- 1 file changed, 97 insertions(+), 85 deletions(-) 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 24efaead..ff4eafeb 100644 --- a/python/llm/test/benchmark/perplexity/ppl_csv_to_html.py +++ b/python/llm/test/benchmark/perplexity/ppl_csv_to_html.py @@ -28,7 +28,7 @@ def highlight_vals(val, max=3.0, color1='red', color2='green', color3='yellow', return 'background-color: %s' % color1 elif val <= -max: return 'background-color: %s' % color2 - elif val != 0.0 and not pd.isna(val) and is_last: + elif val != 0.0 and is_last: return 'background-color: %s' % color3 else: return '' @@ -38,7 +38,7 @@ def nonzero_min(lst): return min(non_zero_lst) if non_zero_lst else None 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 + return not any(diff < (-threshold) for diff in diff_ppl_result if isinstance(diff, float)) def add_to_dict(dict, key, value): if key not in dict: @@ -58,17 +58,11 @@ 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] = { '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): @@ -112,10 +106,9 @@ def main(): # 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) + task = 'ppl_result' + latest_csv[f'{task}_FP16'] = latest_csv['Model'].apply(lambda model: fp16_dict.get(model, {}).get(task, '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: @@ -127,13 +120,8 @@ def main(): 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) - ppl_result = 'ppl_result'#en='en' - #zh='zh' + ppl_result = 'ppl_result' csv_dict = {} for csv_file in csv_files: @@ -143,20 +131,15 @@ def main(): current_csv_precision=current_csv_row['Precision'].strip() 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_ppl_result=latest_csv_row[ppl_result]#latest_en=latest_csv_row[en] + latest_ppl_result=latest_csv_row[ppl_result] print("This is the latest_ppl_result",latest_ppl_result) - # latest_zh=latest_csv_row[zh] in_previous_flag=False @@ -167,84 +150,113 @@ def main(): if latest_csv_model==previous_csv_model and latest_csv_precision==previous_csv_precision: - previous_ppl_result=previous_csv_row[ppl_result] #previous_en=previous_csv_row[en] + previous_ppl_result=previous_csv_row[ppl_result] 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: + + if previous_ppl_result > 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_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=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) + latest_csv.insert(loc=7,column='ppl_result_diff_last(%)',value=diff_ppl_result) - diffs_within_normal_range = is_diffs_within_normal_range(diff_ppl_result, threshold=highlight_threshold)#en, diff_zh, threshold=highlight_threshold) + diffs_within_normal_range = is_diffs_within_normal_range(diff_ppl_result, threshold=highlight_threshold) - subset1=['last_diff_ppl_result(%)']#['diff_en(%)','diff_zh(%)'] + subset1='ppl_result_diff_last(%)' - 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}'} + # columns will be different: columns_sec={'ppl_result': '{:.2f}'} + columns={'ppl_result': '{:.2f}', 'last_ppl_result': '{:.2f}', 'ppl_result_diff_last(%)': '{:.2f}'} + # This is the same latest_csv.drop('Index', axis=1, inplace=True) + # subset is different - styled_df = latest_csv.style.format(columns).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) + styled_df = latest_csv.style.format(columns).applymap(lambda val: highlight_vals(val, max=highlight_threshold, is_last=True), subset=['ppl_result_diff_last(%)']) + styled_df = styled_df.applymap(lambda val: highlight_vals(val, max=highlight_threshold, is_last=False), subset=['ppl_result_diff_FP16(%)']) + else: + columns={'ppl_result': '{:.2f}'} 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) - + styled_df = latest_csv.style.format(columns).applymap(lambda val: highlight_vals(val, max=highlight_threshold, is_last=False), subset=['ppl_result_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) + if args.baseline_path and not diffs_within_normal_range: print("The diffs are outside the normal range: %" + str(highlight_threshold)) - return 1 + return 1 return 0 - + if __name__ == "__main__": sys.exit(main()) + + + + + + + + + + + + + + + +# styled_df = latest_csv.style.format(columns).applymap(lambda val: highlight_vals(val, max=3.0, is_last=True), subset=subset1) + +# # These are the same: +# task = 'ppl_result' +# 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) +# else: +# latest_csv.drop('Index', axis=1, inplace=True) +# columns_sec={'ppl_result': '{:.2f}'} +# styled_df = latest_csv.style.format(columns_sec).applymap(lambda val: highlight_vals(val, max=3.0, is_last=True))#, subset=subset1) +# task = 'ppl_result' +# 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) + +# if args.baseline_path and not diffs_within_normal_range: +# print("The diffs are outside the normal range: %" + str(highlight_threshold)) +# return 1 +# return 0 + +# if __name__ == "__main__": +# sys.exit(main())