from bert_score import BERTScorer import torch import json import argparse import numpy as np from scipy.stats import ks_2samp, mannwhitneyu, anderson_ksamp import matplotlib.pyplot as plt import re import os import pandas as pd def load_jsonl(file_path): data = [] with open(file_path, 'r') as file: for line in file: data.append(json.loads(line.strip())) return data def dump_txt(data, file_path): with open(file_path, 'w') as file: file.write(str(data) + '\n') def compare_distributions(sample1, sample2): # Kolmogorov-Smirnov Test ks_stat, ks_p_value = ks_2samp(sample1, sample2) print(f"Kolmogorov-Smirnov test statistic: {ks_stat}, p-value: {ks_p_value}") if ks_p_value < 0.05: print("Kolmogorov-Smirnov test: The two samples likely come from different distributions.") else: print("Kolmogorov-Smirnov test: The two samples likely come from the same distribution.") # Mann-Whitney U Test mw_stat, mw_p_value = mannwhitneyu(sample1, sample2, alternative='two-sided') print(f"Mann-Whitney U test statistic: {mw_stat}, p-value: {mw_p_value}") if mw_p_value < 0.05: print("Mann-Whitney U test: The two samples likely come from different distributions.") else: print("Mann-Whitney U test: The two samples likely come from the same distribution.") # Anderson-Darling Test ad_stat, critical_values, ad_significance_level = anderson_ksamp([sample1, sample2]) print(f"Anderson-Darling test statistic: {ad_stat}, significance level: {ad_significance_level}") if ad_stat > critical_values[2]: # Using 5% significance level print("Anderson-Darling test: The two samples likely come from different distributions.") else: print("Anderson-Darling test: The two samples likely come from the same distribution.") return ks_p_value, mw_p_value, ad_stat, critical_values[2] def get_num_from_directory(directory_path): # List to store the extracted numbers numbers = [] # Iterate over each file/directory in the specified path for filename in os.listdir(directory_path): # Use regex to find numbers in the filename match = re.search(r'checkpoint-(\d+)', filename) if match: # Append the extracted number to the list as an integer numbers.append(int(match.group(1))) return numbers parser = argparse.ArgumentParser() parser.add_argument('--model', type=str, default='160m',help='model name') #160m 410m 1b 1.4b 2.8b 6.9b 12b parser.add_argument('--epoch', type=int, default=9,help='model name') parser.add_argument('--size', type=int, default=600,help='model name') parser.add_argument('--subname', type=str, default='arxiv', help='subset name') parser.add_argument('--lr', type=float, default=2e-5, help='learning rate') parser.add_argument('--temp', type=float, default=0.0, help='generation temperature') parser.add_argument('--topp', type=float, default=1.0, help='generation top_p') parser.add_argument('--logging', type=str, default='', help='logging of the file') args = parser.parse_args() device = torch.device("cuda" if torch.cuda.is_available() else "cpu") bert_scorer = BERTScorer('roberta-large', device=device, rescale_with_baseline=True, lang='en') loss_file_member = f'/workspace/{args.subname}_dataset/output_ft_more_layers_{args.subname}_epoch_{args.epoch}_{args.logging}/pythia-{args.model}-member-{args.model}-epoch-{args.epoch}-pile-full-{args.size}-subsets-{args.subname}-{args.lr}/checkpoint-675/trainer_state.json' loss_file_nonmember = f'/workspace/{args.subname}_dataset/output_ft_more_layers_{args.subname}_epoch_{args.epoch}_{args.logging}/pythia-{args.model}-nonmember-{args.model}-epoch-{args.epoch}-pile-full-{args.size}-subsets-{args.subname}-{args.lr}/checkpoint-675/trainer_state.json' loss_datafile_member = json.load(open(loss_file_member))['log_history'] loss_datafile_nonmember = json.load(open(loss_file_nonmember))['log_history'] loss_l_member = [] loss_l_nonmember = [] for i in range(len(loss_datafile_member)): try: loss_data_memeber = loss_datafile_member[i]['loss'] loss_l_member.append(loss_data_memeber) except: continue for i in range(len(loss_datafile_nonmember)): try: loss_data_nonmember = loss_datafile_nonmember[i]['loss'] loss_l_nonmember.append(loss_data_nonmember) except: continue # Find the largest value in the list max_value_member = max(loss_l_member) max_value_nonmember = max(loss_l_nonmember) # Divide each value by the largest value normalized_loss_l_member = [x / max_value_member for x in loss_l_member] normalized_loss_l_nonmember = [x / max_value_nonmember for x in loss_l_nonmember] results_dict = {} ks_p_value_l=[] mw_p_value_l=[] directory_path = f"/workspace/{args.subname}_dataset/output_ft_more_layers_{args.subname}_epoch_{args.epoch}_{args.logging}/pythia-{args.model}-member-{args.model}-epoch-{args.epoch}-pile-full-{args.size}-subsets-{args.subname}-{args.lr}" numbers = get_num_from_directory(directory_path) numbers.sort() for num in numbers: for candidate in ['member', 'nonmember']: print(f"#############{num}############") model_name = f'pythia-{args.model}' log_str = f'{candidate}-{args.model}-epoch-{args.epoch}' response_orig = load_jsonl(f'/workspace/responses_ft_more_layers_{args.subname}_epoch_{args.epoch}_{args.logging}/{model_name}-{log_str}-pile-full-{args.size}-subsets-{args.subname}-{args.lr}-orig.jsonl') response_ft = load_jsonl(f'/workspace/responses_ft_more_layers_{args.subname}_epoch_{args.epoch}_{args.logging}/all_checkpoint/{model_name}-{log_str}-pile-full-{args.size}-subsets-{args.subname}-{args.lr}-{num}-ft.jsonl') response_only_orig = [] response_only_ft = [] for i in range(len(response_orig)): response_only_orig.append(response_orig[i]['output_text']) response_only_ft.append(response_ft[i]['output_text']) ctc_scores = bert_scorer.score(response_only_ft, response_only_orig)[2] results_dict[candidate]=ctc_scores ks_p_value, mw_p_value, ad_stat, adcv=compare_distributions(results_dict['member'], results_dict['nonmember']) os.makedirs(f'bert_results_ft_more_layers_{args.subname}_epoch_{args.epoch}_{args.logging}', exist_ok=True) os.makedirs(f'p_value_loss_ft_more_layers_{args.subname}_epoch_{args.epoch}_{args.logging}', exist_ok=True) file_path =f'/workspace/bert_results_ft_more_layers_{args.subname}_epoch_{args.epoch}_{args.logging}/pile_full_bert_{args.model}_{args.epoch}_{args.subname}_{args.size}_{args.lr}_test.txt' txt_info=f''' Kolmogorov-Smirnov test statistic: p-value: {ks_p_value} Mann-Whitney U test statistic: p-value: {mw_p_value} Anderson-Darling test statistic: {ad_stat} critical-value:{adcv} ''' dump_txt(txt_info, file_path) ks_p_value_l.append(ks_p_value) mw_p_value_l.append(mw_p_value) plt.figure(figsize=(10, 6)) plt.plot(ks_p_value_l, marker='o', linestyle='-', color='b', label='P-value') # Plot the second line plt.plot(normalized_loss_l_member, marker='s', linestyle='--', color='g', label='member loss') # Plot the third line plt.plot(normalized_loss_l_nonmember, marker='^', linestyle='-.', color='r', label='nonmember loss') # Add title and labels plt.title(f'P-Value subsets-{args.subname}-{args.lr}') plt.xlabel('Iteration') plt.ylabel('Loss') # Add legend plt.legend() # Show grid plt.grid(True) plt.savefig(f'/workspace/p_value_loss_ft_more_layers_{args.subname}_epoch_{args.epoch}_{args.logging}/{args.model}-{args.epoch}-pile-full-{args.size}-subsets-{args.subname}-{args.lr}-ks.png') plt.figure(figsize=(10, 6)) plt.plot(mw_p_value_l, marker='o', linestyle='-', color='b', label='Loss') plt.plot(normalized_loss_l_member, marker='s', linestyle='--', color='g', label='member loss') # Plot the third line plt.plot(normalized_loss_l_nonmember, marker='^', linestyle='-.', color='r', label='nonmember loss') # Add title and labels plt.title(f'P-Value subsets-{args.subname}-{args.lr}') plt.xlabel('Iteration') plt.ylabel('Loss') # Add legend plt.legend() # Show grid plt.grid(True) plt.savefig(f'/workspace/p_value_loss_ft_more_layers_{args.subname}_epoch_{args.epoch}_{args.logging}/{args.model}-{args.epoch}-pile-full-{args.size}-subsets-{args.subname}-{args.lr}-mw.png') print(len(loss_l_member)) print(len(loss_l_nonmember)) print(len(ks_p_value_l)) print(len(mw_p_value_l)) df_dict = {'member_loss': loss_l_member, 'nonmember_loss': loss_l_nonmember} df_loss = pd.DataFrame(df_dict) df_dict_test = {'ks_pvalue': ks_p_value_l, 'mw_pvalue': mw_p_value_l} df_pvalue = pd.DataFrame(df_dict_test) df_normalized_loss_dict = {'member_loss': normalized_loss_l_member, 'nonmember_loss': normalized_loss_l_nonmember} df_normalized_loss = pd.DataFrame(df_normalized_loss_dict) df_loss.to_csv(f'/workspace/pile_{args.subname}_temp_{args.temp}_topp_{args.topp}_loss_ft_more_layers_{args.subname}_epoch_{args.epoch}_{args.logging}.csv', index=False) df_pvalue.to_csv(f'/workspace/pile_{args.subname}_temp_{args.temp}_topp_{args.topp}_pvalue_ft_more_layers_{args.subname}_epoch_{args.epoch}_{args.logging}.csv', index=False) df_normalized_loss.to_csv(f'/workspace/pile_{args.subname}_temp_{args.temp}_topp_{args.topp}_normalized_loss_ft_more_layers_{args.subname}_epoch_{args.epoch}_{args.logging}.csv', index=False)