import os import torch from transformers import AutoModelForCausalLM, AutoTokenizer, Trainer, TrainingArguments, DataCollatorForLanguageModeling import datasets from datasets import load_dataset from datasets import Dataset, DatasetDict from peft import get_peft_model, LoraConfig, TaskType import json from tqdm import tqdm import pandas as pd from functools import partial import argparse import re import matplotlib matplotlib.use('Agg') # Use the Agg backend for non-interactive plotting import matplotlib.pyplot as plt 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=3,help='model name') #160m 410m 1b 1.4b 2.8b 6.9b 12b parser.add_argument('--subname', type=str, default='arxiv',help='model name') parser.add_argument('--size', type=int, default=600 ,help='model 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('--candidate', type=str, default='member', help='learning rate') args = parser.parse_args() # Disable wandb logging os.environ["WANDB_DISABLED"] = "true" loss_file = f'/workspace/{args.subname}_dataset/output_ft_more_layers_{args.subname}_epoch_{args.epoch}_mlp/pythia-{args.model}-{args.candidate}-{args.model}-epoch-{args.epoch}-pile-full-{args.size}-subsets-{args.subname}-{args.lr}/checkpoint-675/trainer_state.json' loss_datafile = json.load(open(loss_file))['log_history'] loss_l = [] for i in range(len(loss_datafile)): try: loss_data = loss_datafile[i]['loss'] loss_l.append(loss_data) except: continue model_name = f'pythia-{args.model}' # Load the tokenizer and model model_name_hf_ori = f"/workspace/{model_name}" # You can choose other sizes as well tokenizer = AutoTokenizer.from_pretrained(model_name_hf_ori) tokenizer.padding_side = "left" # Add padding token if missing if tokenizer.pad_token is None: tokenizer.pad_token = tokenizer.eos_token tokenizer.pad_token_id = tokenizer.eos_token_id data_files = f"/workspace/dataset_inference/{args.subname}_train.jsonl" raw_train_data_df = pd.read_json(data_files, lines=True) #Pile Validation Set val_data_files = f"/workspace/dataset_inference/{args.subname}_val.jsonl" raw_val_data_df = pd.read_json(val_data_files, lines=True) tds=Dataset.from_pandas(raw_train_data_df) vds=Dataset.from_pandas(raw_val_data_df) raw_data = DatasetDict() raw_data['train'] = tds raw_data['validation'] = vds # Tokenize the input data def tokenize_function(examples,max_length=384): tokens = tokenizer(examples["text"], padding="max_length", truncation=True, max_length=max_length) #tokens["labels"] = tokens["input_ids"].copy() return tokens data_num = 1000 A_members = raw_data['train'].shuffle(seed=42).select(range(0, args.size)).map(partial(tokenize_function,max_length=512), batched=True, remove_columns=["text"]) A_nonmembers = raw_data['validation'].shuffle(seed=42).select(range(0, args.size)).map(partial(tokenize_function,max_length=512), batched=True, remove_columns=["text"]) B_members = raw_data['train'].shuffle(seed=42).select(range(data_num, data_num*2)).map(tokenize_function, batched=True, remove_columns=["text"]) B_nonmembers = raw_data['validation'].shuffle(seed=42).select(range(data_num, data_num*2)).map(tokenize_function, batched=True, remove_columns=["text"]) 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 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_jsonl(data, file_path): with open(file_path, 'w') as file: for item in data: json.dump(item, file) file.write('\n') def generate_responses(model,ds): response_list = [] for item in tqdm(ds): input_ids = torch.tensor(item['input_ids']).reshape(1,-1).to(model.device) input_len = input_ids.shape[1] pred = model.generate(input_ids, max_new_tokens=100) input_text = tokenizer.decode(pred[0][:input_len], skip_special_tokens=True) output_text = tokenizer.decode(pred[0][input_len:], skip_special_tokens=True) response_list.append({'output_text':output_text,'input_text':input_text}) return response_list def generate_responses(model,ds,temperature,top_p): model.eval() #print(type(ds[0])) #print(ds[0]) inputs = torch.tensor([item['input_ids'] for item in ds]).to(model.device) masks = torch.tensor([item['attention_mask'] for item in ds]).to(model.device) num_input,input_len = inputs.shape input_text = [] output_text = [] bs = 10 for i in tqdm(range(0,num_input,bs)): pred = model.generate(inputs=inputs[i:i+bs], attention_mask=masks[i:i+bs],max_new_tokens=100, temperature=temperature, top_p=top_p).detach() input_text += tokenizer.batch_decode(pred[:,:input_len], skip_special_tokens=True) output_text += tokenizer.batch_decode(pred[:,input_len:], skip_special_tokens=True) return [{'output_text':a,'input_text':b} for a,b in zip(output_text,input_text)] def run(train_dataset,eval_dataset,log_str, loss_l, args): directory_path = f"/workspace/{args.subname}_dataset/output_ft_more_layers_{args.subname}_epoch_{args.epoch}_mlp/pythia-{args.model}-{args.candidate}-{args.model}-epoch-{args.epoch}-pile-full-{args.size}-subsets-{args.subname}-{args.lr}" numbers = get_num_from_directory(directory_path) min_loss_index = loss_l.index(min(loss_l)) os.makedirs(f'responses_ft_more_layers_{args.subname}_epoch_{args.epoch}_mlp/all_checkpoint', exist_ok=True) for num in numbers: model_name_hf = f"/workspace/{args.subname}_dataset/output_ft_more_layers_{args.subname}_epoch_{args.epoch}_mlp/pythia-{args.model}-{args.candidate}-{args.model}-epoch-{args.epoch}-pile-full-{args.size}-subsets-{args.subname}-{args.lr}/checkpoint-{num}" # You can choose other sizes as well if torch.cuda.is_available(): device = torch.device("cuda") model = AutoModelForCausalLM.from_pretrained(model_name_hf).to(device) #model.to(device) model.eval() response_list = generate_responses(model,eval_dataset, args.temp, args.topp) if num == numbers[min_loss_index]: dump_jsonl(response_list,f'responses_ft_more_layers_{args.subname}_epoch_{args.epoch}_mlp/all_checkpoint/{model_name}-{log_str}-{num}-ft.jsonl') else: dump_jsonl(response_list,f'responses_ft_more_layers_{args.subname}_epoch_{args.epoch}_mlp/all_checkpoint/{model_name}-{log_str}-{num}-ft.jsonl') run(A_members,B_members,f'member-{args.model}-epoch-{args.epoch}-pile-full-{args.size}-subsets-{args.subname}-{args.lr}', loss_l, args) run(A_nonmembers,B_nonmembers,f'nonmember-{args.model}-epoch-{args.epoch}-pile-full-{args.size}-subsets-{args.subname}-{args.lr}', loss_l, args)