-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathprecompute_document_mia_scores.py
168 lines (142 loc) · 6.26 KB
/
precompute_document_mia_scores.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
# %%
import logging
logging.basicConfig(level='ERROR')
import argparse
import json
import os
import random
import zlib
import datasets
import numpy as np
import torch
from datasets import concatenate_datasets, load_dataset
from tqdm import tqdm
from transformers import AutoModelForCausalLM, AutoTokenizer
'''
This script chunks each document into paragraphs of args.max_length tokens and computes the MIA scores for each paragraph.
'''
# %%
def calculatePerplexity(sentence, model, tokenizer):
"""
exp(loss)
"""
encodings = tokenizer(sentence, return_tensors='pt', truncation=True, max_length=2048)
if model.device.type == "cuda":
encodings = {k: v.cuda() for k, v in encodings.items()}
with torch.no_grad():
outputs = model(**encodings, labels=encodings['input_ids'])
loss, logits = outputs[:2]
'''
extract logits:
'''
# Apply softmax to the logits to get probabilities
probabilities = torch.nn.functional.log_softmax(logits, dim=-1)
# probabilities = torch.nn.functional.softmax(logits, dim=-1)
all_prob = []
input_ids_processed = encodings['input_ids'][0][1:]
for i, token_id in enumerate(input_ids_processed):
probability = probabilities[0, i, token_id].item()
all_prob.append(probability)
return torch.exp(loss).item(), all_prob, loss.item()
# %%
# in run.py you have a variant of this function with one more MIA: ppl/Ref_ppl
def inference(model1, tokenizer1, text):
pred = {}
p1, all_prob, p1_likelihood = calculatePerplexity(text, model1, tokenizer1)
p_lower, _, p_lower_likelihood = calculatePerplexity(text.lower(), model1, tokenizer1)
# ppl
pred["ppl"] = p1
# Ratio of log ppl of lower-case and normal-case
pred["ppl/lowercase_ppl"] = -(np.log(p_lower) / np.log(p1)).item()
# Ratio of log ppl of large and zlib
zlib_entropy = len(zlib.compress(bytes(text, 'utf-8')))
pred["ppl/zlib"] = np.log(p1)/zlib_entropy
# min-k prob
for ratio in [0.05, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6]:
k_length = int(len(all_prob)*ratio)
topk_prob = np.sort(all_prob)[:k_length]
pred[f"Min_{ratio*100}% Prob"] = -np.mean(topk_prob).item()
return pred
# %%
def create_text(x):
conversation = x['conversations']
text = ""
for message in conversation:
text += message['from'] + ": " + message['value'] + "\n"
return {"text": text}
def create_chunks(text, tokenizer1, max_length):
tokens = tokenizer1.encode(text, add_special_tokens=True)
chunks = [tokenizer1.decode(tokens[i:i+max_length], skip_special_tokens=True) for i in range(0, len(tokens), max_length)]
return chunks
def parse_args():
parser = argparse.ArgumentParser()
parser.add_argument("--model_name", type=str, default="EleutherAI/pythia-2.8b")
parser.add_argument("--dataset_name", type=str, default="haritzpuerto/the_pile_00_arxiv")
parser.add_argument("--max_length", type=int, default=2048)
parser.add_argument("--filter_outliers", action="store_true")
parser.add_argument("--min_chars", type=int, default=100)
parser.add_argument("--output_path", type=str)
parser.add_argument("--cache_dir", type=str, default="/tmp")
parser.add_argument("--seed", type=int, default=0)
parser.add_argument("--sample_size", type=int, default=None)
return parser.parse_args()
if __name__ == "__main__":
'''
How to run
python precompute_mia_docs.py \
--model_name EleutherAI/pythia-2.8b \
--output_path out/doc_mia/EleutherAI/pythia-2.8b/arxiv/2048_tokens \
--filter_outliers
'''
args = parse_args()
random.seed(args.seed)
# %%
model1 = AutoModelForCausalLM.from_pretrained(args.model_name, return_dict=True, device_map='auto', torch_dtype=torch.bfloat16, cache_dir=args.cache_dir)
model1.eval()
tokenizer1 = AutoTokenizer.from_pretrained(args.model_name)
# %%
ds = load_dataset(args.dataset_name)
if args.filter_outliers:
# removing outlier docs
ds['train'] = ds['train'].filter(lambda x: len(x["text"]) > args.min_chars)
ds['validation'] = ds['validation'].filter(lambda x: len(x["text"]) > args.min_chars)
ds['test'] = ds['test'].filter(lambda x: len(x["text"]) > args.min_chars)
nonmembers = concatenate_datasets([ds["validation"], ds["test"]])
if args.sample_size is not None:
# pick the largest sample_size from the nonmembers
nonmembers = sorted(nonmembers, key=lambda x: len(x['text']), reverse=True)[:args.sample_size]
# pick the largest sample_size from the members
# PICK A RANDOM SAMPLE OF MEMBERS WHOSE LENGTH IS IN THE RANGE OF THE LENGTH OF THE DOCS OF THE NONMEMBERS
doc_lengths = [len(text['text']) for text in nonmembers]
min_len = min(doc_lengths)
max_len = max(doc_lengths)
nonmembers = datasets.Dataset.from_list(nonmembers)
members = ds['train'].filter(lambda x: min_len <= len(x["text"]) and len(x["text"]) <= max_len).shuffle(seed=args.seed).select(range(len(nonmembers)))
nonmembers.save_to_disk(os.path.join(args.output_path, "nonmembers"))
members.save_to_disk(os.path.join(args.output_path, "members"))
# %%
data_points_members = []
for text in tqdm(members['text']):
chunks = create_chunks(text, tokenizer1, args.max_length)
doc_features = []
for chunk in chunks:
mia_features = inference(model1, tokenizer1, chunk)
doc_features.append({'pred': mia_features, 'label': 1})
data_points_members.append(doc_features)
torch.cuda.empty_cache()
with open(os.path.join(args.output_path, "mia_members.jsonl"), "w") as f:
for dp in data_points_members:
f.write(json.dumps(dp) + "\n")
data_points_nonmembers = []
for text in tqdm(nonmembers['text']):
chunks = create_chunks(text, tokenizer1, args.max_length)
doc_features = []
for chunk in chunks:
if len(chunk) > args.min_chars:
mia_features = inference(model1, tokenizer1, chunk)
doc_features.append({'pred': mia_features, 'label': 0})
data_points_nonmembers.append(doc_features)
torch.cuda.empty_cache()
with open(os.path.join(args.output_path, "mia_nonmembers.jsonl"), "w") as f:
for dp in data_points_nonmembers:
f.write(json.dumps(dp) + "\n")