Spaces:
Runtime error
Runtime error
Update app.py
Browse files
app.py
CHANGED
@@ -1,18 +1,939 @@
|
|
1 |
-
import
|
|
|
|
|
|
|
|
|
2 |
import torch
|
3 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
4 |
|
5 |
-
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
6 |
|
7 |
-
tokenizer = AutoTokenizer.from_pretrained("roberta-large-openai-detector")
|
8 |
-
model = AutoModelForSequenceClassification.from_pretrained("roberta-large-openai-detector").to(device)
|
9 |
|
10 |
-
|
|
|
|
|
|
|
11 |
|
12 |
-
|
13 |
-
|
14 |
-
predictions = dict([ (x['label'], x['score']) for x in outputs ])
|
15 |
-
return predictions["LABEL_1"]
|
16 |
|
17 |
-
|
18 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import matplotlib.pyplot as plt
|
2 |
+
import numpy as np
|
3 |
+
import datasets
|
4 |
+
import transformers
|
5 |
+
import re
|
6 |
import torch
|
7 |
+
import torch.nn.functional as F
|
8 |
+
import tqdm
|
9 |
+
import random
|
10 |
+
from sklearn.metrics import roc_curve, precision_recall_curve, auc
|
11 |
+
import argparse
|
12 |
+
import datetime
|
13 |
+
import os
|
14 |
+
import json
|
15 |
+
import functools
|
16 |
+
import custom_datasets
|
17 |
+
from multiprocessing.pool import ThreadPool
|
18 |
+
import time
|
19 |
|
|
|
20 |
|
|
|
|
|
21 |
|
22 |
+
# 15 colorblind-friendly colors
|
23 |
+
COLORS = ["#0072B2", "#009E73", "#D55E00", "#CC79A7", "#F0E442",
|
24 |
+
"#56B4E9", "#E69F00", "#000000", "#0072B2", "#009E73",
|
25 |
+
"#D55E00", "#CC79A7", "#F0E442", "#56B4E9", "#E69F00"]
|
26 |
|
27 |
+
# define regex to match all <extra_id_*> tokens, where * is an integer
|
28 |
+
pattern = re.compile(r"<extra_id_\d+>")
|
|
|
|
|
29 |
|
30 |
+
|
31 |
+
def load_base_model():
|
32 |
+
print('MOVING BASE MODEL TO GPU...', end='', flush=True)
|
33 |
+
start = time.time()
|
34 |
+
try:
|
35 |
+
mask_model.cpu()
|
36 |
+
except NameError:
|
37 |
+
pass
|
38 |
+
if args.openai_model is None:
|
39 |
+
base_model.to(DEVICE)
|
40 |
+
print(f'DONE ({time.time() - start:.2f}s)')
|
41 |
+
|
42 |
+
|
43 |
+
def load_mask_model():
|
44 |
+
print('MOVING MASK MODEL TO GPU...', end='', flush=True)
|
45 |
+
start = time.time()
|
46 |
+
|
47 |
+
if args.openai_model is None:
|
48 |
+
base_model.cpu()
|
49 |
+
if not args.random_fills:
|
50 |
+
mask_model.to(DEVICE)
|
51 |
+
print(f'DONE ({time.time() - start:.2f}s)')
|
52 |
+
|
53 |
+
|
54 |
+
def tokenize_and_mask(text, span_length, pct, ceil_pct=False):
|
55 |
+
tokens = text.split(' ')
|
56 |
+
mask_string = '<<<mask>>>'
|
57 |
+
|
58 |
+
n_spans = pct * len(tokens) / (span_length + args.buffer_size * 2)
|
59 |
+
if ceil_pct:
|
60 |
+
n_spans = np.ceil(n_spans)
|
61 |
+
n_spans = int(n_spans)
|
62 |
+
|
63 |
+
n_masks = 0
|
64 |
+
while n_masks < n_spans:
|
65 |
+
start = np.random.randint(0, len(tokens) - span_length)
|
66 |
+
end = start + span_length
|
67 |
+
search_start = max(0, start - args.buffer_size)
|
68 |
+
search_end = min(len(tokens), end + args.buffer_size)
|
69 |
+
if mask_string not in tokens[search_start:search_end]:
|
70 |
+
tokens[start:end] = [mask_string]
|
71 |
+
n_masks += 1
|
72 |
+
|
73 |
+
# replace each occurrence of mask_string with <extra_id_NUM>, where NUM increments
|
74 |
+
num_filled = 0
|
75 |
+
for idx, token in enumerate(tokens):
|
76 |
+
if token == mask_string:
|
77 |
+
tokens[idx] = f'<extra_id_{num_filled}>'
|
78 |
+
num_filled += 1
|
79 |
+
assert num_filled == n_masks, f"num_filled {num_filled} != n_masks {n_masks}"
|
80 |
+
text = ' '.join(tokens)
|
81 |
+
return text
|
82 |
+
|
83 |
+
|
84 |
+
def count_masks(texts):
|
85 |
+
return [len([x for x in text.split() if x.startswith("<extra_id_")]) for text in texts]
|
86 |
+
|
87 |
+
|
88 |
+
# replace each masked span with a sample from T5 mask_model
|
89 |
+
def replace_masks(texts):
|
90 |
+
n_expected = count_masks(texts)
|
91 |
+
stop_id = mask_tokenizer.encode(f"<extra_id_{max(n_expected)}>")[0]
|
92 |
+
tokens = mask_tokenizer(texts, return_tensors="pt", padding=True).to(DEVICE)
|
93 |
+
outputs = mask_model.generate(**tokens, max_length=150, do_sample=True, top_p=args.mask_top_p, num_return_sequences=1, eos_token_id=stop_id)
|
94 |
+
return mask_tokenizer.batch_decode(outputs, skip_special_tokens=False)
|
95 |
+
|
96 |
+
|
97 |
+
def extract_fills(texts):
|
98 |
+
# remove <pad> from beginning of each text
|
99 |
+
texts = [x.replace("<pad>", "").replace("</s>", "").strip() for x in texts]
|
100 |
+
|
101 |
+
# return the text in between each matched mask token
|
102 |
+
extracted_fills = [pattern.split(x)[1:-1] for x in texts]
|
103 |
+
|
104 |
+
# remove whitespace around each fill
|
105 |
+
extracted_fills = [[y.strip() for y in x] for x in extracted_fills]
|
106 |
+
|
107 |
+
return extracted_fills
|
108 |
+
|
109 |
+
|
110 |
+
def apply_extracted_fills(masked_texts, extracted_fills):
|
111 |
+
# split masked text into tokens, only splitting on spaces (not newlines)
|
112 |
+
tokens = [x.split(' ') for x in masked_texts]
|
113 |
+
|
114 |
+
n_expected = count_masks(masked_texts)
|
115 |
+
|
116 |
+
# replace each mask token with the corresponding fill
|
117 |
+
for idx, (text, fills, n) in enumerate(zip(tokens, extracted_fills, n_expected)):
|
118 |
+
if len(fills) < n:
|
119 |
+
tokens[idx] = []
|
120 |
+
else:
|
121 |
+
for fill_idx in range(n):
|
122 |
+
text[text.index(f"<extra_id_{fill_idx}>")] = fills[fill_idx]
|
123 |
+
|
124 |
+
# join tokens back into text
|
125 |
+
texts = [" ".join(x) for x in tokens]
|
126 |
+
return texts
|
127 |
+
|
128 |
+
|
129 |
+
def perturb_texts_(texts, span_length, pct, ceil_pct=False):
|
130 |
+
if not args.random_fills:
|
131 |
+
masked_texts = [tokenize_and_mask(x, span_length, pct, ceil_pct) for x in texts]
|
132 |
+
raw_fills = replace_masks(masked_texts)
|
133 |
+
extracted_fills = extract_fills(raw_fills)
|
134 |
+
perturbed_texts = apply_extracted_fills(masked_texts, extracted_fills)
|
135 |
+
|
136 |
+
# Handle the fact that sometimes the model doesn't generate the right number of fills and we have to try again
|
137 |
+
attempts = 1
|
138 |
+
while '' in perturbed_texts:
|
139 |
+
idxs = [idx for idx, x in enumerate(perturbed_texts) if x == '']
|
140 |
+
print(f'WARNING: {len(idxs)} texts have no fills. Trying again [attempt {attempts}].')
|
141 |
+
masked_texts = [tokenize_and_mask(x, span_length, pct, ceil_pct) for idx, x in enumerate(texts) if idx in idxs]
|
142 |
+
raw_fills = replace_masks(masked_texts)
|
143 |
+
extracted_fills = extract_fills(raw_fills)
|
144 |
+
new_perturbed_texts = apply_extracted_fills(masked_texts, extracted_fills)
|
145 |
+
for idx, x in zip(idxs, new_perturbed_texts):
|
146 |
+
perturbed_texts[idx] = x
|
147 |
+
attempts += 1
|
148 |
+
else:
|
149 |
+
if args.random_fills_tokens:
|
150 |
+
# tokenize base_tokenizer
|
151 |
+
tokens = base_tokenizer(texts, return_tensors="pt", padding=True).to(DEVICE)
|
152 |
+
valid_tokens = tokens.input_ids != base_tokenizer.pad_token_id
|
153 |
+
replace_pct = args.pct_words_masked * (args.span_length / (args.span_length + 2 * args.buffer_size))
|
154 |
+
|
155 |
+
# replace replace_pct of input_ids with random tokens
|
156 |
+
random_mask = torch.rand(tokens.input_ids.shape, device=DEVICE) < replace_pct
|
157 |
+
random_mask &= valid_tokens
|
158 |
+
random_tokens = torch.randint(0, base_tokenizer.vocab_size, (random_mask.sum(),), device=DEVICE)
|
159 |
+
# while any of the random tokens are special tokens, replace them with random non-special tokens
|
160 |
+
while any(base_tokenizer.decode(x) in base_tokenizer.all_special_tokens for x in random_tokens):
|
161 |
+
random_tokens = torch.randint(0, base_tokenizer.vocab_size, (random_mask.sum(),), device=DEVICE)
|
162 |
+
tokens.input_ids[random_mask] = random_tokens
|
163 |
+
perturbed_texts = base_tokenizer.batch_decode(tokens.input_ids, skip_special_tokens=True)
|
164 |
+
else:
|
165 |
+
masked_texts = [tokenize_and_mask(x, span_length, pct, ceil_pct) for x in texts]
|
166 |
+
perturbed_texts = masked_texts
|
167 |
+
# replace each <extra_id_*> with args.span_length random words from FILL_DICTIONARY
|
168 |
+
for idx, text in enumerate(perturbed_texts):
|
169 |
+
filled_text = text
|
170 |
+
for fill_idx in range(count_masks([text])[0]):
|
171 |
+
fill = random.sample(FILL_DICTIONARY, span_length)
|
172 |
+
filled_text = filled_text.replace(f"<extra_id_{fill_idx}>", " ".join(fill))
|
173 |
+
assert count_masks([filled_text])[0] == 0, "Failed to replace all masks"
|
174 |
+
perturbed_texts[idx] = filled_text
|
175 |
+
|
176 |
+
return perturbed_texts
|
177 |
+
|
178 |
+
|
179 |
+
def perturb_texts(texts, span_length, pct, ceil_pct=False):
|
180 |
+
chunk_size = args.chunk_size
|
181 |
+
if '11b' in mask_filling_model_name:
|
182 |
+
chunk_size //= 2
|
183 |
+
|
184 |
+
outputs = []
|
185 |
+
for i in tqdm.tqdm(range(0, len(texts), chunk_size), desc="Applying perturbations"):
|
186 |
+
outputs.extend(perturb_texts_(texts[i:i + chunk_size], span_length, pct, ceil_pct=ceil_pct))
|
187 |
+
return outputs
|
188 |
+
|
189 |
+
|
190 |
+
def drop_last_word(text):
|
191 |
+
return ' '.join(text.split(' ')[:-1])
|
192 |
+
|
193 |
+
|
194 |
+
def _openai_sample(p):
|
195 |
+
if args.dataset != 'pubmed': # keep Answer: prefix for pubmed
|
196 |
+
p = drop_last_word(p)
|
197 |
+
|
198 |
+
# sample from the openai model
|
199 |
+
kwargs = { "engine": args.openai_model, "max_tokens": 200 }
|
200 |
+
if args.do_top_p:
|
201 |
+
kwargs['top_p'] = args.top_p
|
202 |
+
|
203 |
+
r = openai.Completion.create(prompt=f"{p}", **kwargs)
|
204 |
+
return p + r['choices'][0].text
|
205 |
+
|
206 |
+
|
207 |
+
# sample from base_model using ****only**** the first 30 tokens in each example as context
|
208 |
+
def sample_from_model(texts, min_words=55, prompt_tokens=30):
|
209 |
+
# encode each text as a list of token ids
|
210 |
+
if args.dataset == 'pubmed':
|
211 |
+
texts = [t[:t.index(custom_datasets.SEPARATOR)] for t in texts]
|
212 |
+
all_encoded = base_tokenizer(texts, return_tensors="pt", padding=True).to(DEVICE)
|
213 |
+
else:
|
214 |
+
all_encoded = base_tokenizer(texts, return_tensors="pt", padding=True).to(DEVICE)
|
215 |
+
all_encoded = {key: value[:, :prompt_tokens] for key, value in all_encoded.items()}
|
216 |
+
|
217 |
+
if args.openai_model:
|
218 |
+
# decode the prefixes back into text
|
219 |
+
prefixes = base_tokenizer.batch_decode(all_encoded['input_ids'], skip_special_tokens=True)
|
220 |
+
pool = ThreadPool(args.batch_size)
|
221 |
+
|
222 |
+
decoded = pool.map(_openai_sample, prefixes)
|
223 |
+
else:
|
224 |
+
decoded = ['' for _ in range(len(texts))]
|
225 |
+
|
226 |
+
# sample from the model until we get a sample with at least min_words words for each example
|
227 |
+
# this is an inefficient way to do this (since we regenerate for all inputs if just one is too short), but it works
|
228 |
+
tries = 0
|
229 |
+
while (m := min(len(x.split()) for x in decoded)) < min_words:
|
230 |
+
if tries != 0:
|
231 |
+
print()
|
232 |
+
print(f"min words: {m}, needed {min_words}, regenerating (try {tries})")
|
233 |
+
|
234 |
+
sampling_kwargs = {}
|
235 |
+
if args.do_top_p:
|
236 |
+
sampling_kwargs['top_p'] = args.top_p
|
237 |
+
elif args.do_top_k:
|
238 |
+
sampling_kwargs['top_k'] = args.top_k
|
239 |
+
min_length = 50 if args.dataset in ['pubmed'] else 150
|
240 |
+
outputs = base_model.generate(**all_encoded, min_length=min_length, max_length=200, do_sample=True, **sampling_kwargs, pad_token_id=base_tokenizer.eos_token_id, eos_token_id=base_tokenizer.eos_token_id)
|
241 |
+
decoded = base_tokenizer.batch_decode(outputs, skip_special_tokens=True)
|
242 |
+
tries += 1
|
243 |
+
|
244 |
+
if args.openai_model:
|
245 |
+
global API_TOKEN_COUNTER
|
246 |
+
|
247 |
+
# count total number of tokens with GPT2_TOKENIZER
|
248 |
+
total_tokens = sum(len(GPT2_TOKENIZER.encode(x)) for x in decoded)
|
249 |
+
API_TOKEN_COUNTER += total_tokens
|
250 |
+
|
251 |
+
return decoded
|
252 |
+
|
253 |
+
|
254 |
+
def get_likelihood(logits, labels):
|
255 |
+
assert logits.shape[0] == 1
|
256 |
+
assert labels.shape[0] == 1
|
257 |
+
|
258 |
+
logits = logits.view(-1, logits.shape[-1])[:-1]
|
259 |
+
labels = labels.view(-1)[1:]
|
260 |
+
log_probs = torch.nn.functional.log_softmax(logits, dim=-1)
|
261 |
+
log_likelihood = log_probs.gather(dim=-1, index=labels.unsqueeze(-1)).squeeze(-1)
|
262 |
+
return log_likelihood.mean()
|
263 |
+
|
264 |
+
|
265 |
+
# Get the log likelihood of each text under the base_model
|
266 |
+
def get_ll(text):
|
267 |
+
if args.openai_model:
|
268 |
+
kwargs = { "engine": args.openai_model, "temperature": 0, "max_tokens": 0, "echo": True, "logprobs": 0}
|
269 |
+
r = openai.Completion.create(prompt=f"<|endoftext|>{text}", **kwargs)
|
270 |
+
result = r['choices'][0]
|
271 |
+
tokens, logprobs = result["logprobs"]["tokens"][1:], result["logprobs"]["token_logprobs"][1:]
|
272 |
+
|
273 |
+
assert len(tokens) == len(logprobs), f"Expected {len(tokens)} logprobs, got {len(logprobs)}"
|
274 |
+
|
275 |
+
return np.mean(logprobs)
|
276 |
+
else:
|
277 |
+
with torch.no_grad():
|
278 |
+
tokenized = base_tokenizer(text, return_tensors="pt").to(DEVICE)
|
279 |
+
labels = tokenized.input_ids
|
280 |
+
return -base_model(**tokenized, labels=labels).loss.item()
|
281 |
+
|
282 |
+
|
283 |
+
def get_lls(texts):
|
284 |
+
if not args.openai_model:
|
285 |
+
return [get_ll(text) for text in texts]
|
286 |
+
else:
|
287 |
+
global API_TOKEN_COUNTER
|
288 |
+
|
289 |
+
# use GPT2_TOKENIZER to get total number of tokens
|
290 |
+
total_tokens = sum(len(GPT2_TOKENIZER.encode(text)) for text in texts)
|
291 |
+
API_TOKEN_COUNTER += total_tokens * 2 # multiply by two because OpenAI double-counts echo_prompt tokens
|
292 |
+
|
293 |
+
pool = ThreadPool(args.batch_size)
|
294 |
+
return pool.map(get_ll, texts)
|
295 |
+
|
296 |
+
|
297 |
+
# get the average rank of each observed token sorted by model likelihood
|
298 |
+
def get_rank(text, log=False):
|
299 |
+
assert args.openai_model is None, "get_rank not implemented for OpenAI models"
|
300 |
+
|
301 |
+
with torch.no_grad():
|
302 |
+
tokenized = base_tokenizer(text, return_tensors="pt").to(DEVICE)
|
303 |
+
logits = base_model(**tokenized).logits[:,:-1]
|
304 |
+
labels = tokenized.input_ids[:,1:]
|
305 |
+
|
306 |
+
# get rank of each label token in the model's likelihood ordering
|
307 |
+
matches = (logits.argsort(-1, descending=True) == labels.unsqueeze(-1)).nonzero()
|
308 |
+
|
309 |
+
assert matches.shape[1] == 3, f"Expected 3 dimensions in matches tensor, got {matches.shape}"
|
310 |
+
|
311 |
+
ranks, timesteps = matches[:,-1], matches[:,-2]
|
312 |
+
|
313 |
+
# make sure we got exactly one match for each timestep in the sequence
|
314 |
+
assert (timesteps == torch.arange(len(timesteps)).to(timesteps.device)).all(), "Expected one match per timestep"
|
315 |
+
|
316 |
+
ranks = ranks.float() + 1 # convert to 1-indexed rank
|
317 |
+
if log:
|
318 |
+
ranks = torch.log(ranks)
|
319 |
+
|
320 |
+
return ranks.float().mean().item()
|
321 |
+
|
322 |
+
|
323 |
+
# get average entropy of each token in the text
|
324 |
+
def get_entropy(text):
|
325 |
+
assert args.openai_model is None, "get_entropy not implemented for OpenAI models"
|
326 |
+
|
327 |
+
with torch.no_grad():
|
328 |
+
tokenized = base_tokenizer(text, return_tensors="pt").to(DEVICE)
|
329 |
+
logits = base_model(**tokenized).logits[:,:-1]
|
330 |
+
neg_entropy = F.softmax(logits, dim=-1) * F.log_softmax(logits, dim=-1)
|
331 |
+
return -neg_entropy.sum(-1).mean().item()
|
332 |
+
|
333 |
+
|
334 |
+
def get_roc_metrics(real_preds, sample_preds):
|
335 |
+
fpr, tpr, _ = roc_curve([0] * len(real_preds) + [1] * len(sample_preds), real_preds + sample_preds)
|
336 |
+
roc_auc = auc(fpr, tpr)
|
337 |
+
return fpr.tolist(), tpr.tolist(), float(roc_auc)
|
338 |
+
|
339 |
+
|
340 |
+
def get_precision_recall_metrics(real_preds, sample_preds):
|
341 |
+
precision, recall, _ = precision_recall_curve([0] * len(real_preds) + [1] * len(sample_preds), real_preds + sample_preds)
|
342 |
+
pr_auc = auc(recall, precision)
|
343 |
+
return precision.tolist(), recall.tolist(), float(pr_auc)
|
344 |
+
|
345 |
+
|
346 |
+
# save the ROC curve for each experiment, given a list of output dictionaries, one for each experiment, using colorblind-friendly colors
|
347 |
+
def save_roc_curves(experiments):
|
348 |
+
# first, clear plt
|
349 |
+
plt.clf()
|
350 |
+
|
351 |
+
for experiment, color in zip(experiments, COLORS):
|
352 |
+
metrics = experiment["metrics"]
|
353 |
+
plt.plot(metrics["fpr"], metrics["tpr"], label=f"{experiment['name']}, roc_auc={metrics['roc_auc']:.3f}", color=color)
|
354 |
+
# print roc_auc for this experiment
|
355 |
+
print(f"{experiment['name']} roc_auc: {metrics['roc_auc']:.3f}")
|
356 |
+
plt.plot([0, 1], [0, 1], color='black', lw=2, linestyle='--')
|
357 |
+
plt.xlim([0.0, 1.0])
|
358 |
+
plt.ylim([0.0, 1.05])
|
359 |
+
plt.xlabel('False Positive Rate')
|
360 |
+
plt.ylabel('True Positive Rate')
|
361 |
+
plt.title(f'ROC Curves ({base_model_name} - {args.mask_filling_model_name})')
|
362 |
+
plt.legend(loc="lower right", fontsize=6)
|
363 |
+
plt.savefig(f"{SAVE_FOLDER}/roc_curves.png")
|
364 |
+
|
365 |
+
|
366 |
+
# save the histogram of log likelihoods in two side-by-side plots, one for real and real perturbed, and one for sampled and sampled perturbed
|
367 |
+
def save_ll_histograms(experiments):
|
368 |
+
# first, clear plt
|
369 |
+
plt.clf()
|
370 |
+
|
371 |
+
for experiment in experiments:
|
372 |
+
try:
|
373 |
+
results = experiment["raw_results"]
|
374 |
+
# plot histogram of sampled/perturbed sampled on left, original/perturbed original on right
|
375 |
+
plt.figure(figsize=(20, 6))
|
376 |
+
plt.subplot(1, 2, 1)
|
377 |
+
plt.hist([r["sampled_ll"] for r in results], alpha=0.5, bins='auto', label='sampled')
|
378 |
+
plt.hist([r["perturbed_sampled_ll"] for r in results], alpha=0.5, bins='auto', label='perturbed sampled')
|
379 |
+
plt.xlabel("log likelihood")
|
380 |
+
plt.ylabel('count')
|
381 |
+
plt.legend(loc='upper right')
|
382 |
+
plt.subplot(1, 2, 2)
|
383 |
+
plt.hist([r["original_ll"] for r in results], alpha=0.5, bins='auto', label='original')
|
384 |
+
plt.hist([r["perturbed_original_ll"] for r in results], alpha=0.5, bins='auto', label='perturbed original')
|
385 |
+
plt.xlabel("log likelihood")
|
386 |
+
plt.ylabel('count')
|
387 |
+
plt.legend(loc='upper right')
|
388 |
+
plt.savefig(f"{SAVE_FOLDER}/ll_histograms_{experiment['name']}.png")
|
389 |
+
except:
|
390 |
+
pass
|
391 |
+
|
392 |
+
|
393 |
+
# save the histograms of log likelihood ratios in two side-by-side plots, one for real and real perturbed, and one for sampled and sampled perturbed
|
394 |
+
def save_llr_histograms(experiments):
|
395 |
+
# first, clear plt
|
396 |
+
plt.clf()
|
397 |
+
|
398 |
+
for experiment in experiments:
|
399 |
+
try:
|
400 |
+
results = experiment["raw_results"]
|
401 |
+
# plot histogram of sampled/perturbed sampled on left, original/perturbed original on right
|
402 |
+
plt.figure(figsize=(20, 6))
|
403 |
+
plt.subplot(1, 2, 1)
|
404 |
+
|
405 |
+
# compute the log likelihood ratio for each result
|
406 |
+
for r in results:
|
407 |
+
r["sampled_llr"] = r["sampled_ll"] - r["perturbed_sampled_ll"]
|
408 |
+
r["original_llr"] = r["original_ll"] - r["perturbed_original_ll"]
|
409 |
+
|
410 |
+
plt.hist([r["sampled_llr"] for r in results], alpha=0.5, bins='auto', label='sampled')
|
411 |
+
plt.hist([r["original_llr"] for r in results], alpha=0.5, bins='auto', label='original')
|
412 |
+
plt.xlabel("log likelihood ratio")
|
413 |
+
plt.ylabel('count')
|
414 |
+
plt.legend(loc='upper right')
|
415 |
+
plt.savefig(f"{SAVE_FOLDER}/llr_histograms_{experiment['name']}.png")
|
416 |
+
except:
|
417 |
+
pass
|
418 |
+
|
419 |
+
|
420 |
+
def get_perturbation_results(span_length=10, n_perturbations=1, n_samples=500):
|
421 |
+
load_mask_model()
|
422 |
+
|
423 |
+
torch.manual_seed(0)
|
424 |
+
np.random.seed(0)
|
425 |
+
|
426 |
+
results = []
|
427 |
+
original_text = data["original"]
|
428 |
+
sampled_text = data["sampled"]
|
429 |
+
|
430 |
+
perturb_fn = functools.partial(perturb_texts, span_length=span_length, pct=args.pct_words_masked)
|
431 |
+
|
432 |
+
p_sampled_text = perturb_fn([x for x in sampled_text for _ in range(n_perturbations)])
|
433 |
+
p_original_text = perturb_fn([x for x in original_text for _ in range(n_perturbations)])
|
434 |
+
for _ in range(n_perturbation_rounds - 1):
|
435 |
+
try:
|
436 |
+
p_sampled_text, p_original_text = perturb_fn(p_sampled_text), perturb_fn(p_original_text)
|
437 |
+
except AssertionError:
|
438 |
+
break
|
439 |
+
|
440 |
+
assert len(p_sampled_text) == len(sampled_text) * n_perturbations, f"Expected {len(sampled_text) * n_perturbations} perturbed samples, got {len(p_sampled_text)}"
|
441 |
+
assert len(p_original_text) == len(original_text) * n_perturbations, f"Expected {len(original_text) * n_perturbations} perturbed samples, got {len(p_original_text)}"
|
442 |
+
|
443 |
+
for idx in range(len(original_text)):
|
444 |
+
results.append({
|
445 |
+
"original": original_text[idx],
|
446 |
+
"sampled": sampled_text[idx],
|
447 |
+
"perturbed_sampled": p_sampled_text[idx * n_perturbations: (idx + 1) * n_perturbations],
|
448 |
+
"perturbed_original": p_original_text[idx * n_perturbations: (idx + 1) * n_perturbations]
|
449 |
+
})
|
450 |
+
|
451 |
+
load_base_model()
|
452 |
+
|
453 |
+
for res in tqdm.tqdm(results, desc="Computing log likelihoods"):
|
454 |
+
p_sampled_ll = get_lls(res["perturbed_sampled"])
|
455 |
+
p_original_ll = get_lls(res["perturbed_original"])
|
456 |
+
res["original_ll"] = get_ll(res["original"])
|
457 |
+
res["sampled_ll"] = get_ll(res["sampled"])
|
458 |
+
res["all_perturbed_sampled_ll"] = p_sampled_ll
|
459 |
+
res["all_perturbed_original_ll"] = p_original_ll
|
460 |
+
res["perturbed_sampled_ll"] = np.mean(p_sampled_ll)
|
461 |
+
res["perturbed_original_ll"] = np.mean(p_original_ll)
|
462 |
+
res["perturbed_sampled_ll_std"] = np.std(p_sampled_ll) if len(p_sampled_ll) > 1 else 1
|
463 |
+
res["perturbed_original_ll_std"] = np.std(p_original_ll) if len(p_original_ll) > 1 else 1
|
464 |
+
|
465 |
+
return results
|
466 |
+
|
467 |
+
|
468 |
+
def run_perturbation_experiment(results, criterion, span_length=10, n_perturbations=1, n_samples=500):
|
469 |
+
# compute diffs with perturbed
|
470 |
+
predictions = {'real': [], 'samples': []}
|
471 |
+
for res in results:
|
472 |
+
if criterion == 'd':
|
473 |
+
predictions['real'].append(res['original_ll'] - res['perturbed_original_ll'])
|
474 |
+
predictions['samples'].append(res['sampled_ll'] - res['perturbed_sampled_ll'])
|
475 |
+
elif criterion == 'z':
|
476 |
+
if res['perturbed_original_ll_std'] == 0:
|
477 |
+
res['perturbed_original_ll_std'] = 1
|
478 |
+
print("WARNING: std of perturbed original is 0, setting to 1")
|
479 |
+
print(f"Number of unique perturbed original texts: {len(set(res['perturbed_original']))}")
|
480 |
+
print(f"Original text: {res['original']}")
|
481 |
+
if res['perturbed_sampled_ll_std'] == 0:
|
482 |
+
res['perturbed_sampled_ll_std'] = 1
|
483 |
+
print("WARNING: std of perturbed sampled is 0, setting to 1")
|
484 |
+
print(f"Number of unique perturbed sampled texts: {len(set(res['perturbed_sampled']))}")
|
485 |
+
print(f"Sampled text: {res['sampled']}")
|
486 |
+
predictions['real'].append((res['original_ll'] - res['perturbed_original_ll']) / res['perturbed_original_ll_std'])
|
487 |
+
predictions['samples'].append((res['sampled_ll'] - res['perturbed_sampled_ll']) / res['perturbed_sampled_ll_std'])
|
488 |
+
|
489 |
+
fpr, tpr, roc_auc = get_roc_metrics(predictions['real'], predictions['samples'])
|
490 |
+
p, r, pr_auc = get_precision_recall_metrics(predictions['real'], predictions['samples'])
|
491 |
+
name = f'perturbation_{n_perturbations}_{criterion}'
|
492 |
+
print(f"{name} ROC AUC: {roc_auc}, PR AUC: {pr_auc}")
|
493 |
+
return {
|
494 |
+
'name': name,
|
495 |
+
'predictions': predictions,
|
496 |
+
'info': {
|
497 |
+
'pct_words_masked': args.pct_words_masked,
|
498 |
+
'span_length': span_length,
|
499 |
+
'n_perturbations': n_perturbations,
|
500 |
+
'n_samples': n_samples,
|
501 |
+
},
|
502 |
+
'raw_results': results,
|
503 |
+
'metrics': {
|
504 |
+
'roc_auc': roc_auc,
|
505 |
+
'fpr': fpr,
|
506 |
+
'tpr': tpr,
|
507 |
+
},
|
508 |
+
'pr_metrics': {
|
509 |
+
'pr_auc': pr_auc,
|
510 |
+
'precision': p,
|
511 |
+
'recall': r,
|
512 |
+
},
|
513 |
+
'loss': 1 - pr_auc,
|
514 |
+
}
|
515 |
+
|
516 |
+
|
517 |
+
def run_baseline_threshold_experiment(criterion_fn, name, n_samples=500):
|
518 |
+
torch.manual_seed(0)
|
519 |
+
np.random.seed(0)
|
520 |
+
|
521 |
+
results = []
|
522 |
+
for batch in tqdm.tqdm(range(n_samples // batch_size), desc=f"Computing {name} criterion"):
|
523 |
+
original_text = data["original"][batch * batch_size:(batch + 1) * batch_size]
|
524 |
+
sampled_text = data["sampled"][batch * batch_size:(batch + 1) * batch_size]
|
525 |
+
|
526 |
+
for idx in range(len(original_text)):
|
527 |
+
results.append({
|
528 |
+
"original": original_text[idx],
|
529 |
+
"original_crit": criterion_fn(original_text[idx]),
|
530 |
+
"sampled": sampled_text[idx],
|
531 |
+
"sampled_crit": criterion_fn(sampled_text[idx]),
|
532 |
+
})
|
533 |
+
|
534 |
+
# compute prediction scores for real/sampled passages
|
535 |
+
predictions = {
|
536 |
+
'real': [x["original_crit"] for x in results],
|
537 |
+
'samples': [x["sampled_crit"] for x in results],
|
538 |
+
}
|
539 |
+
|
540 |
+
fpr, tpr, roc_auc = get_roc_metrics(predictions['real'], predictions['samples'])
|
541 |
+
p, r, pr_auc = get_precision_recall_metrics(predictions['real'], predictions['samples'])
|
542 |
+
print(f"{name}_threshold ROC AUC: {roc_auc}, PR AUC: {pr_auc}")
|
543 |
+
return {
|
544 |
+
'name': f'{name}_threshold',
|
545 |
+
'predictions': predictions,
|
546 |
+
'info': {
|
547 |
+
'n_samples': n_samples,
|
548 |
+
},
|
549 |
+
'raw_results': results,
|
550 |
+
'metrics': {
|
551 |
+
'roc_auc': roc_auc,
|
552 |
+
'fpr': fpr,
|
553 |
+
'tpr': tpr,
|
554 |
+
},
|
555 |
+
'pr_metrics': {
|
556 |
+
'pr_auc': pr_auc,
|
557 |
+
'precision': p,
|
558 |
+
'recall': r,
|
559 |
+
},
|
560 |
+
'loss': 1 - pr_auc,
|
561 |
+
}
|
562 |
+
|
563 |
+
|
564 |
+
# strip newlines from each example; replace one or more newlines with a single space
|
565 |
+
def strip_newlines(text):
|
566 |
+
return ' '.join(text.split())
|
567 |
+
|
568 |
+
|
569 |
+
# trim to shorter length
|
570 |
+
def trim_to_shorter_length(texta, textb):
|
571 |
+
# truncate to shorter of o and s
|
572 |
+
shorter_length = min(len(texta.split(' ')), len(textb.split(' ')))
|
573 |
+
texta = ' '.join(texta.split(' ')[:shorter_length])
|
574 |
+
textb = ' '.join(textb.split(' ')[:shorter_length])
|
575 |
+
return texta, textb
|
576 |
+
|
577 |
+
|
578 |
+
def truncate_to_substring(text, substring, idx_occurrence):
|
579 |
+
# truncate everything after the idx_occurrence occurrence of substring
|
580 |
+
assert idx_occurrence > 0, 'idx_occurrence must be > 0'
|
581 |
+
idx = -1
|
582 |
+
for _ in range(idx_occurrence):
|
583 |
+
idx = text.find(substring, idx + 1)
|
584 |
+
if idx == -1:
|
585 |
+
return text
|
586 |
+
return text[:idx]
|
587 |
+
|
588 |
+
|
589 |
+
def generate_samples(raw_data, batch_size):
|
590 |
+
torch.manual_seed(42)
|
591 |
+
np.random.seed(42)
|
592 |
+
data = {
|
593 |
+
"original": [],
|
594 |
+
"sampled": [],
|
595 |
+
}
|
596 |
+
|
597 |
+
for batch in range(len(raw_data) // batch_size):
|
598 |
+
print('Generating samples for batch', batch, 'of', len(raw_data) // batch_size)
|
599 |
+
original_text = raw_data[batch * batch_size:(batch + 1) * batch_size]
|
600 |
+
sampled_text = sample_from_model(original_text, min_words=30 if args.dataset in ['pubmed'] else 55)
|
601 |
+
|
602 |
+
for o, s in zip(original_text, sampled_text):
|
603 |
+
if args.dataset == 'pubmed':
|
604 |
+
s = truncate_to_substring(s, 'Question:', 2)
|
605 |
+
o = o.replace(custom_datasets.SEPARATOR, ' ')
|
606 |
+
|
607 |
+
o, s = trim_to_shorter_length(o, s)
|
608 |
+
|
609 |
+
# add to the data
|
610 |
+
data["original"].append(o)
|
611 |
+
data["sampled"].append(s)
|
612 |
+
|
613 |
+
if args.pre_perturb_pct > 0:
|
614 |
+
print(f'APPLYING {args.pre_perturb_pct}, {args.pre_perturb_span_length} PRE-PERTURBATIONS')
|
615 |
+
load_mask_model()
|
616 |
+
data["sampled"] = perturb_texts(data["sampled"], args.pre_perturb_span_length, args.pre_perturb_pct, ceil_pct=True)
|
617 |
+
load_base_model()
|
618 |
+
|
619 |
+
return data
|
620 |
+
|
621 |
+
|
622 |
+
def generate_data(dataset, key):
|
623 |
+
# load data
|
624 |
+
if dataset in custom_datasets.DATASETS:
|
625 |
+
data = custom_datasets.load(dataset, cache_dir)
|
626 |
+
else:
|
627 |
+
data = datasets.load_dataset(dataset, split='train', cache_dir=cache_dir)[key]
|
628 |
+
|
629 |
+
# get unique examples, strip whitespace, and remove newlines
|
630 |
+
# then take just the long examples, shuffle, take the first 5,000 to tokenize to save time
|
631 |
+
# then take just the examples that are <= 512 tokens (for the mask model)
|
632 |
+
# then generate n_samples samples
|
633 |
+
|
634 |
+
# remove duplicates from the data
|
635 |
+
data = list(dict.fromkeys(data)) # deterministic, as opposed to set()
|
636 |
+
|
637 |
+
# strip whitespace around each example
|
638 |
+
data = [x.strip() for x in data]
|
639 |
+
|
640 |
+
# remove newlines from each example
|
641 |
+
data = [strip_newlines(x) for x in data]
|
642 |
+
|
643 |
+
# try to keep only examples with > 250 words
|
644 |
+
if dataset in ['writing', 'squad', 'xsum']:
|
645 |
+
long_data = [x for x in data if len(x.split()) > 250]
|
646 |
+
if len(long_data) > 0:
|
647 |
+
data = long_data
|
648 |
+
|
649 |
+
random.seed(0)
|
650 |
+
random.shuffle(data)
|
651 |
+
|
652 |
+
data = data[:5_000]
|
653 |
+
|
654 |
+
# keep only examples with <= 512 tokens according to mask_tokenizer
|
655 |
+
# this step has the extra effect of removing examples with low-quality/garbage content
|
656 |
+
tokenized_data = preproc_tokenizer(data)
|
657 |
+
data = [x for x, y in zip(data, tokenized_data["input_ids"]) if len(y) <= 512]
|
658 |
+
|
659 |
+
# print stats about remainining data
|
660 |
+
print(f"Total number of samples: {len(data)}")
|
661 |
+
print(f"Average number of words: {np.mean([len(x.split()) for x in data])}")
|
662 |
+
|
663 |
+
return generate_samples(data[:n_samples], batch_size=batch_size)
|
664 |
+
|
665 |
+
|
666 |
+
def load_base_model_and_tokenizer(name):
|
667 |
+
if args.openai_model is None:
|
668 |
+
print(f'Loading BASE model {args.base_model_name}...')
|
669 |
+
base_model_kwargs = {}
|
670 |
+
if 'gpt-j' in name or 'neox' in name:
|
671 |
+
base_model_kwargs.update(dict(torch_dtype=torch.float16))
|
672 |
+
if 'gpt-j' in name:
|
673 |
+
base_model_kwargs.update(dict(revision='float16'))
|
674 |
+
base_model = transformers.AutoModelForCausalLM.from_pretrained(name, **base_model_kwargs, cache_dir=cache_dir)
|
675 |
+
else:
|
676 |
+
base_model = None
|
677 |
+
|
678 |
+
optional_tok_kwargs = {}
|
679 |
+
if "facebook/opt-" in name:
|
680 |
+
print("Using non-fast tokenizer for OPT")
|
681 |
+
optional_tok_kwargs['fast'] = False
|
682 |
+
if args.dataset in ['pubmed']:
|
683 |
+
optional_tok_kwargs['padding_side'] = 'left'
|
684 |
+
base_tokenizer = transformers.AutoTokenizer.from_pretrained(name, **optional_tok_kwargs, cache_dir=cache_dir)
|
685 |
+
base_tokenizer.pad_token_id = base_tokenizer.eos_token_id
|
686 |
+
|
687 |
+
return base_model, base_tokenizer
|
688 |
+
|
689 |
+
|
690 |
+
def eval_supervised(data, model):
|
691 |
+
print(f'Beginning supervised evaluation with {model}...')
|
692 |
+
detector = transformers.AutoModelForSequenceClassification.from_pretrained(model, cache_dir=cache_dir).to(DEVICE)
|
693 |
+
tokenizer = transformers.AutoTokenizer.from_pretrained(model, cache_dir=cache_dir)
|
694 |
+
|
695 |
+
real, fake = data['original'], data['sampled']
|
696 |
+
|
697 |
+
with torch.no_grad():
|
698 |
+
# get predictions for real
|
699 |
+
real_preds = []
|
700 |
+
for batch in tqdm.tqdm(range(len(real) // batch_size), desc="Evaluating real"):
|
701 |
+
batch_real = real[batch * batch_size:(batch + 1) * batch_size]
|
702 |
+
batch_real = tokenizer(batch_real, padding=True, truncation=True, max_length=512, return_tensors="pt").to(DEVICE)
|
703 |
+
real_preds.extend(detector(**batch_real).logits.softmax(-1)[:,0].tolist())
|
704 |
+
|
705 |
+
# get predictions for fake
|
706 |
+
fake_preds = []
|
707 |
+
for batch in tqdm.tqdm(range(len(fake) // batch_size), desc="Evaluating fake"):
|
708 |
+
batch_fake = fake[batch * batch_size:(batch + 1) * batch_size]
|
709 |
+
batch_fake = tokenizer(batch_fake, padding=True, truncation=True, max_length=512, return_tensors="pt").to(DEVICE)
|
710 |
+
fake_preds.extend(detector(**batch_fake).logits.softmax(-1)[:,0].tolist())
|
711 |
+
|
712 |
+
predictions = {
|
713 |
+
'real': real_preds,
|
714 |
+
'samples': fake_preds,
|
715 |
+
}
|
716 |
+
|
717 |
+
fpr, tpr, roc_auc = get_roc_metrics(real_preds, fake_preds)
|
718 |
+
p, r, pr_auc = get_precision_recall_metrics(real_preds, fake_preds)
|
719 |
+
print(f"{model} ROC AUC: {roc_auc}, PR AUC: {pr_auc}")
|
720 |
+
|
721 |
+
# free GPU memory
|
722 |
+
del detector
|
723 |
+
torch.cuda.empty_cache()
|
724 |
+
|
725 |
+
return {
|
726 |
+
'name': model,
|
727 |
+
'predictions': predictions,
|
728 |
+
'info': {
|
729 |
+
'n_samples': n_samples,
|
730 |
+
},
|
731 |
+
'metrics': {
|
732 |
+
'roc_auc': roc_auc,
|
733 |
+
'fpr': fpr,
|
734 |
+
'tpr': tpr,
|
735 |
+
},
|
736 |
+
'pr_metrics': {
|
737 |
+
'pr_auc': pr_auc,
|
738 |
+
'precision': p,
|
739 |
+
'recall': r,
|
740 |
+
},
|
741 |
+
'loss': 1 - pr_auc,
|
742 |
+
}
|
743 |
+
|
744 |
+
|
745 |
+
if __name__ == '__main__':
|
746 |
+
DEVICE = "cuda"
|
747 |
+
|
748 |
+
parser = argparse.ArgumentParser()
|
749 |
+
parser.add_argument('--dataset', type=str, default="xsum")
|
750 |
+
parser.add_argument('--dataset_key', type=str, default="document")
|
751 |
+
parser.add_argument('--pct_words_masked', type=float, default=0.3) # pct masked is actually pct_words_masked * (span_length / (span_length + 2 * buffer_size))
|
752 |
+
parser.add_argument('--span_length', type=int, default=2)
|
753 |
+
parser.add_argument('--n_samples', type=int, default=200)
|
754 |
+
parser.add_argument('--n_perturbation_list', type=str, default="1,10")
|
755 |
+
parser.add_argument('--n_perturbation_rounds', type=int, default=1)
|
756 |
+
parser.add_argument('--base_model_name', type=str, default="gpt2-medium")
|
757 |
+
parser.add_argument('--scoring_model_name', type=str, default="")
|
758 |
+
parser.add_argument('--mask_filling_model_name', type=str, default="t5-large")
|
759 |
+
parser.add_argument('--batch_size', type=int, default=50)
|
760 |
+
parser.add_argument('--chunk_size', type=int, default=20)
|
761 |
+
parser.add_argument('--n_similarity_samples', type=int, default=20)
|
762 |
+
parser.add_argument('--int8', action='store_true')
|
763 |
+
parser.add_argument('--half', action='store_true')
|
764 |
+
parser.add_argument('--base_half', action='store_true')
|
765 |
+
parser.add_argument('--do_top_k', action='store_true')
|
766 |
+
parser.add_argument('--top_k', type=int, default=40)
|
767 |
+
parser.add_argument('--do_top_p', action='store_true')
|
768 |
+
parser.add_argument('--top_p', type=float, default=0.96)
|
769 |
+
parser.add_argument('--output_name', type=str, default="")
|
770 |
+
parser.add_argument('--openai_model', type=str, default=None)
|
771 |
+
parser.add_argument('--openai_key', type=str)
|
772 |
+
parser.add_argument('--baselines_only', action='store_true')
|
773 |
+
parser.add_argument('--skip_baselines', action='store_true')
|
774 |
+
parser.add_argument('--buffer_size', type=int, default=1)
|
775 |
+
parser.add_argument('--mask_top_p', type=float, default=1.0)
|
776 |
+
parser.add_argument('--pre_perturb_pct', type=float, default=0.0)
|
777 |
+
parser.add_argument('--pre_perturb_span_length', type=int, default=5)
|
778 |
+
parser.add_argument('--random_fills', action='store_true')
|
779 |
+
parser.add_argument('--random_fills_tokens', action='store_true')
|
780 |
+
parser.add_argument('--cache_dir', type=str, default="~/.cache")
|
781 |
+
args = parser.parse_args()
|
782 |
+
|
783 |
+
API_TOKEN_COUNTER = 0
|
784 |
+
|
785 |
+
if args.openai_model is not None:
|
786 |
+
import openai
|
787 |
+
assert args.openai_key is not None, "Must provide OpenAI API key as --openai_key"
|
788 |
+
openai.api_key = args.openai_key
|
789 |
+
|
790 |
+
START_DATE = datetime.datetime.now().strftime('%Y-%m-%d')
|
791 |
+
START_TIME = datetime.datetime.now().strftime('%H-%M-%S-%f')
|
792 |
+
|
793 |
+
# define SAVE_FOLDER as the timestamp - base model name - mask filling model name
|
794 |
+
# create it if it doesn't exist
|
795 |
+
precision_string = "int8" if args.int8 else ("fp16" if args.half else "fp32")
|
796 |
+
sampling_string = "top_k" if args.do_top_k else ("top_p" if args.do_top_p else "temp")
|
797 |
+
output_subfolder = f"{args.output_name}/" if args.output_name else ""
|
798 |
+
if args.openai_model is None:
|
799 |
+
base_model_name = args.base_model_name.replace('/', '_')
|
800 |
+
else:
|
801 |
+
base_model_name = "openai-" + args.openai_model.replace('/', '_')
|
802 |
+
scoring_model_string = (f"-{args.scoring_model_name}" if args.scoring_model_name else "").replace('/', '_')
|
803 |
+
SAVE_FOLDER = f"tmp_results/{output_subfolder}{base_model_name}{scoring_model_string}-{args.mask_filling_model_name}-{sampling_string}/{START_DATE}-{START_TIME}-{precision_string}-{args.pct_words_masked}-{args.n_perturbation_rounds}-{args.dataset}-{args.n_samples}"
|
804 |
+
if not os.path.exists(SAVE_FOLDER):
|
805 |
+
os.makedirs(SAVE_FOLDER)
|
806 |
+
print(f"Saving results to absolute path: {os.path.abspath(SAVE_FOLDER)}")
|
807 |
+
|
808 |
+
# write args to file
|
809 |
+
with open(os.path.join(SAVE_FOLDER, "args.json"), "w") as f:
|
810 |
+
json.dump(args.__dict__, f, indent=4)
|
811 |
+
|
812 |
+
mask_filling_model_name = args.mask_filling_model_name
|
813 |
+
n_samples = args.n_samples
|
814 |
+
batch_size = args.batch_size
|
815 |
+
n_perturbation_list = [int(x) for x in args.n_perturbation_list.split(",")]
|
816 |
+
n_perturbation_rounds = args.n_perturbation_rounds
|
817 |
+
n_similarity_samples = args.n_similarity_samples
|
818 |
+
|
819 |
+
cache_dir = args.cache_dir
|
820 |
+
os.environ["XDG_CACHE_HOME"] = cache_dir
|
821 |
+
if not os.path.exists(cache_dir):
|
822 |
+
os.makedirs(cache_dir)
|
823 |
+
print(f"Using cache dir {cache_dir}")
|
824 |
+
|
825 |
+
GPT2_TOKENIZER = transformers.GPT2Tokenizer.from_pretrained('gpt2', cache_dir=cache_dir)
|
826 |
+
|
827 |
+
# generic generative model
|
828 |
+
base_model, base_tokenizer = load_base_model_and_tokenizer(args.base_model_name)
|
829 |
+
|
830 |
+
# mask filling t5 model
|
831 |
+
if not args.baselines_only and not args.random_fills:
|
832 |
+
int8_kwargs = {}
|
833 |
+
half_kwargs = {}
|
834 |
+
if args.int8:
|
835 |
+
int8_kwargs = dict(load_in_8bit=True, device_map='auto', torch_dtype=torch.bfloat16)
|
836 |
+
elif args.half:
|
837 |
+
half_kwargs = dict(torch_dtype=torch.bfloat16)
|
838 |
+
print(f'Loading mask filling model {mask_filling_model_name}...')
|
839 |
+
mask_model = transformers.AutoModelForSeq2SeqLM.from_pretrained(mask_filling_model_name, **int8_kwargs, **half_kwargs, cache_dir=cache_dir)
|
840 |
+
try:
|
841 |
+
n_positions = mask_model.config.n_positions
|
842 |
+
except AttributeError:
|
843 |
+
n_positions = 512
|
844 |
+
else:
|
845 |
+
n_positions = 512
|
846 |
+
preproc_tokenizer = transformers.AutoTokenizer.from_pretrained('t5-small', model_max_length=512, cache_dir=cache_dir)
|
847 |
+
mask_tokenizer = transformers.AutoTokenizer.from_pretrained(mask_filling_model_name, model_max_length=n_positions, cache_dir=cache_dir)
|
848 |
+
if args.dataset in ['english', 'german']:
|
849 |
+
preproc_tokenizer = mask_tokenizer
|
850 |
+
|
851 |
+
load_base_model()
|
852 |
+
|
853 |
+
print(f'Loading dataset {args.dataset}...')
|
854 |
+
data = generate_data(args.dataset, args.dataset_key)
|
855 |
+
if args.random_fills:
|
856 |
+
FILL_DICTIONARY = set()
|
857 |
+
for texts in data.values():
|
858 |
+
for text in texts:
|
859 |
+
FILL_DICTIONARY.update(text.split())
|
860 |
+
FILL_DICTIONARY = sorted(list(FILL_DICTIONARY))
|
861 |
+
|
862 |
+
if args.scoring_model_name:
|
863 |
+
print(f'Loading SCORING model {args.scoring_model_name}...')
|
864 |
+
del base_model
|
865 |
+
del base_tokenizer
|
866 |
+
torch.cuda.empty_cache()
|
867 |
+
base_model, base_tokenizer = load_base_model_and_tokenizer(args.scoring_model_name)
|
868 |
+
load_base_model() # Load again because we've deleted/replaced the old model
|
869 |
+
|
870 |
+
# write the data to a json file in the save folder
|
871 |
+
with open(os.path.join(SAVE_FOLDER, "raw_data.json"), "w") as f:
|
872 |
+
print(f"Writing raw data to {os.path.join(SAVE_FOLDER, 'raw_data.json')}")
|
873 |
+
json.dump(data, f)
|
874 |
+
|
875 |
+
if not args.skip_baselines:
|
876 |
+
baseline_outputs = [run_baseline_threshold_experiment(get_ll, "likelihood", n_samples=n_samples)]
|
877 |
+
if args.openai_model is None:
|
878 |
+
rank_criterion = lambda text: -get_rank(text, log=False)
|
879 |
+
baseline_outputs.append(run_baseline_threshold_experiment(rank_criterion, "rank", n_samples=n_samples))
|
880 |
+
logrank_criterion = lambda text: -get_rank(text, log=True)
|
881 |
+
baseline_outputs.append(run_baseline_threshold_experiment(logrank_criterion, "log_rank", n_samples=n_samples))
|
882 |
+
entropy_criterion = lambda text: get_entropy(text)
|
883 |
+
baseline_outputs.append(run_baseline_threshold_experiment(entropy_criterion, "entropy", n_samples=n_samples))
|
884 |
+
|
885 |
+
baseline_outputs.append(eval_supervised(data, model='roberta-base-openai-detector'))
|
886 |
+
baseline_outputs.append(eval_supervised(data, model='roberta-large-openai-detector'))
|
887 |
+
|
888 |
+
outputs = []
|
889 |
+
|
890 |
+
if not args.baselines_only:
|
891 |
+
# run perturbation experiments
|
892 |
+
for n_perturbations in n_perturbation_list:
|
893 |
+
perturbation_results = get_perturbation_results(args.span_length, n_perturbations, n_samples)
|
894 |
+
for perturbation_mode in ['d', 'z']:
|
895 |
+
output = run_perturbation_experiment(
|
896 |
+
perturbation_results, perturbation_mode, span_length=args.span_length, n_perturbations=n_perturbations, n_samples=n_samples)
|
897 |
+
outputs.append(output)
|
898 |
+
with open(os.path.join(SAVE_FOLDER, f"perturbation_{n_perturbations}_{perturbation_mode}_results.json"), "w") as f:
|
899 |
+
json.dump(output, f)
|
900 |
+
|
901 |
+
if not args.skip_baselines:
|
902 |
+
# write likelihood threshold results to a file
|
903 |
+
with open(os.path.join(SAVE_FOLDER, f"likelihood_threshold_results.json"), "w") as f:
|
904 |
+
json.dump(baseline_outputs[0], f)
|
905 |
+
|
906 |
+
if args.openai_model is None:
|
907 |
+
# write rank threshold results to a file
|
908 |
+
with open(os.path.join(SAVE_FOLDER, f"rank_threshold_results.json"), "w") as f:
|
909 |
+
json.dump(baseline_outputs[1], f)
|
910 |
+
|
911 |
+
# write log rank threshold results to a file
|
912 |
+
with open(os.path.join(SAVE_FOLDER, f"logrank_threshold_results.json"), "w") as f:
|
913 |
+
json.dump(baseline_outputs[2], f)
|
914 |
+
|
915 |
+
# write entropy threshold results to a file
|
916 |
+
with open(os.path.join(SAVE_FOLDER, f"entropy_threshold_results.json"), "w") as f:
|
917 |
+
json.dump(baseline_outputs[3], f)
|
918 |
+
|
919 |
+
# write supervised results to a file
|
920 |
+
with open(os.path.join(SAVE_FOLDER, f"roberta-base-openai-detector_results.json"), "w") as f:
|
921 |
+
json.dump(baseline_outputs[-2], f)
|
922 |
+
|
923 |
+
# write supervised results to a file
|
924 |
+
with open(os.path.join(SAVE_FOLDER, f"roberta-large-openai-detector_results.json"), "w") as f:
|
925 |
+
json.dump(baseline_outputs[-1], f)
|
926 |
+
|
927 |
+
outputs += baseline_outputs
|
928 |
+
|
929 |
+
save_roc_curves(outputs)
|
930 |
+
save_ll_histograms(outputs)
|
931 |
+
save_llr_histograms(outputs)
|
932 |
+
|
933 |
+
# move results folder from tmp_results/ to results/, making sure necessary directories exist
|
934 |
+
new_folder = SAVE_FOLDER.replace("tmp_results", "results")
|
935 |
+
if not os.path.exists(os.path.dirname(new_folder)):
|
936 |
+
os.makedirs(os.path.dirname(new_folder))
|
937 |
+
os.rename(SAVE_FOLDER, new_folder)
|
938 |
+
|
939 |
+
print(f"Used an *estimated* {API_TOKEN_COUNTER} API tokens (may be inaccurate)")
|