Spaces:
Running
Running
Upload folder using huggingface_hub
Browse files- masking_methods_trial.py +188 -0
- requirements.txt +2 -1
masking_methods_trial.py
ADDED
@@ -0,0 +1,188 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
from transformers import AutoTokenizer, AutoModelForMaskedLM
|
3 |
+
from transformers import pipeline
|
4 |
+
import random
|
5 |
+
from nltk.corpus import stopwords
|
6 |
+
import nltk
|
7 |
+
nltk.download('stopwords')
|
8 |
+
import math
|
9 |
+
from vocabulary_split import split_vocabulary, filter_logits
|
10 |
+
import abc
|
11 |
+
from typing import List
|
12 |
+
|
13 |
+
# Load tokenizer and model for masked language model
|
14 |
+
tokenizer = AutoTokenizer.from_pretrained("bert-large-cased-whole-word-masking")
|
15 |
+
model = AutoModelForMaskedLM.from_pretrained("bert-large-cased-whole-word-masking")
|
16 |
+
fill_mask = pipeline("fill-mask", model=model, tokenizer=tokenizer)
|
17 |
+
|
18 |
+
# Get permissible vocabulary
|
19 |
+
permissible, _ = split_vocabulary(seed=42)
|
20 |
+
permissible_indices = torch.tensor([i in permissible.values() for i in range(len(tokenizer))])
|
21 |
+
|
22 |
+
def get_logits_for_mask(model, tokenizer, sentence):
|
23 |
+
inputs = tokenizer(sentence, return_tensors="pt")
|
24 |
+
mask_token_index = torch.where(inputs["input_ids"] == tokenizer.mask_token_id)[1]
|
25 |
+
|
26 |
+
with torch.no_grad():
|
27 |
+
outputs = model(**inputs)
|
28 |
+
|
29 |
+
logits = outputs.logits
|
30 |
+
mask_token_logits = logits[0, mask_token_index, :]
|
31 |
+
return mask_token_logits.squeeze()
|
32 |
+
|
33 |
+
# Abstract Masking Strategy
|
34 |
+
class MaskingStrategy(abc.ABC):
|
35 |
+
@abc.abstractmethod
|
36 |
+
def select_words_to_mask(self, words: List[str], **kwargs) -> List[int]:
|
37 |
+
"""
|
38 |
+
Given a list of words, return the indices of words to mask.
|
39 |
+
"""
|
40 |
+
pass
|
41 |
+
|
42 |
+
# Specific Masking Strategies
|
43 |
+
class RandomNonStopwordMasking(MaskingStrategy):
|
44 |
+
def __init__(self, num_masks: int = 1):
|
45 |
+
self.num_masks = num_masks
|
46 |
+
self.stop_words = set(stopwords.words('english'))
|
47 |
+
|
48 |
+
def select_words_to_mask(self, words: List[str], **kwargs) -> List[int]:
|
49 |
+
non_stop_indices = [i for i, word in enumerate(words) if word.lower() not in self.stop_words]
|
50 |
+
if not non_stop_indices:
|
51 |
+
return []
|
52 |
+
num_masks = min(self.num_masks, len(non_stop_indices))
|
53 |
+
return random.sample(non_stop_indices, num_masks)
|
54 |
+
|
55 |
+
class HighEntropyMasking(MaskingStrategy):
|
56 |
+
def __init__(self, num_masks: int = 1):
|
57 |
+
self.num_masks = num_masks
|
58 |
+
|
59 |
+
def select_words_to_mask(self, words: List[str], sentence: str, model, tokenizer, permissible_indices) -> List[int]:
|
60 |
+
candidate_indices = [i for i, word in enumerate(words) if word.lower() not in set(stopwords.words('english'))]
|
61 |
+
if not candidate_indices:
|
62 |
+
return []
|
63 |
+
|
64 |
+
entropy_scores = {}
|
65 |
+
for idx in candidate_indices:
|
66 |
+
masked_sentence = ' '.join(words[:idx] + ['[MASK]'] + words[idx+1:])
|
67 |
+
logits = get_logits_for_mask(model, tokenizer, masked_sentence)
|
68 |
+
filtered_logits = filter_logits(logits, permissible_indices)
|
69 |
+
probs = torch.softmax(filtered_logits, dim=-1)
|
70 |
+
top_5_probs = probs.topk(5).values
|
71 |
+
entropy = -torch.sum(top_5_probs * torch.log(top_5_probs + 1e-10)).item()
|
72 |
+
entropy_scores[idx] = entropy
|
73 |
+
|
74 |
+
# Select top N indices with highest entropy
|
75 |
+
sorted_indices = sorted(entropy_scores, key=entropy_scores.get, reverse=True)
|
76 |
+
return sorted_indices[:self.num_masks]
|
77 |
+
|
78 |
+
class PseudoRandomNonStopwordMasking(MaskingStrategy):
|
79 |
+
def __init__(self, num_masks: int = 1, seed: int = 10):
|
80 |
+
self.num_masks = num_masks
|
81 |
+
self.seed = seed
|
82 |
+
self.stop_words = set(stopwords.words('english'))
|
83 |
+
|
84 |
+
def select_words_to_mask(self, words: List[str], **kwargs) -> List[int]:
|
85 |
+
non_stop_indices = [i for i, word in enumerate(words) if word.lower() not in self.stop_words]
|
86 |
+
if not non_stop_indices:
|
87 |
+
return []
|
88 |
+
random.seed(self.seed)
|
89 |
+
num_masks = min(self.num_masks, len(non_stop_indices))
|
90 |
+
return random.sample(non_stop_indices, num_masks)
|
91 |
+
|
92 |
+
class CompositeMaskingStrategy(MaskingStrategy):
|
93 |
+
def __init__(self, strategies: List[MaskingStrategy]):
|
94 |
+
self.strategies = strategies
|
95 |
+
|
96 |
+
def select_words_to_mask(self, words: List[str], **kwargs) -> List[int]:
|
97 |
+
selected_indices = []
|
98 |
+
for strategy in self.strategies:
|
99 |
+
if isinstance(strategy, HighEntropyMasking):
|
100 |
+
selected = strategy.select_words_to_mask(words, **kwargs)
|
101 |
+
else:
|
102 |
+
selected = strategy.select_words_to_mask(words)
|
103 |
+
selected_indices.extend(selected)
|
104 |
+
return list(set(selected_indices)) # Remove duplicates
|
105 |
+
|
106 |
+
# Refactored mask_between_lcs function
|
107 |
+
def mask_between_lcs(sentence, lcs_points, masking_strategy: MaskingStrategy, model, tokenizer, permissible_indices):
|
108 |
+
words = sentence.split()
|
109 |
+
masked_indices = []
|
110 |
+
|
111 |
+
segments = []
|
112 |
+
|
113 |
+
# Define segments based on LCS points
|
114 |
+
previous = 0
|
115 |
+
for point in lcs_points:
|
116 |
+
if point > previous:
|
117 |
+
segments.append((previous, point))
|
118 |
+
previous = point + 1
|
119 |
+
if previous < len(words):
|
120 |
+
segments.append((previous, len(words)))
|
121 |
+
|
122 |
+
# Collect all indices to mask from each segment
|
123 |
+
for start, end in segments:
|
124 |
+
segment_words = words[start:end]
|
125 |
+
if isinstance(masking_strategy, HighEntropyMasking):
|
126 |
+
selected = masking_strategy.select_words_to_mask(segment_words, sentence, model, tokenizer, permissible_indices)
|
127 |
+
else:
|
128 |
+
selected = masking_strategy.select_words_to_mask(segment_words)
|
129 |
+
|
130 |
+
# Adjust indices relative to the whole sentence
|
131 |
+
for idx in selected:
|
132 |
+
masked_idx = start + idx
|
133 |
+
if masked_idx not in masked_indices:
|
134 |
+
masked_indices.append(masked_idx)
|
135 |
+
|
136 |
+
# Apply masking
|
137 |
+
for idx in masked_indices:
|
138 |
+
words[idx] = '[MASK]'
|
139 |
+
|
140 |
+
masked_sentence = ' '.join(words)
|
141 |
+
logits = get_logits_for_mask(model, tokenizer, masked_sentence)
|
142 |
+
|
143 |
+
# Process each masked token
|
144 |
+
top_words_list = []
|
145 |
+
logits_list = []
|
146 |
+
for i, idx in enumerate(masked_indices):
|
147 |
+
logits_i = logits[i]
|
148 |
+
if logits_i.dim() > 1:
|
149 |
+
logits_i = logits_i.squeeze()
|
150 |
+
filtered_logits_i = filter_logits(logits_i, permissible_indices)
|
151 |
+
logits_list.append(filtered_logits_i.tolist())
|
152 |
+
top_5_indices = filtered_logits_i.topk(5).indices.tolist()
|
153 |
+
top_words = [tokenizer.decode([i]) for i in top_5_indices]
|
154 |
+
top_words_list.append(top_words)
|
155 |
+
|
156 |
+
return masked_sentence, logits_list, top_words_list
|
157 |
+
|
158 |
+
# Example Usage
|
159 |
+
if __name__ == "__main__":
|
160 |
+
# Example sentence and LCS points
|
161 |
+
sentence = "This is a sample sentence with some LCS points"
|
162 |
+
lcs_points = [2, 5, 8] # Indices of LCS points
|
163 |
+
|
164 |
+
# Initialize masking strategies
|
165 |
+
random_non_stopword_strategy = RandomNonStopwordMasking(num_masks=1)
|
166 |
+
high_entropy_strategy = HighEntropyMasking(num_masks=1)
|
167 |
+
pseudo_random_strategy = PseudoRandomNonStopwordMasking(num_masks=1, seed=10)
|
168 |
+
composite_strategy = CompositeMaskingStrategy([
|
169 |
+
RandomNonStopwordMasking(num_masks=1),
|
170 |
+
HighEntropyMasking(num_masks=1)
|
171 |
+
])
|
172 |
+
|
173 |
+
# Choose a strategy
|
174 |
+
chosen_strategy = composite_strategy # You can choose any initialized strategy
|
175 |
+
|
176 |
+
# Apply masking
|
177 |
+
masked_sentence, logits_list, top_words_list = mask_between_lcs(
|
178 |
+
sentence,
|
179 |
+
lcs_points,
|
180 |
+
masking_strategy=chosen_strategy,
|
181 |
+
model=model,
|
182 |
+
tokenizer=tokenizer,
|
183 |
+
permissible_indices=permissible_indices
|
184 |
+
)
|
185 |
+
|
186 |
+
print("Masked Sentence:", masked_sentence)
|
187 |
+
for idx, top_words in enumerate(top_words_list):
|
188 |
+
print(f"Top words for mask {idx+1}:", top_words)
|
requirements.txt
CHANGED
@@ -17,4 +17,5 @@ graphviz==0.20.3
|
|
17 |
gradio==4.29.0
|
18 |
openai
|
19 |
python-dotenv
|
20 |
-
scikit-learn
|
|
|
|
17 |
gradio==4.29.0
|
18 |
openai
|
19 |
python-dotenv
|
20 |
+
scikit-learn
|
21 |
+
sentence-transformers
|