|
import torch |
|
from transformers import AutoTokenizer, AutoModelForMaskedLM |
|
from transformers import pipeline |
|
import random |
|
from nltk.corpus import stopwords |
|
import nltk |
|
nltk.download('stopwords') |
|
import math |
|
from vocabulary_split import split_vocabulary, filter_logits |
|
import abc |
|
from typing import List |
|
|
|
|
|
tokenizer = AutoTokenizer.from_pretrained("bert-large-cased-whole-word-masking") |
|
model = AutoModelForMaskedLM.from_pretrained("bert-large-cased-whole-word-masking") |
|
fill_mask = pipeline("fill-mask", model=model, tokenizer=tokenizer) |
|
|
|
|
|
permissible, _ = split_vocabulary(seed=42) |
|
permissible_indices = torch.tensor([i in permissible.values() for i in range(len(tokenizer))]) |
|
|
|
def get_logits_for_mask(model, tokenizer, sentence): |
|
inputs = tokenizer(sentence, return_tensors="pt") |
|
mask_token_index = torch.where(inputs["input_ids"] == tokenizer.mask_token_id)[1] |
|
|
|
with torch.no_grad(): |
|
outputs = model(**inputs) |
|
|
|
logits = outputs.logits |
|
mask_token_logits = logits[0, mask_token_index, :] |
|
return mask_token_logits.squeeze() |
|
|
|
|
|
class MaskingStrategy(abc.ABC): |
|
@abc.abstractmethod |
|
def select_words_to_mask(self, words: List[str], **kwargs) -> List[int]: |
|
""" |
|
Given a list of words, return the indices of words to mask. |
|
""" |
|
pass |
|
|
|
|
|
class RandomNonStopwordMasking(MaskingStrategy): |
|
def __init__(self, num_masks: int = 1): |
|
self.num_masks = num_masks |
|
self.stop_words = set(stopwords.words('english')) |
|
|
|
def select_words_to_mask(self, words: List[str], **kwargs) -> List[int]: |
|
non_stop_indices = [i for i, word in enumerate(words) if word.lower() not in self.stop_words] |
|
if not non_stop_indices: |
|
return [] |
|
num_masks = min(self.num_masks, len(non_stop_indices)) |
|
return random.sample(non_stop_indices, num_masks) |
|
|
|
class HighEntropyMasking(MaskingStrategy): |
|
def __init__(self, num_masks: int = 1): |
|
self.num_masks = num_masks |
|
|
|
def select_words_to_mask(self, words: List[str], sentence: str, model, tokenizer, permissible_indices) -> List[int]: |
|
candidate_indices = [i for i, word in enumerate(words) if word.lower() not in set(stopwords.words('english'))] |
|
if not candidate_indices: |
|
return [] |
|
|
|
entropy_scores = {} |
|
for idx in candidate_indices: |
|
masked_sentence = ' '.join(words[:idx] + ['[MASK]'] + words[idx+1:]) |
|
logits = get_logits_for_mask(model, tokenizer, masked_sentence) |
|
filtered_logits = filter_logits(logits, permissible_indices) |
|
probs = torch.softmax(filtered_logits, dim=-1) |
|
top_5_probs = probs.topk(5).values |
|
entropy = -torch.sum(top_5_probs * torch.log(top_5_probs + 1e-10)).item() |
|
entropy_scores[idx] = entropy |
|
|
|
|
|
sorted_indices = sorted(entropy_scores, key=entropy_scores.get, reverse=True) |
|
return sorted_indices[:self.num_masks] |
|
|
|
class PseudoRandomNonStopwordMasking(MaskingStrategy): |
|
def __init__(self, num_masks: int = 1, seed: int = 10): |
|
self.num_masks = num_masks |
|
self.seed = seed |
|
self.stop_words = set(stopwords.words('english')) |
|
|
|
def select_words_to_mask(self, words: List[str], **kwargs) -> List[int]: |
|
non_stop_indices = [i for i, word in enumerate(words) if word.lower() not in self.stop_words] |
|
if not non_stop_indices: |
|
return [] |
|
random.seed(self.seed) |
|
num_masks = min(self.num_masks, len(non_stop_indices)) |
|
return random.sample(non_stop_indices, num_masks) |
|
|
|
class CompositeMaskingStrategy(MaskingStrategy): |
|
def __init__(self, strategies: List[MaskingStrategy]): |
|
self.strategies = strategies |
|
|
|
def select_words_to_mask(self, words: List[str], **kwargs) -> List[int]: |
|
selected_indices = [] |
|
for strategy in self.strategies: |
|
if isinstance(strategy, HighEntropyMasking): |
|
selected = strategy.select_words_to_mask(words, **kwargs) |
|
else: |
|
selected = strategy.select_words_to_mask(words) |
|
selected_indices.extend(selected) |
|
return list(set(selected_indices)) |
|
|
|
|
|
def mask_between_lcs(sentence, lcs_points, masking_strategy: MaskingStrategy, model, tokenizer, permissible_indices): |
|
words = sentence.split() |
|
masked_indices = [] |
|
|
|
segments = [] |
|
|
|
|
|
previous = 0 |
|
for point in lcs_points: |
|
if point > previous: |
|
segments.append((previous, point)) |
|
previous = point + 1 |
|
if previous < len(words): |
|
segments.append((previous, len(words))) |
|
|
|
|
|
for start, end in segments: |
|
segment_words = words[start:end] |
|
if isinstance(masking_strategy, HighEntropyMasking): |
|
selected = masking_strategy.select_words_to_mask(segment_words, sentence, model, tokenizer, permissible_indices) |
|
else: |
|
selected = masking_strategy.select_words_to_mask(segment_words) |
|
|
|
|
|
for idx in selected: |
|
masked_idx = start + idx |
|
if masked_idx not in masked_indices: |
|
masked_indices.append(masked_idx) |
|
|
|
|
|
for idx in masked_indices: |
|
words[idx] = '[MASK]' |
|
|
|
masked_sentence = ' '.join(words) |
|
logits = get_logits_for_mask(model, tokenizer, masked_sentence) |
|
|
|
|
|
top_words_list = [] |
|
logits_list = [] |
|
for i, idx in enumerate(masked_indices): |
|
logits_i = logits[i] |
|
if logits_i.dim() > 1: |
|
logits_i = logits_i.squeeze() |
|
filtered_logits_i = filter_logits(logits_i, permissible_indices) |
|
logits_list.append(filtered_logits_i.tolist()) |
|
top_5_indices = filtered_logits_i.topk(5).indices.tolist() |
|
top_words = [tokenizer.decode([i]) for i in top_5_indices] |
|
top_words_list.append(top_words) |
|
|
|
return masked_sentence, logits_list, top_words_list |
|
|
|
|
|
if __name__ == "__main__": |
|
|
|
sentence = "This is a sample sentence with some LCS points" |
|
lcs_points = [2, 5, 8] |
|
|
|
|
|
random_non_stopword_strategy = RandomNonStopwordMasking(num_masks=1) |
|
high_entropy_strategy = HighEntropyMasking(num_masks=1) |
|
pseudo_random_strategy = PseudoRandomNonStopwordMasking(num_masks=1, seed=10) |
|
composite_strategy = CompositeMaskingStrategy([ |
|
RandomNonStopwordMasking(num_masks=1), |
|
HighEntropyMasking(num_masks=1) |
|
]) |
|
|
|
|
|
chosen_strategy = composite_strategy |
|
|
|
|
|
masked_sentence, logits_list, top_words_list = mask_between_lcs( |
|
sentence, |
|
lcs_points, |
|
masking_strategy=chosen_strategy, |
|
model=model, |
|
tokenizer=tokenizer, |
|
permissible_indices=permissible_indices |
|
) |
|
|
|
print("Masked Sentence:", masked_sentence) |
|
for idx, top_words in enumerate(top_words_list): |
|
print(f"Top words for mask {idx+1}:", top_words) |
|
|