OlympIA / webrankings_helper.py
johannoriel's picture
Initial relase. Tested. Working
f34a6fd
import requests
from bs4 import BeautifulSoup
import pandas as pd
import torch
from transformers import pipeline
from sentence_transformers import SentenceTransformer, util
import concurrent.futures
import time
import sys
from sklearn.feature_extraction.text import TfidfVectorizer
from sklearn.metrics.pairwise import cosine_similarity
from transformers import AutoTokenizer, AutoModel
import numpy as np
from scipy import stats
from PyDictionary import PyDictionary
import matplotlib.pyplot as plt
from scipy import stats
import litellm
import re
import sentencepiece
import random
def score_with_llm(title, topic, llm_model):
prompt = f"""Evaluate the relevance of the following article to the topic '{topic}'.
Article title: {title}
Give a final relevance score between 0 and 1, where 1 is very relevant and 0 is not relevant at all.
Respond only with a number between 0 and 1."""
try:
response = litellm.completion(
model=llm_model,
messages=[{"role": "user", "content": prompt}],
max_tokens=10
)
score_match = re.search(r'\d+(\.\d+)?', response.choices[0].message.content.strip())
if score_match:
score = float(score_match.group())
print(f"Score LLM : {score}")
return max(0, min(score, 1))
else:
print(f"Could not extract a score from LLM response: {response.choices[0].message.content}")
return None
except Exception as e:
print(f"Error in scoring with LLM {llm_model}: {str(e)}")
return None
def expand_keywords_llm(keyword, max_synonyms=3, llm_model="ollama/qwen2"):
prompt = f"""Please provide up to {max_synonyms} synonyms or closely related terms for the word or phrase: "{keyword}".
Return only the list of synonyms, separated by commas, without any additional explanation."""
try:
response = litellm.completion(
model=llm_model,
messages=[{"role": "user", "content": prompt}],
max_tokens=50
)
synonyms = [s.strip() for s in response.choices[0].message.content.split(',')]
return [keyword] + synonyms[:max_synonyms]
except Exception as e:
print(f"Error in expanding keywords with LLM {llm_model}: {str(e)}")
return [keyword]
# Fonction pour obtenir les liens de la page d'accueil
def get_homepage_links(url):
response = requests.get(url)
soup = BeautifulSoup(response.text, 'html.parser')
links = soup.find_all('a', href=True)
return [(link.text.strip(), link['href']) for link in links if link.text.strip()]
# Fonction pour obtenir le contenu d'un article
def get_article_content(url):
try:
print(f"Récupération du contenu de : {url}")
response = requests.get(url)
print(f"Taille de la réponse HTTP : {len(response.content)} octets") # Affiche le nombre d'octets de la réponse HTTP
soup = BeautifulSoup(response.text, 'html.parser')
print(f"Taille de l'objet soup : {sys.getsizeof(soup)} octets") # Affiche la taille en mémoire de l'objet soup
article = soup.find('article')
if article:
paragraphs = article.find_all('p')
content = ' '.join([p.text for p in paragraphs])
print(f"Paragraphes récupéré : {len(content)} caractères")
return content
print("Aucun contenu d'article trouvé")
return ""
except Exception as e:
print(f"Erreur lors de la récupération du contenu : {str(e)}")
return ""
# Fonction pour l'analyse zero-shot
def zero_shot_analysis(text, topic, classifier):
if not text:
print("Texte vide pour l'analyse zero-shot")
return 0.0
result = classifier(text, candidate_labels=[topic, f"not {topic}"], multi_label=False)
print(f"Score zero-shot : {result['scores'][0]}")
return result['scores'][0]
# Fonction pour l'analyse par embeddings
def embedding_analysis(text, topic_embedding, model):
if not text:
print("Texte vide pour l'analyse par embeddings")
return 0.0
text_embedding = model.encode([text], convert_to_tensor=True)
similarity = util.pytorch_cos_sim(text_embedding, topic_embedding).item()
print(f"Score embedding : {similarity}")
return similarity
from sklearn.feature_extraction.text import TfidfVectorizer
from sklearn.metrics.pairwise import cosine_similarity
#import nltk
#from nltk.corpus import wordnet
#nltk.download('wordnet')
def preprocess_text(text):
# Tokenize the text
tokens = text.lower().split()
# Expand each token with its synonyms
expanded_tokens = []
for token in tokens:
synonyms = set()
for syn in wordnet.synsets(token):
for lemma in syn.lemmas():
synonyms.add(lemma.name())
expanded_tokens.extend(list(synonyms))
return ' '.join(expanded_tokens)
def improved_tfidf_similarity(texts, query):
# Preprocess texts and query
preprocessed_texts = [preprocess_text(text) for text in texts]
preprocessed_query = preprocess_text(query)
# Combine texts and query for vectorization
all_texts = preprocessed_texts + [preprocessed_query]
# Use TfidfVectorizer with custom parameters
vectorizer = TfidfVectorizer(ngram_range=(1, 2), min_df=1, smooth_idf=True)
tfidf_matrix = vectorizer.fit_transform(all_texts)
# Calculate cosine similarity
cosine_similarities = cosine_similarity(tfidf_matrix[-1], tfidf_matrix[:-1]).flatten()
# Normalize similarities to avoid zero scores
normalized_similarities = (cosine_similarities - cosine_similarities.min()) / (cosine_similarities.max() - cosine_similarities.min())
return normalized_similarities
from sklearn.feature_extraction.text import TfidfVectorizer
from sklearn.metrics.pairwise import cosine_similarity
import numpy as np
def improved_tfidf_similarity_v2(texts, query):
# Combine texts and query, treating each word or phrase as a separate document
all_docs = [word.strip() for text in texts for word in text.split(',')] + [word.strip() for word in query.split(',')]
# Create TF-IDF matrix
vectorizer = TfidfVectorizer()
tfidf_matrix = vectorizer.fit_transform(all_docs)
# Calculate document vectors by summing the TF-IDF vectors of their words
doc_vectors = []
query_vector = np.zeros((1, tfidf_matrix.shape[1]))
current_doc = 0
for i, doc in enumerate(all_docs):
if i < len(all_docs) - len(query.split(',')): # If it's part of the texts
if current_doc == len(texts):
break
if doc in texts[current_doc]:
doc_vectors.append(tfidf_matrix[i].toarray())
else:
current_doc += 1
doc_vectors.append(tfidf_matrix[i].toarray())
else: # If it's part of the query
query_vector += tfidf_matrix[i].toarray()
doc_vectors = np.array([np.sum(doc, axis=0) for doc in doc_vectors])
# Calculate cosine similarity
similarities = cosine_similarity(query_vector, doc_vectors).flatten()
# Normalize similarities to avoid zero scores
normalized_similarities = (similarities - similarities.min()) / (similarities.max() - similarities.min() + 1e-8)
return normalized_similarities
# Example usage:
# texts = ["longevity, health, aging", "computer science, AI"]
# query = "longevity, life extension, anti-aging"
# results = improved_tfidf_similarity_v2(texts, query)
# print(results)
# Nouvelles fonctions
def tfidf_similarity(texts, query):
vectorizer = TfidfVectorizer()
tfidf_matrix = vectorizer.fit_transform(texts + [query])
cosine_similarities = cosine_similarity(tfidf_matrix[-1], tfidf_matrix[:-1]).flatten()
return cosine_similarities
def bert_similarity(texts, query, model_name='bert-base-uncased'):
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModel.from_pretrained(model_name)
def get_embedding(text):
inputs = tokenizer(text, return_tensors='pt', truncation=True, padding=True, max_length=512)
with torch.no_grad():
outputs = model(**inputs)
return outputs.last_hidden_state.mean(dim=1).squeeze().numpy()
query_embedding = get_embedding(query)
text_embeddings = [get_embedding(text) for text in texts]
similarities = [cosine_similarity([query_embedding], [text_embedding])[0][0] for text_embedding in text_embeddings]
return similarities
# Fonction principale d'analyse modifiée
def analyze_link(title, link, topic, zero_shot_classifiers, embedding_models, expanded_query, llm_models, testcontent):
print(f"\nAnalyse de : {title}")
results = {
"Titre": title,
#"TF-IDF (titre)": improved_tfidf_similarity_v2([title], expanded_query)[0],
#"BERT (titre)": bert_similarity([title], expanded_query)[0],
}
# Zero-shot analysis
for name, classifier in zero_shot_classifiers.items():
results[f"Zero-shot (titre) - {name}"] = zero_shot_analysis(title, topic, classifier)
# Embedding analysis
for name, model in embedding_models.items():
topic_embedding = model.encode([expanded_query], convert_to_tensor=True)
results[f"Embeddings (titre) - {name}"] = embedding_analysis(title, topic_embedding, model)
# LLM analysis
for model in llm_models:
results[f"LLM Score - {model}"] = score_with_llm(title, topic, model)
if testcontent:
content = get_article_content(link)
#results["TF-IDF (contenu)"] = improved_tfidf_similarity_v2([content], expanded_query)[0]
#results["BERT (contenu)"]= bert_similarity([content], expanded_query)[0]
# Zero-shot analysis
for name, classifier in zero_shot_classifiers.items():
results[f"Zero-shot (contenu) - {name}"] = zero_shot_analysis(content, topic, classifier)
# Embedding analysis
for name, model in embedding_models.items():
topic_embedding = model.encode([expanded_query], convert_to_tensor=True)
results[f"Embeddings (contenu) - {name}"] = embedding_analysis(content, topic_embedding, model)
# LLM analysis
for model in llm_models:
results[f"LLM Content Score - {model}"] = score_with_llm(content, topic, model)
return results
from scipy import stats
def evaluate_ranking(reference_data_valid, reference_data_rejected, method_scores, threshold, silent):
simple_score = 0
true_positives = 0
false_positives = 0
true_negatives = 0
false_negatives = 0
# Créer une liste de tous les éléments avec leur statut (1 pour valide, 0 pour rejeté)
all_items = [(item, 1) for item in reference_data_valid] + [(item, 0) for item in reference_data_rejected]
# Trier les éléments selon leur score dans la méthode
all_items_temp = all_items.copy()
# correct false positive if method spit out same score for all
#random.shuffle(all_items_temp)
all_items_temp.reverse()
sorted_method = sorted([(item, method_scores.get(item, 0)) for item, _ in all_items_temp],
key=lambda x: x[1], reverse=True)
# Créer des listes pour le calcul de la corrélation de Spearman
reference_ranks = []
method_ranks = []
for i, (item, status) in enumerate(all_items):
method_score = method_scores.get(item, 0)
method_rank = next(j for j, (it, score) in enumerate(sorted_method) if it == item)
reference_ranks.append(i)
method_ranks.append(method_rank)
if status == 1: # Item valide
if method_score >= threshold:
simple_score += 1
true_positives += 1
else:
simple_score -= 1
false_negatives += 1
else: # Item rejeté
if method_score < threshold:
simple_score += 1
true_negatives += 1
else:
simple_score -= 1
false_positives += 1
# Calculer le coefficient de corrélation de Spearman
if not silent:
print("+++")
print(reference_ranks)
print("---")
print(method_ranks)
spearman_corr, _ = stats.spearmanr(reference_ranks, method_ranks)
# Calculer la précision, le rappel et le F1-score
precision = true_positives / (true_positives + false_positives) if (true_positives + false_positives) > 0 else 0
recall = true_positives / (true_positives + false_negatives) if (true_positives + false_negatives) > 0 else 0
f1_score = 2 * (precision * recall) / (precision + recall) if (precision + recall) > 0 else 0
return {
"simple_score": simple_score,
"spearman_correlation": spearman_corr,
"precision": precision,
"recall": recall,
"f1_score": f1_score,
}
def find_optimal_threshold(reference_data_valid, reference_data_rejected, method_scores):
best_score = float('-inf')
best_threshold = 0
for threshold in np.arange(0, 1.05, 0.05):
result = evaluate_ranking(
reference_data_valid,
reference_data_rejected,
method_scores,
threshold, True
)
if result['simple_score'] > best_score:
best_score = result['simple_score']
best_threshold = threshold
return best_threshold
def reset_cuda_context():
torch.cuda.empty_cache()
torch.cuda.ipc_collect()
if torch.cuda.is_available():
torch.cuda.set_device(torch.cuda.current_device())
torch.cuda.synchronize()
import gc
def clear_models():
global zero_shot_classifiers, embedding_models_dict, bert_models, tfidf_objects
for classifier in zero_shot_classifiers.values():
del classifier
zero_shot_classifiers.clear()
for model in embedding_models_dict.values():
del model
embedding_models_dict.clear()
for model in bert_models:
del model
bert_models.clear()
for vectorizer in tfidf_objects:
del vectorizer
tfidf_objects.clear()
torch.cuda.empty_cache()
gc.collect()
def clear_globals():
for name in list(globals()):
if isinstance(globals()[name], (torch.nn.Module, torch.Tensor)):
del globals()[name]
def release_vram(zero_shot_classifiers, embedding_models, bert_models, tfidf_objects):
# Supprimer les objets zero-shot classifiers
for model in zero_shot_classifiers.values():
del model
# Supprimer les objets embedding models
for model in embedding_models.values():
del model
# Supprimer les objets bert models
for model in bert_models:
del model
# Supprimer les objets tfidf objects
for obj in tfidf_objects:
del obj
# Vider le cache de la mémoire GPU
torch.cuda.empty_cache()
torch.cuda.synchronize()
gc.collect()
clear_globals()
reset_cuda_context()
def load_finetuned_model(model_path):
checkpoint = torch.load(model_path)
base_model = AutoModel.from_pretrained(checkpoint['base_model_name'])
model = EmbeddingModel(base_model)
model.load_state_dict(checkpoint['model_state_dict'])
return model