multi-qa-mpnet-base-cos-v1

Uso (Sentence-Transformers)

Instalar sentence-transformers: pip install -U sentence-transformers

from sentence_transformers import SentenceTransformer, util

query = "How many people live in London?"
docs = ["Around 9 Million people live in London", "London is known for its financial district"]

#Cargar el modelo
model = SentenceTransformer('sentence-transformers/multi-qa-mpnet-base-cos-v1')

#Codificar pregunta y documentos
query_emb = model.encode(query)
doc_emb = model.encode(docs)

#Calcular la puntuación de punto entre la pregunta y todos los embeddings de documentos
scores = util.dot_score(query_emb, doc_emb)[0].cpu().tolist()

#Combinar documentos y puntuaciones
doc_score_pairs = list(zip(docs, scores))

#Ordenar por puntuación decreciente
doc_score_pairs = sorted(doc_score_pairs, key=lambda x: x[1], reverse=True)

#Mostrar pasajes y puntuaciones
for doc, score in doc_score_pairs:
    print(score, doc)

Uso (HuggingFace Transformers)

from transformers import AutoTokenizer, AutoModel
import torch
import torch.nn.functional as F

# Mean Pooling - Tomar el promedio de todos los tokens
def mean_pooling(model_output, attention_mask):
    token_embeddings = model_output.last_hidden_state  # Primer elemento de model_output contiene todos los embeddings de tokens
    input_mask_expanded = attention_mask.unsqueeze(-1).expand(token_embeddings.size()).float()
    return torch.sum(token_embeddings * input_mask_expanded, 1) / torch.clamp(input_mask_expanded.sum(1), min=1e-9)

# Codificar texto
def encode(texts):
    # Tokenizar oraciones
    encoded_input = tokenizer(texts, padding=True, truncation=True, return_tensors='pt')

    # Calcular los embeddings de los tokens
    with torch.no_grad():
        model_output = model(**encoded_input, return_dict=True)

    # Realizar pooling
    embeddings = mean_pooling(model_output, encoded_input['attention_mask'])

    # Normalizar embeddings
    embeddings = F.normalize(embeddings, p=2, dim=1)

    return embeddings

# Oraciones para las cuales queremos embeddings de oraciones
query = "How many people live in London?"
docs = ["Around 9 Million people live in London", "London is known for its financial district"]

# Cargar el modelo desde HuggingFace Hub
tokenizer = AutoTokenizer.from_pretrained("sentence-transformers/multi-qa-mpnet-base-cos-v1")
model = AutoModel.from_pretrained("sentence-transformers/multi-qa-mpnet-base-cos-v1")

#Codificar pregunta y documentos
def encode(texts):
    encoded_input = tokenizer(texts, padding=True, truncation=True, return_tensors='pt')
    with torch.no_grad():
        model_output = model(**encoded_input, return_dict=True)

    embeddings = mean_pooling(model_output, encoded_input['attention_mask'])
    embeddings = F.normalize(embeddings, p=2, dim=1)

    return embeddings

query_emb = encode(query)
doc_emb = encode(docs)

#Calcular la puntuación de punto entre la pregunta y todos los embeddings de documentos
scores = torch.mm(query_emb, doc_emb.transpose(0, 1))[0].cpu().tolist()

#Combinar documentos & puntuaciones
doc_score_pairs = list(zip(docs, scores))

#Ordenar por puntuación decreciente
doc_score_pairs = sorted(doc_score_pairs, key=lambda x: x[1], reverse=True)

#Mostrar pasajes & puntuaciones
for doc, score in doc_score_pairs:
    print(score, doc)