semantic_xlmr

Usando Sentence-Transformers

Instalar el paquete de sentence-transformers:

pip install -U sentence-transformers

Luego usar el modelo de la siguiente manera:

from sentence_transformers import SentenceTransformer
sentences = ["I sing in bengali", "আমি বাংলায় গান গাই"]
model = SentenceTransformer('headlesstech/semantic_xlmr')
embeddings = model.encode(sentences)
print(embeddings)

Usando HuggingFace Transformers

Usar el modelo sin sentence-transformers:

from transformers import AutoTokenizer, AutoModel
import torch
# Mean Pooling - Tomar en cuenta la máscara de atención para una media correcta

#Function to mean pool

def mean_pooling(model_output, attention_mask):
    token_embeddings = model_output[0] # Primer elemento de model_output contiene todas las token embeddings
    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)
# Oraciones para las cuales queremos las embeddings de oraciones
sentences = ["I sing in bengali", "আমি বাংলায় গান গাই"]
# Cargar el modelo desde HuggingFace Hub

tokenizer = AutoTokenizer.from_pretrained('headlesstech/semantic_xlmr')
model = AutoModel.from_pretrained('headlesstech/semantic_xlmr')
# Tokenizar oraciones
encoded_input = tokenizer(sentences, padding=True, truncation=True, return_tensors='pt')
# Calcular las embeddings de tokens
with torch.no_grad():
    model_output = model(**encoded_input)
# Realizar pooling. En este caso, mean pooling.
sentence_embeddings = mean_pooling(model_output, encoded_input['attention_mask'])
print("Sentence embeddings:")
print(sentence_embeddings)

Arquitectura Completa del Modelo

SentenceTransformer(
(0): Transformer({'max_seq_length': 512, 'do_lower_case': False}) with Transformer model: XLMRobertaModel
(1): Pooling({'word_embedding_dimension': 768, 'pooling_mode_cls_token': False, 'pooling_mode_mean_tokens': True, 'pooling_mode_max_tokens': False, 'pooling_mode_mean_sqrt_len_tokens': False})
)