README.md

metadata

pipeline_tag: sentence-similarity
language: en
license: apache-2.0
tags:
  - sentence-transformers
  - feature-extraction
  - sentence-similarity
  - transformers
  - onnx

This is the ONNX model of sentence-transformers/all-MiniLM-L6-v2 [https://seb.sbert.net]. Currently, Hugging Face does not support downloading ONNX model and generate embeddings. I have created a workaround using sbert and optimum together to generate embeddings.

pip install onnx
pip install onnxruntime==1.10.0
pip install transformers>4.6.1
pip install sentencepiece
pip install sentence-transformers
pip install optimum
pip install torch==1.9.0

Then you can use the model like this:

import os
from sentence_transformers.util import snapshot_download
from transformers import AutoTokenizer
from optimum.onnxruntime import ORTModelForFeatureExtraction
from sentence_transformers.models import Transformer, Pooling, Dense
import torch
from transformers.modeling_outputs import BaseModelOutput
import torch.nn.functional as F
import shutil

model_name = 'vamsibanda/sbert-onnx-all-MiniLM-L12-v2'
cache_folder = './'
model_path =  os.path.join(cache_folder, model_name.replace("/", "_"))

def download_onnx_model(model_name, cache_folder, model_path, force_download = False):
    if force_download and os.path.exists(model_path):
       shutil.rmtree(model_path)
    elif os.path.exists(model_path):
       return 
    snapshot_download(model_name,
                     cache_dir=cache_folder,
                     library_name='sentence-transformers'
                     )
    return
      
def mean_pooling(model_output, attention_mask):
    token_embeddings = model_output[0] #First element of model_output contains all 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)

def generate_embedding(text):
    token = tokenizer(text, return_tensors='pt')
    embedding = model(input_ids=token['input_ids'], attention_mask=token['attention_mask'])
    embedding = mean_pooling(embedding, token['attention_mask'])
    embedding = F.normalize(embedding, p=2, dim=1)
    return embedding.tolist()[0]                    
     
_ = download_onnx_model(model_name, cache_folder, model_path)                
tokenizer = AutoTokenizer.from_pretrained(model_path)
model = ORTModelForFeatureExtraction.from_pretrained(model_path, force_download=False)
pooling_layer = Pooling.load(f"{model_path}/1_Pooling")

generate_embedding('That is a happy person')