---
license: apache-2.0
---

## Convert pytorch model to onnx format.

import torch
import onnx
import onnxruntime
from onnxruntime import InferenceSession
from transformers import RobertaTokenizer, RobertaModel
from transformers.convert_graph_to_onnx import convert
import numpy as np
from onnxruntime.transformers import optimizer
from pathlib import Path
from onnxruntime.quantization import quantize_dynamic, QuantType
from sentence_transformers import SentenceTransformer, util

sbert = SentenceTransformer('sentence-transformers/all-roberta-large-v1')
sbert.save('sbert-all-roberta-large-v1')

tokenizer = RobertaTokenizer.from_pretrained('sentence-transformers/all-roberta-large-v1')
model = RobertaModel.from_pretrained('sentence-transformers/all-roberta-large-v1')
model.save_pretrained('./all-roberta-large-v1/')
tokenizer.save_pretrained('./all-roberta-large-v1/')

opt_model_path = "onnx-model/sbert-roberta-large.onnx"
convert(framework='pt', model='./all-roberta-large-v1/', output= Path(opt_model_path), opset=12, use_external_format=False, pipeline_name='feature-extraction')

quantize_dynamic(
    model_input='onnx-model/sbert-roberta-large.onnx',
    model_output='onnx-model/sbert-roberta-large-quant.onnx',
    per_channel=True,
    reduce_range=True,
    activation_type=QuantType.QUInt8,
    weight_type=QuantType.QInt8,
    optimize_model=False,
    use_external_data_format=False
) 
##Copy pooling layer and tokenizer files to the output directory


## How to generate embeddings?

from onnxruntime import InferenceSession
import torch
from transformers.modeling_outputs import BaseModelOutput
from transformers import RobertaTokenizerFast
import torch.nn.functional as F
from sentence_transformers.models import Transformer, Pooling, Dense

class RobertaEncoder(torch.nn.Module):
    def __init__(self, encoder_sess):
        super().__init__()
        self.encoder = encoder_sess

    def forward(
        self,
        input_ids,
        attention_mask,
        inputs_embeds=None,
        head_mask=None,
        output_attentions=None,
        output_hidden_states=None,
        return_dict=None,
    ):

        encoder_hidden_state = torch.from_numpy(
            self.encoder.run(
                None,
                {
                    "input_ids": input_ids.cpu().numpy(),
                    "attention_mask": attention_mask.cpu().numpy(),
                    
                },
            )[0]
        )

        return BaseModelOutput(encoder_hidden_state)

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 sbert_onnx_encode(sentence_input):
    token = roberta_tokenizer(sentence_input, return_tensors='pt')
    encoder_outputs = encoder_layer(input_ids=token['input_ids'], attention_mask=token['attention_mask'])
    sbert_embeddings = mean_pooling(encoder_outputs, token['attention_mask'])
    sbert_embeddings = F.normalize(sbert_embeddings, p=2, dim=1)
    return sbert_embeddings.tolist()[0]

roberta_tokenizer = RobertaTokenizerFast.from_pretrained('sbert-onnx-all-roberta-large-v1')
encoder_sess = InferenceSession('sbert-onnx-all-roberta-large-v1/sbert-roberta-large-quant.onnx')
encoder_layer = RobertaEncoder(encoder_sess)
pooling_layer = Pooling.load('./sbert-onnx-all-roberta-large-v1/1_Pooling/')

m1 = sbert_onnx_encode('That is a happy person')
m2 = sbert.encode('That is a happy person').tolist()
print(util.cos_sim(m1,m2))
##tensor([[0.9925]])