---
license: apache-2.0
---


```
##How to download the model?
!sudo apt-get install git-lfs
!git lfs install

# Then
!git clone https://huggingface.co/vamsibanda/sbert-onnx-all-roberta-large-v1
```

## 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]])
```