README.md

---
language: en
tags:
- feature-extraction
- onnx
- use
- text-embedding
- tensorflow-hub
license: apache-2.0
inference: false
widget:
- text: Thank goodness ONNX is available, it is lots faster!
---
### Universal Sentence Encoder Large v5

ONNX version of [https://tfhub.dev/google/universal-sentence-encoder-large/5](https://tfhub.dev/google/universal-sentence-encoder-large/5)

The original TFHub version of the model is referenced in other models here E.g. [https://huggingface.co/vprelovac/universal-sentence-encoder-large-5](https://huggingface.co/vprelovac/universal-sentence-encoder-large-5) 

### Overview

See overview and license details at [https://tfhub.dev/google/universal-sentence-encoder-large/5](https://tfhub.dev/google/universal-sentence-encoder-large/5)

This model is a full precision version of the TFHub original, in ONNX format.

It uses the [ONNXRuntime Extensions](https://github.com/microsoft/onnxruntime-extensions) to embed the tokenizer within the ONNX model, so no seperate tokenizer is needed, and text is fed directly into the ONNX model.

Post-processing (E.g. pooling, normalization) is also implemented within the ONNX model, so no separate processing is necessary.

### How to use

```python
import onnxruntime as ort
from onnxruntime_extensions import get_library_path
from os import cpu_count

sentences = ["hello world"]

def load_onnx_model(model_filepath):
  _options = ort.SessionOptions()
  _options.inter_op_num_threads, _options.intra_op_num_threads = cpu_count(), cpu_count()
  _options.register_custom_ops_library(get_library_path())
  _providers = ["CPUExecutionProvider"]  # could use ort.get_available_providers()
  return ort.InferenceSession(path_or_bytes=model_filepath, sess_options=_options, providers=_providers)

model = load_onnx_model("filepath_for_model_dot_onnx")

model_outputs = model.run(output_names=["outputs"], input_feed={"inputs": sentences})[0]
print(model_outputs)
```