textualization / sentence-transphormers

Compute RoBERTa sentence embeddings in PHP using ONNX framework

Compute RoBERTa sentence embeddings in PHP using ONNX framework.

This brings the power of Sentence Transformers to the PHP world.

Installation

Add this project to your dependencies

composer require textualization/sentence-transphormers
composer update

Before using it, you will need to install the ONNX framework:

composer exec -- php -r "require 'vendor/autoload.php'; OnnxRuntime\Vendor::check();"

and download the Distill RoBERTa v1 ONNX model (this takes a while, the model is 362Mb in size):

composer exec -- php -r "require 'vendor/autoload.php'; Textualization\SentenceTransphormers\Vendor::check();"

Multilingual model

To use the multilingual model. Additional dependencies are needed.

Download the SentencePiece library:

composer exec -- php -r "require 'vendor/autoload.php'; Textualization\SentencePiece\Vendor::check();"

Download the XLM Tokenizer SentencePiece BPE model:

composer exec -- php -r "require 'vendor/autoload.php'; Textualization\Ropherta\Tokenizer\Vendor::check();"

Download the Multilingual-E5-small model (471Mb in size):

composer exec -- php -r "require 'vendor/autoload.php'; Textualization\SentenceTransphormers\Vendor::check(true);"

Please note: if you had downloaded the monolingual model you'll need to delete it first. Currently only one model is possible, this limitation will be lifted in future versions.

Computing embeddings

$model = new SentenceRophertaModel();

$emb = $model->embeddings("Text");

Check \Textualization\Ropherta\Distances to check whether two embeddings are closer to each other.

Model employed

The model being used is an ONNX export from sentence-transformers/all-distilroberta-v1, hosted at HuggingFace Hub: textualization/all-distilroberta-v1.

The multilingual model is intfloat/multilingual-e5-small, exported to ONNX by the authors.

sentence-transphormers's Issues

Test cases are needed

Some test cases running the models against some texts and comparing the output to Python output would be great.