linkorgsonnx

linkorgsonnx is an R package that allows you to join datasets using neural networks in R.

Installation

Prerequisites:

To use the development version of the package, you must have the following installed:

1. Cargo, the package manager for Rust
2. The ONNX Runtime, needed for running the neural networks.

Installation for both should require copying a few commands in your terminal. If you are looking to run the models on a GPU, please make sure that you install the relevant GPU version. Please see the following note reguarding a common issue compiling packages using Rust on Windows.

Installing Rust on Windows:

To install Rust on windows, you can use the Rust installation wizard, rustup-init.exe, found at this site. Depending on your version of Windows, you are likely to see an error that looks something like the following when you try and compile the package:

error: toolchain 'stable-x86_64-pc-windows-gnu' is not installed

In this case, you should run rustup install stable-x86_64-pc_windows-gnu to install the missing toolchain. If you’re missing another toolchain, simply type this in the place of stable-x86_64-pc_windows-gnu in the command above.

Installing from GitHub

At the moment, the only way to install this package is from this github repository, although we hope to have pre-built binaries available on CRAN / R-Universe soon.

devtools::install_github("beniaminogreen/linkorgsonnx")

Obtaining Models From HuggingFace

We have designed linkedorgsonnx to be compatible with most of the sentence-similarity models available on HuggingFace, so browsing the sentence similarity task is a great place to start.

Huggingface distributes a tool called Optimum that can be used to download models from the huggingface hub and convert them into the ONNX format. The code below can be used to download the all-MiniLM-L6-v2 model (the most popular model trained for this task), convert it to ONNX, and save it to ./mini_lm_l6/.

# pip install --upgrade --upgrade-strategy eager optimum[onnxruntime]

# If using CUDA GPU
optimum-cli export onnx --model "sentence-transformers/all-MiniLM-L6-v2" mini_lm_l6/ --optimize O4 --device cuda --task sentence-similarity


# If using CPU
optimum-cli export onnx --model "sentence-transformers/all-MiniLM-L6-v2" mini_lm_l6/ --optimize O3 --device cpu --task sentence-similarity

Generating Embeddings

Now that you have your hands on a model, it should be easy to generate embeddings for text inputs. Simply point linkorgsonnx to where the ONNX Runtime is installed, load the model into memory, and get to generating embeddings!

# Point Package to where the ONNX Runtime is installed
Sys.setenv(ORT_DYLIB_PATH = "path/to/libonnxruntime.so")

# Loads model from a given directory
model <- new_inference_session("mini_lm_l6/")

inputs <- c(
            "The quick brown fox jumped over the lazy dog",
            "Sphynx of black quartx, judge my vow",
            "The five boxing wizards jump quickly"
)
generate_embeddings(model, inputs)

Neural Joins

The flagship feature of linkorgsonnx is it’s capacity to join datasets on the basis of neural network embeddings.

To (inner) join two dataframes on the basis of a single column, you can run the following code, which generates embeddings for the name_1 column in dataset 1 and name_2 in dataset, and returns pairs that have a cosine distance less than .1 in the embedding space.

neural_inner_join(dataframe_1, dataframe_2, model, by = c("name_1", "name_2"), radius = .1)