Skip to content

user_guide.md

Latest commit

 

History

History
105 lines (66 loc) · 6.31 KB

user_guide.md

File metadata and controls

105 lines (66 loc) · 6.31 KB

User Guide

Introduction

The Transformers Extension provides a Python library with UDFs that allow the use of pre-trained NLP models provided by the Transformers API.

The extension provides two types of UDFs:

  • Utility UDFs: Manage Transformers models in the Exasol BucketFS.
  • Prediction UDFs: Use a downloaded pre-trained model to perform prediction.

A Note on UDFs :

User Defined Functions (UDFs) are scripts in various programming languages that can be executed in the Exasol database. They can be used by a user for more flexibility in data processing. With the Transformers Extension, we provide multiple UDFs for you to use on your Exasol database. You can find more detailed documentation on UDFs on the UDF Scripts page.

UDFs and the necessary Script language Container are stored in Exasol's file system BucketFS, and we also use this to store the Hugging Face models on the Exasol cluster.

More information on the BucketFS can be found at docs.exasol.com.

For prerequisites and the setup-guide, please visit the Getting started file.

Utility UDFs

This Exasol Extension provides UDFs for interacting with Hugging Face's Transformers API to use pre-trained models on an Exasol cluster.

These UDFs deal with the installation and deletion of pretrained Transformers models in the Exasol BucketFS.

We provide the following UDFs:

UDF Name Use
TE_MODEL_DOWNLOADER_UDF Download a specific model from Hugging Face.
TE_LIST_MODELS_UDF List all models available in the BucketFS.
INSTALL_AI_DEFAULT_MODEL_UDF Install all default models used in the Transformers Extension.
TE_DELETE_MODEL_UDF Delete a specific model from the BucketFS.

You can find further information on these UDFs here.

Prediction UDFs

These UDFs call a model stored in the BucketFS and use it to make predictions on the given input data.

We have selected a curated list of models, which are used in our UDFs. You can install these models using the install_default_models- command or the INSTALL_AI_DEFAULT_MODEL_UDF. These UDFs require only minimal configuation to use:

UDF Name task_type Use
AI Sentiment text-classification Classifies the given text according the sentiment found in the text.
AI Classify zero-shot-classification This UDF classifies the input text into classes defined by the user.
AI Extract Entities token-classification This UDF finds tokens in a given text, and assigns a label to found tokens.

However, if you want to configure a task to your specific needs, UDFs with the suffix "Extended" in the name allow you to specify all available parameters.

You may specify different models/configuration parameters for each input row. Having many different models will however increase the processing time, as the models need to be loaded.

For example, you may want to select a specific model to be used.

These are the available UDFs:

UDF Name task_type Use
AI Custom Classify Extended text-classification Classifies the given text into classes known to the model.
AI Entailment Extended text-classification Takes two input texts and compares them.
AI Answer Extended question-answering Extracts answer(s) from a given question text.
AI Fill Mask Extended fill-mask Replace <mask> tokens in the input with predicted text.
AI Complete Extended text-generation Predict the continuation of the given text.
AI Extract Extended token-classification Find and label tokens in a given text.
AI Translate Extended translation This UDF translates a given text from one language to another.
AI Classify Extended zero-shot-classification This UDF classifies the input text into classes defined by the user.

Each UDF uses models for a predefined Transformers task. Models which do not support this task will perform poorly. So take care to match your selected model to the UDF for your desired task.

More details on the available parameters can be found in detailed documentation.

Common UDF Parameters

Many UDFs use a set of common parameters:

  • device_id: To run on a GPU, specify the valid cuda device ID. Value NULL means to use the CPU instead.
  • bucketfs_conn: The BucketFS connection name.
  • sub_dir: The directory where the model is stored in the BucketFS.
  • model_name: The name of the model to use for prediction. You can find the details of the models on the Hugging Face models page.

Common Output Behavior

Each of the UDFs generates an output containing the original input columns passed to the UDF plus additional columns containing the inference results.

In case of any error during model loading or prediction, the additional output columns are set to NULL and column ERROR_MESSAGE is set to the stacktrace of the error.

What's Next:

For information on managing the models, please visit the Manage Models in the BucketFS guide.

For information on invoking the models, please visit the guide on Using Prediction UDFs.