Skip to content

Releases: exasol/transformers-extension

0.6.0: Cleanup

27 Nov 14:33
42ae75c
Compare
Choose a tag to compare

Summary

Bug fixes regarding SLC download and input validation. On to of that dependency and documentation updates.

Bug Fixes

  • #134: Fixed slc download still assuming existence of slc-parts
  • #124: Fixed input validation in extract_unique_model_dataframes_from_batch

Documentation

  • #132: Fixed outdated information in documentation

Security

  • #148: Security updates of dependencies
  • #136: Removed configobj

0.5.0: Support for transformers v.4.31.0

12 Sep 13:18
1321af1
Compare
Choose a tag to compare

Summary

This release makes the extension compatible with Huggingface transformers v.4.31.0 and their new model cache format.
Furthermore, it makes the deployment scripts compatible with Exasol v8 by enabling encryption
and allows the user to configure the TLS verification.

Features

  • #88: Added custom matcher functions for unit tests
  • #103: Added option to toggle use of TLS certificate validation for Database connection
  • #42: Update transformers to 4.31 and adapt the model uploader

Bug Fixes

  • #89: Fixed the content of error code config file
  • #100: Enabled encryption for all pyexasol connection to be compatible with Exasol 8
  • #84: Reactivated test after move to AWS
  • #128: Fix release workflow and remove splitting the SLC

Refactorings

  • #24: Added model counters to unit tests of prediction UDFs
  • #95: Removed setup.py
  • #107: Use SLCT api for building the language container
  • #108: Use itde pytest plugin for tests
  • #110: Splitted SLC into dependency and release build step
  • #8: Moved CI-tests to AWS
  • #115: Refactored ModelDownloaderUDF
  • #121: Use matchers in without db integration tests

Documentation

  • #93: Added the Developer Guide
  • #126: Add documentation for token for private models in prediction UDFs

Added Zero-Shot model and error handling structure

31 Mar 07:18
71c6611
Compare
Choose a tag to compare

Summary

This release introduces a new UDF script for Zero-Shot text classification. Moreover, this version enables users to use custom models located in local filesystem or private repositories. In addition, this release includes an error handling mechanism to handle errors that may occur during model loading or one of the prediction stages.

Features

  • #11: Converted DownloadUDF to SET UDF
  • #58: Added setup to upload models from local filesystem
  • #47: Added rank column to model results returning top-k predictions
  • #72: Added authentication token to download private models
  • #64: Added Zero-Shot test classification
  • #25: Added error handling structure

Documentation

  • #87: Updated User Guide with error_message column

Simplify language container setup and fix torch version 1.11.0

03 Feb 15:07
d7bae80
Compare
Choose a tag to compare

Summary

This version allows users to install the language container in one step. The version of language container to be installed is given to the installation script, therefore there is no need to download the container file separately. Moreover, this release fixed torch version to 1.11.0, so that we avoided to package unused nvidia dependencies existing later versions of torch.

Features

n/a

Bug Fixes

n/a

Refactoring

  • #76: Updated torch version to 1.11.0
  • #73 : Deployed SLC in one step

Documentation

n/a

0.2.0: Bug fix and user guide improvements on top of 0.1.0

28 Oct 10:58
dcd2133
Compare
Choose a tag to compare

Summary

This release fixes the script deployment. Previously, it failed when it can't create the schema.
Now it tries to open the schema. Furthermore, this release also improves the user guide.

Features

n/a

Bug Fixes

  • #62: Fixed that ScriptsDeployer fails when it can't create the schema

Refactoring

n/a

Documentation

  • #60: Updated the User Guide

Add a downloader UDF and a set of prediction UDFs using the transformers API

26 Sep 08:43
ffdc33e
Compare
Choose a tag to compare

Summary

This is the initial release of the transformers-extension which provides a
downloader UDF that allows us to store the pre-trained machine learning model
by transformers and a set of Prediction UDFs that allow the downloaded model
to be cached and used through the transformers API.

This version provides the following machine learning tasks:

  • Sequence Classification
  • Question Answering,
  • Filling Mask
  • Text Generation
  • Token Classification
  • Text Translation.

Features

  • #1: Added the initial setup of the project
  • #5: Prepared the skeleton of the project
  • #4: Added model downloader UDF
  • #9: Created sequence classification UDF for single text
  • #14: Created sequence classification UDF for pair text
  • #16: Created question answering UDF
  • #21: Added parameter specifying GPU device
  • #26: Created masked language modelling UDF
  • #29: Created text generation UDF
  • #31: Created token classification UDF
  • #28: Added top_k result returning feature to question answering UDF
  • #33: Added text translation UDF
  • #48: Prepared the initial release
  • #56: Split language container into parts to upload release artifacts

Bug Fixes

  • #2: Renamed master branch to main
  • #18: Corrected model filtering in prediction UDFs
  • #50: Fixed release_droid configuration
  • #52: Reduced disk space used by the machine during releasing

Refactoring

  • #12: Updated method for generating bucket udf path
  • #35: Setup masked language modelling pipeline once
  • #20: Applied same API call across all prediction UDFs
  • #19: Inherent prediction UDF classes from same base class
  • #54: Updated exasol-udf-mock import

Documentation

  • #44: Added User Guide