Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

[v0.2.0] - 11/11/2020

Added

Scripts to register results of any model as a run in any AML experiment source MR
Ability to load embeddings using azure ml dataset AML177 MR
Calibration on training set by default AML177 MR
Log metrics to Azure ML AML177 MR
Output calibration set predictions on optimal threshold AML177 MR
AML configuration and scripts to train models on azure
Experiment tags AML177 MR
RunID in slack message upon experiment completion AML177 MR

Changed

Copy embedding to model dir on serialization !1
Calibration tqdm update progress bar every 1/10th time AML177 MR
Avoid uploading accuracy tables because they don't load in UI AML177 MR
Moved aml.docker to respective cpu, gpu configs AML177 MR
Renamed val_path to valid_path AML177 MR

Removed

Gensim dependency, instead use numpy matrix to load pretrained embeddings AML177 MR
Unwanted nested checkpoint saving dirs lightning_logs/version_0 AML177 MR

[v0.1.1] - 07/15/2020

Added

Label smoothing
Label squeezing
Add new model that uses 1d convolution instead of 2d increasing performance by 46% and reduced training time from 3h to 2h for 360k records.
Auto build sphinx docs and push to web server using CI
Add gitlab badges for test coverage

Changed

Make tuning on training set default instead of tuning set
Moved tests directory outside of source directory
Moved one-off scripts into their own "scripts" module from "vayu"

Removed

Configurable truth configs because they are detrimental to performance
Task yaml because it was not needed
requirements.txt