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.
- 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
- 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
tovalid_path
AML177 MR
- Gensim dependency, instead use numpy matrix to load pretrained embeddings AML177 MR
- Unwanted nested checkpoint saving dirs
lightning_logs/version_0
AML177 MR
- 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
- 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"
- Configurable truth configs because they are detrimental to performance
- Task yaml because it was not needed
- requirements.txt