changelogs.md


shoprunner/octopod

Repository  -  API  -  Source

3.1.4

April 2, 2021

Added

  • PyPI version badge to README.md

3.1.3

April 2, 2021

Added

  • CI testing and documentation badges to README.md

3.1.2

March 31, 2021

Fixed

  • Rebuild reqirements.txt to use m2r2

3.1.1

March 31, 2021

Fixed

  • sphinx and m2r were no longer in requirements files so docs failed added sphinx and had to update to m2r2 because m2r is not actively supported
  • Changelog dates were still for 2020

3.1.0

February 3, 2021

Changed

  • Datasets can now take string or encoded labels using sklearn label encoders.

    Added

  • Tests for new functionality in the Dataloaders for image, text, & ensemble

3.0.1

January 13, 2021

Fixed

  • transformers was listed twice in setup.py

3.0.0

January 4, 2021

Added

  • Report smoothed training losses in progress bar during fitting.

    Changed

  • Report exponentially weighted moving average rather than simple average of training loss over batches at the end of each epoch.

2.2.6

December 31, 2020

Fixed

  • Added matplotlib dependency to display progress bars in notebooks
  • Removed parameter that no longer exists from MultiTaskLearner.fit docstring

    Changed

  • Use pip-tools

2.2.5

December 15, 2020

Fixed

  • Multi Task BERT model save works for models without new classifiers

2.2.4

December 14, 2020

Fixed

  • Load method in ensemble model loads to image_dense_layers

2.2.3

December 10, 2020

Fixed

  • Multi-label datasets are imported in top level init

2.2.2

October 29, 2020

Fixed

  • Update creevey to new name wildebeest

2.2.1

October 29, 2020

Added

  • note to notebook tutorials about potential out-of-memory issue when DataLoaders have too high a value for num_workers

    Changed

  • included Dockerfile now installs the octopod library

    Fixed

  • included Dockerfile now installs libgl1-mesa-glx to avoid ImportError: libGL.so.1: cannot open shared object file: No such file or directory when importing octopod
  • batch sizes of size 1 are skipped correctly for all models

2.2.0

August 20, 2020

Changed

  • Switched to using Github Actions for CI/CD

2.1.0

July 15, 2020

Changed

  • Torch is no longer pinned to version 1.2. This allows user to use Octopod with python 3.8.

2.0.2

July 14, 2020

Fixed

  • Support for torch.optim.lr_scheduler.ReduceLROnPlateau for scheduler argument in MultiTaskLearner.fit
  • Learner object skips and ResnetForMultiTaskClassification forward method automatically handles batches of size 1 to avoid ValueError with nn.BatchNorm1d failing on batches of size 1

2.0.1

July 14, 2020

Added

  • logo now in octopod readme

2.0.0

July 10, 2020

Added

  • Octopod learners can now use multiple loss functions and do multi-label in addition to multi-class tasks

    Changed

  • loss functions and metrics are specified via a dictionary of tasks and corresponding loss functions

1.0.0

July 9, 2020

Changed

  • Tonks is now called Ocotopod

==== Below is Tonks development ====

1.0.0

July 8, 2020

Added

  • Warnings to switch to new name

0.4.4

July 5, 2020

Fixed

  • A few paren locations in multi_task_ensemble.py

0.4.3

June 24, 2020

Added

  • Version pin for sentencepiece

0.4.2

June 23, 2020

Added

  • Dockerfile

0.4.1

June 16, 2020

Added

  • Missing 's' in extras_requires in setup.py

0.4.0

June 12, 2020

Added

  • warning about upcoming rename of library

0.3.0

April 14, 2020

Added

  • read the docs fixes

0.2.0

April 13, 2020

Added

  • devs requirements in setup.py

0.1.0

April 13, 2020

Added

  • Link to docs in README

0.0.0

April 13, 2020

Added

  • Open-sourcing Tonks codebase for multi-dataset multi-task learning