Home

Welcome to the doctr wiki!

Maintainer notes

Release process

• Check that the versions in setup.py & .conda/meta.yaml are correct
• Create a release on Github where the name and especially the tag matches the version specified in setup.py (e.g.: v0.11.0)
• After releasing apply post-release modifications ref.: https://github.com/mindee/doctr/commit/251aa22bdcac263767c6957c8f18c18bb2d20ffe

Then build the conda package and upload to anaconda:

conda build purge
conda build purge-all
rm -rf conda-dist/
BUILD_VERSION='X.Y.Z' python setup.py sdist
mkdir conda-dist
conda-build ./conda/ -c pytorch --output-folder conda-dist
ls -l conda-dist/noarch/*conda

anaconda upload conda-dist/noarch/*conda -u mindee

Training process

Classification & Orientation

Acc: 99% +

--epochs 40
- default config (`batch_size` adjusted to machine capacity)

Recognition

exact + partial: 99,9% +

--epochs 20
- default config (`batch_size` adjusted to machine capacity)

Detection

Precision: 82% ++ Recall: 76% ++ Mean IoU: 60% +

NOTE: polylr power is set to 0.5 instead of default 1.0

--epochs 25
-b 2
--rotation
--eval-straight (faster eval)
--save-interval-epoch
--optim adamw
--lr 0.001
--sched polylr