- Tutorial on image augmentations for computer vision: img-augmentation-demo.ipynb
- Tutorial on image and text augmentations for ML: img-text-augmentation-demo.ipynb
Both tutorial notebooks use the albumentations
package to do standard geometric, colorspace, kernel filter, and
random erasing (Cutout) transformations. Gives examples of image mixing
(Mixup and CutMix) augmentations, and combining multiple image augmentations together.
The image-specific notebook, img-augmentation-demo.ipynb
, concludes with examples of
techniques applied by top teams in the Kaggle Recursion Cellular Image
Classification
Machine Learning competition.
The second tutorial notebook,img-text-augmentation-demo.ipynb
, includes most of the former content on image augmentations, but also dives into text augmentations using the nlpaug
package.
To try out the notebooks, you can create a working conda
enviroment from the provided environment file, activate it, and launch jupyter notebook:
$ conda create env -f environment.yml
$ conda activate augment
(augment) $ jupyter notebook
Note that if you are only interested in running the image augmentations, you don't need most of these dependencies and can just pip install albumentations
to get started.
The notebook is intended to be rendered as slides using the RISE extension, which can be configured in the Jupyter Notebook Extensions UI.
The image augmentation talk was originally created for a presention to the SLC Python Meetup on June 3, 2020. A more concise version of it was presented at PluralsightLIVE 2020, and can be accessed on the pslive-version branch.
The image and text augmentations tutorial was presented at the Women in Data Science Conference, Salt Lake City chapter, on April 2, 2022.