image-augmentation-ml

Finding the optimal set of image augmentations for image classification.

Installation Instructions

First, ensure that you have the latest build tools installed on your machine:

python3 -m pip install --upgrade build
python3 -m build

Then, you can install the image-augmentation-ml package as follows:

python3 -m pip install -e .

Development Tools

Install the pre-commit checker by running the following:

python3 -m pip install pre-commit
pre-commit install

From this point forward, whenever you commit to this repository, an autoformatter (black) and a linter (flake8) will run. If either come up with errors, you must fix them, add them to the staging area, and commit again.

Testing Instructions

You can run the test suite with the following command:

tox

Imageset Download

You can download the imageset using the following command.

make img_download

To specify the image size (either 160px or 320px), add on the PX flag. By default, this command will download the 160px dataset.

make img_download PX=320

Image Augmentation

The set of augmented images can be created by running the augment script. To create the entire set of augmented images, you can run the following command:

python scripts/augment.py configs/augmentation/all.yaml

Image Classification

To train an image classifier, you can run the following command. This command will train an image classifier on the set of all augmented images:

python scripts/classify.py configs/augmentation/all.yaml configs/classifier/default.yaml

Classifier Evaluation

To evaluate the results of a classifier, you can run the following command. This command will evaluate the 30th epoch of the classifier that trained on all augmented images:

python scripts/evaluate.py configs/augmentation/all.yaml configs/classifier/default.yaml --load_epoch 30

Then, to visualize the results, you can plot them as follows:

python scripts/plot_results.py