We have conducted a fair benchmark of several augmentation libraries by
comparing how many images they process per second. In this benchmark, we measured
the transform itself, as well as the conversion to torch.Tensor, and
a subtraction of the ImageNet mean.
Here is how you can run the benchmark yourself on a validation set from ImageNet resized to 256x256x:
export DATA_DIR="<PATH to ImageNet val>"
conda env create -f benchmark/augbench.yaml
conda activate augbench
pip install git+https://github.com/MIPT-Oulu/solt@master#egg-name=solt
pip install -e benchmark
python -u -m augbench.benchmark -i 500 -r 20 --deterministic --markdown