Image augmentation library working with TPU

Image augmentation is an important technique to improve model accuracy.
Many people use libraries such as albumentations and imgaug to apply various augmentations, but since these libraries are created using non-TF libraries, you need to use tf.numpy_function to use them in tf.data.
However, this approach cannot work in a TPU Node environment such as Google Colab. Google Colab TPU is often used in competitions such as kaggle because it is very strong and inexpensive.
So I have created a new image augmentation library that works with TPU! I have actually used it in a kaggle competition and it works!
Please give it a try!

library: GitHub - hirune924/imgaug-tf: An image augmentation library for tensorflow.
kaggle solution: UW-Madison GI Tract Image Segmentation | Kaggle

1 Like

This is very inspiring. I will do well to try it out and provide feedback if necessary.

1 Like

@hirune924 san, this is really great and useful library.

FYI, you may want to look at the recent keras-cv library which provides many augmentations and is actively developing. I think it would be great if you share some insight and feedback on it, HERE.

2 Likes

@innat Yes I found it recently. It is great with a good variety of transforms for batch. It would be interesting to use a combination of both as there are many transforms that are not duplicated compared to imgaugtf

1 Like