Proper test-time data augmentation in production usually does not apply the same augmentation pipeline as during training. In production, the computational budget is usually limited and we only have time to run inference on a few augmented images.
Therefore, the random data augmentation that works well during model training is not ideal. It is important to get the highest accuracy boost with only a few augmented samples. In practice this means that the best test-time data augmentation has to be found using experimentation
Run the Colab notebook to experiment for yourself and learn how to find the best augmentation parameters.
Don't forget to tag @dufourpascal in your comment, otherwise they may not be notified.