R Deep Learning Essentials - Second Edition

Test time augmentation

We can also use data augmentation during test time. In the augment.R file, it created a file with the original test set of 4,200 rows (data/test0.csv), which was used that to evaluate the model. The augment.R file also created a file called test_augment.csv, which has the original 4,200 rows and 2 copies for each image. The copies are similar to what we did to augment the training data, that is, a row with data rotated 15 degrees left and a row with data rotated 15 degrees right. The three rows are outputted sequentially and we will use these 3 rows to vote for the winner. We need to take 3 records at a time from test_augment.csv and calculate the prediction value as the average of these three values. Here is the code ...