The variations that the training process can use to generate more training data from the training data you provide.
- macOS 10.14+
- Xcode 10.0+
- Create ML
Augmentation generates new images from the training data you supply to increase the amount of training data available to the model. The number of augmented images per original depends on the specific option you choose. For example, if you use
rotation, you get four new images with a random rotation angle for each original. If you use
flip, you get three new images: one that is flipped horizontally, one flipped vertically, and one flipped along both axes.
Augmentation options combine multiplicatively to produce potentially very large data sets. If you use both of the options described above, you get 4 × 3 = 12 images: each of four rotated images is flipped three ways. Adding more options further multiplies the data set size by the number of variants for that option.
To keep the training time manageable, the image classifier restricts the total number of augmentation combinations to 100. The classifier applies the augmentation options you specify until applying more would cause the total count to exceed that threshold. The classifier then restricts the number of augmented variants for any further options. For example, if the first three options you use have four variants each, resulting in 4 × 4 × 4 = 64 combinations, any further options are limited to one variant to avoid exceeding 100 total combinations.
To optimize performance in the face of this restriction, when you use more than one option, the classifier applies them in the order of greatest to least training effectiveness. Specifially, it uses this ordering:
See Improving Your Model’s Accuracy for a discussion about when to use augmentation.