# When augmenting data, shall the dataset keep a balanced ratio?

When training a model it is more and more common to augment data. posts indicate that only the training set shall be augmented. On the other hand, it is common to split datasets in a fashion following ratios like 70% (train), 15%(validation), 15% (test)

My question is:

• When using augmentation techniques, shall this ratio still be respected after augmentation (meaning that the number of items included in validation and test determines the augmentation ratio)
• or should the dataset be split before the augmentation process (meaning that dataset ratios are unbalanced)?

Any publications regarding this topic?

Using your ratios, it would go something like this.

2. Select $$700,000$$ images for training, another $$150,000$$ for validation, and the final $$150,000$$ for test.

3. Augment the $$700,000$$ training images to give yourself $$700,001$$ or $$700,000,000$$ images in the training set.

4. Train on the $$700,001$$ or $$700,000,000$$ training images.

5. Validated and test your model on the validation set of $$150,000$$ and test set of $$150,000$$, respectively.

You have your data. Create a training set. Now pretend nothing but the training set exists. Augment the training set; train your model on the augmented data. Validate and test on the holdout data sets.

• Not sure that I understood the answer (sorry..) Is that the way I should build my augmentation? Or are you pointing an aberration in my question (which i don't see)? Commented Oct 20, 2021 at 19:51
• You have your data. Create a training set. Now pretend nothing but the training set exists. Augment the training set; train your model on the augmented data. Validate and test on the holdout data sets.
– Dave
Commented Oct 20, 2021 at 19:53
• My intuition - The motivation to augment is to improve the fitting process(reduce variance, etc..). No logic in augmenting the test/validation data. Commented Jan 11 at 12:30