C2W2 Assignment - Why augment data if results degrade?

Course Q&A TensorFlow: Data and Deployment Specialization Device-based Models with TensorFlow Lite
1

Hi,

In C2W2 we learn, among many other things, how to augment image data in order to avoid over fitting. In the Assignment exercise, we use the cats and dogs data to test the previously learnt principles. The accuracy I get after 15 epochs and using the same model is:

  • Without augmentation: Training = 98% and Validation = 89%
  • With augmentation: Training = 76% and Validation = 82%

Given the above results, why should we augment the training set since the results degrade on the validation set? In the end the result on the validation is what counts no?

Could it be that, due to the augmentation, the training set becomes more general than the test set? The model tries to capture this generalization without success resulting in a degradation of the results on the validation set.

Could a solution be to augment the validation set in the same way as the training set? Most information I find on the internet however says that augmenting the validation set is very uncommon.

Any thoughts?

Best regards,

Wouter

2

Hello there,

All the set’s metrics count, its an iterative chained process.

This sounds about right.

The purpose of the validation set is to test and adapt by changing hyperameters. The main goal is to fit the training dataset at best possible, since that has the most examples. If for example you augment the entire dataset and you choose a part for validation that might be better I think, as long as that part is representative of the entire distribution. The danger is that if you split a part for validation and augment it, it might drift way from original dataset.