7
$\begingroup$

I would like to use a small dataset to create CNN model. So, I am using data augmentation to increase the train dataset. Should I use all augmentation techniques (arguments) that listed here?

I have noticed that adding many arguments decrease the accuracy of the model and make the training set harder than the testing set.

What is the best practices to use data augmentation when use flow_from_directory?

$\endgroup$

1 Answer 1

6
$\begingroup$

Augmentations often rely on the nature of your data. Imagine if an the result of an augmentation would be logical in your context.

For example let's say you have a cats vs dogs dataset. The images here can be flipped left to right. On the other hand in the MNIST dataset it makes no sense to flip the images at all (what good would it be feeding a flipped '3' to your model). Some augmentations might actually confuse the model (e.g. an upside down '9' might look like a '6').

Furthermore, if augmetation is making your model worse, try smaller augmentations (e.g. stick to affine transformations with small ranges - $[-5%, +5%]$ rotation/translation/scaling).

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.

Not the answer you're looking for? Browse other questions tagged or ask your own question.