I have a question about the aug_transform function, does this function generate new samples? If so, what parameter indicates how many new samples it generates?
If it does not generate new samples, it modify the images in the dataset?
I have read the documentation and I can’t find where it is explained.
(I am using the function ImageDataLoaders.from_csv)
The function takes the original images and modifies them randomly. The modifications are different for each batch.
In your above example, both, affine and lighting transformations are applied with a 75% chance. Affine transformations are flipping or warping, lighting is lighting and contrast.
You can also control to what extend the images can be modified. In your example, images are zoomed in but not zoomed out, as you set
min_zoom to 1.0 and
max_zoom to 1.1. That means that your images can be zoomed to every value between 1.0 and 1.1.
It is also possible, that multiple transformations are applied to the same image, such as flipping and zooming. So there is a large number of possible modifications which are applied to the images.
Thank you BresNet! This information is great and it helped me to understand the aug_transform function