Add ability to duplicate augmentor_images

[scott-zockoll]

Instead of just being able to use all of the original images (n=0), this allows cloning the original images (by using a negative number) before applying the transformations.

This is my first time asking for a pull request on GitHub. This is a very small change. When I was using this package in my project I wanted a way to scale my original dataset, but also keep the data even over all the categories. In sample(), you specify the number of samples you would like to generate (the n parameter), but this is stochastic and does not keep the categories even. You can also use n=0 apply transformations on every image in the original dataset, but I wanted multiple transformations on the same image (say I was scaling my dataset 3x). I added a couple lines for the ability to specify negative numbers for n. The absolute value of n is the amount you wish to scale the dataset before applying the transformations.

Again this is my first pull request, please let me know what I did wrong.

[mdbloice]

Hi @scott-zockoll that is not a bad idea, and actually that exact feature was actually default behaviour in a previous version of Augmentor, as the images were not randomised (basically it would iterate over all images repeatedly until n samples were generated). If I remember correctly a few people complained about that, and didn't like this behaviour so I changed it to be random. I think having the option to not do this randomly makes sense, and one could provide a scale parameter instead of a n value. I think it would be best if we made a new function for this rather than making the sample() any way ambiguous. Perhaps instead we should decide on a good name for this feature, perhaps something like scale(factor) or something?