Add keep_aspect_ratio flag to flow_from_directory

[qdbp]

I had a collection of images with non-standard, non-square sizes, and found that ImageDataGenerator would stretch them significantly.

I decided to add a keep_aspect_ratio kwarg to ImageDataGenerator.flow_from_directory to allow keeping images at their original aspect ratio, padding them to target_size with the value specified by the new cval kwarg.

#3338

[fchollet]

Closing outdated PR. If you still care about the content of the PR, please submit a new PR to master, updated for the Keras 2.0 API.

[paulo-raca]

Any particular reason why it has never been merged?
The code is very straightforward, and I believe many other have the same issue. (Well, I do 🙋‍♂️ )

If there is a better way of doing this, please let me know -- Otherwise, I can go on and make an updated pull request.

[small-yellow-duck]

I agree that having a keep_aspect_ratio flag is a good idea and I would like to see another PR opened.

[mark-meyer]

I also agreed that this would be nice to have. When dealing with classifying images it is easy to images scenarios where deforming the images id going to be a problem. Imagine a image-based iris dataset where sepal width and height are important but you are starting with inputs of different aspect ratios.

[EliasHasle]

There are even datasets where rescaling seriously breaks the data. See the ALASKA steganalysis contest.

I would expect there to be an option of padding/cropping instead of resampling, and then the fill_mode could apply naturally.

[codingpy]

Hi, I think this functionality doesn't work when the original image is smaller than the target image. But appreciate your great work. @qdbp

[jsl303]

@xunnv, then would it be possible to enlarge smaller images as much as possible to keep the aspect ratio and then add padding?