Keras: How to use data augmentation with regression?

[ademenet]

Hi!

I'm quite new to Keras - and machine learning. I used to practice deep learning with Lasagne and attempt Kaggle Facial Keypoints Detection competition. I would like to improve my score by using Keras - and learn how to use it.

For this I need to solve regression problem in order to set several keypoints. But, I need data-augmentation (rotations, vertical flip, contrast jittering) in order to avoid overfitting. I couldn't be able to figure out how to apply real-time data augmentation on inputs and ground truth (e.g. if I rotate my image, I need to re-compute keypoints new coordinates).

I tried several recipes, using fit_generator or flow. But none of them work. Here my last attempt using fit_generator : https://gist.github.com/moannuo/98c40f18e73749b32d45563ce5081cd1

Is someone succeed in integrating data augmentation for both inputs and ground truth values? I keep looking for answers, if you have tips I would be grateful.

Thank you very much! 😄

[ademenet]

Ok, I figured it out my self. I used a generator with fit_generator. The tricky thing was to define a good generator that is able to send inputs and targets no matter the batch size is. I'm not sure I made myself clear. Anyway, if some people wan't to have a look to a solution: look at my repository.