Create data object (for ConvLearner ) using custom data generator

I am wondering if there is any way to have ConvLearner working with custom data generator. So that learner trains on data generated on the fly instead of reading images from disk.

Working on a Kaggle competition that has 50 millions images with drawing strokes stored in CSV format.
Save all training images to files on disk would take up about 0.5 terabytes of disk space.

Did you mange to find a solution?