Binary learning with --disk learning

[4ound]

Hey! We have big dataset for training (few TB), so we use your library with --disk learning option. We spend significant time on dataset parsing. On practice, if generate binary file and start RAM learning (with --no-bin and without), second variant would be faster. While on disk mode doesn't use/generate binary file, why?

[aksnzhy]

@4ound The binary file could beyond you RAM capability, so we only generate binary file for in-memory training.