Multi-Scale Architecture and Device-Aware Data-Random-Drop Based Fine-Tuning Method for Acoustic Scene Classification.