Classifying Variable-Length Audio Files with All-Convolutional Networks and Masked Global Pooling.