Stagewise Enlargement of Batch Size for SGD-based Learning