Soft Weight-Sharing for Neural Network Compression