Deep Compression: Compressing Deep Neural Networks with Pruning, Trained Quantization and Huffman Coding