Deep Compression of Pre-trained Transformer Models