Training Acceleration of Low-Rank Decomposed Networks using Sequential Freezing and Rank Quantization