Neural Network Compression via Low Frequency Preference