Bring gain of top-1 accuracy (69.54% -> 70.20%)

1. Quantize weights
2. Add quantize-dequantze nodes