how does the quantization book avoid trivial solution?

I tried training my own module, but the quantization stage always collapsed into trivial solution, i.e. all features are indexed to the same entry, I use all the same training procedure, has anyone come across this issues? or it is just me configuring something wrong?