TODO

1. Remove all duplication caused by porting from C.
2. Add faster reader.
3. gzip support.
4. Move dual problems with based on dot-product to Kernel infrastucture.
   Kernel are not normal kernels, but a thing that computes <w, phi(x)> and w + phi(x) w/o transforming the original dataset.
5. Integration with mallet
6. Poly-2 “kernel” support
7. Hash “kernel” support