Some improvements

[procyontao]

It seems that the software works.

There are still several things to improve:

1. patch edge artifact in deconvolution
2. Tilt angle offset and xtilt: The solution should be rotating subtomograms during 'isonet.py extract' using this two angles, along y and x respectively. During 'isonet.py predict', rotate entire tomograms then predict and rotate back.
3. Reduce overlap of subtomograms as much as possible: The solution might be generating a lot more more random coordinates and select a subset of them with largest min interparticle distance.
4. check grammar "The program crop the tomogram in multiple tiles (z,y,x) for multiprocessing and assembly them into one. e.g. (1,2,2)"