Skip to content

Latest commit

 

History

History
63 lines (41 loc) · 2.62 KB

TODO.md

File metadata and controls

63 lines (41 loc) · 2.62 KB

  • For typical inverse problems where f(x) = fdata(x) + µ fprior(x), set convergebce criterion to be ‖∇f(x)‖ ≪ max(‖∇fdata(x)‖, |µ|⋅‖∇fprior(x)‖).

  • Rename HyperbolicTotalVariation as EdgePreservingSmoothness.

  • Force a cubic step for the initial step or after a restart in LBFGS.

  • Optimize methods like sum, increment, scan for multi-dimensional arrays when getdata() returns non-null.

  • mitiv.io.BufferedImage: most methods here are deprecated.

  • Add support for boolean arrays.

  • Have public methods of CostFunction and descendants check the arguments and short-circuit if alpha = 0, not the low level implementations.

  • Decide that byte are interpreted as being unsigned (there is no real use of signed byte in image processing and having both signed and unsigned version of integers is not worth the added complexity, and bugs), the cleanest way to do that is to change the get() method for byte arrays. Side effects: MDA format, etc. Note that this is similar to how is interpreted the TYPE_BYTE in DataBuffer objects. ALMOST DONE

  • Optimize methods for flat arrays as for the conversion routines.

  • Use ImageWriteParam and ImageReadParam to improve reading/writing of images. This can be set in FormatOptions.

  • Deal with NaN and Infinite when converting arrays to images. In particular with the initial image.

  • Implement broadcasting rules for assign.

  • Move mitiv.array.ArrayUtils.java into mitiv.utils.ArrayUtils.java.

  • Use varargs, e.g. int... dims to simplify the code of, e.g. Shape.java, and suppress the needs of TPP in some cases.

  • Support CFitsIO syntax for specifying a particular HDU or extension and a ROI: filename.fits[HDU][RNG1,RNG2,...] where HDU is the HDU number or extension name (automatically trimmed and converted to upper-case letters) and RNGn are ranges (possibly * or : for all).

  • Use WeightedData class in WeightedConvolution.

  • In FlatArray.javax and StriddenArray.javax there is no checking nor wrapping of the slicing index (contrarily to the dim arg).

  • Many operations such as slice, view, etc. could have their args tested at an higher level and a low level, e.g. _slice, method is then called whith different implementations.

  • In the cost functions, evaluate and computeCostAndGradient should check arguments? Otherwise, make them low level (with an underscore).

  • Make Flat{Double,Float}* and ShapedVector more intricate?

  • space.create() or vec.similar() to create a new vector.

  • Automatically build JavaDoc.

  • Rename toplevel package as tipi.