Performance Tips


The regularizers are built during calling with metaprogramming. Every time you call TV() it creates a new version which is evaluated with eval. In the first deconvolution routine it has to compile this piece of code. To prevent the compilation every time, define

reg = TV()

which is then later used as a variable. In a notebook or REPL environment just define it in a different cell.

No Regularizer

Often the results are good without regularizer but then need to be early stopped (e.g. like iterations=20). This increases the performance drastically, but might lead to more artifacts in certain regions.



You can also try to adjust the settings of the L-BFGS algorithm of Optim.jl Try to change m in opt=LBFGS(linesearch=BackTracking(), m=10). m is the history value of the L-BFGS algorithm. Smaller is usually faster, but might lead to worse results. See also Wikipedia.

L-BFGS uses LineSearches.jl. In our examples BackTracking turned out to be the fastest, but it might be worth to try different ones.


Try to set the keyword iterations=20 to a lower number if you want to early stop the deconvolution. Of course, the results might be worse then.