Release v0.7.7 · unfoldtoolbox/Unfold.jl

Unfold v0.7.7

new structured backbone for the solver. Added ~6-7 new solvers. Split up the solver into prepare & solve functions that can be chained / modified easily. Added a reference document for exactly this.
major bug: likelihoodratiotest was calculating wrong results (it was calculating a wrong likelihood). This lead to a breaking change in API, where now the data need to be provided as well.
Bumped JLD to 0.5 - this means one cannot save files now and load it in older Unfold versions, but old unfold version files can still be loaded
several small docstring fixes
new CUDAextension, allowing to provide CuArray to automatically choose a better solver (still needs to be tested more though)
fit now has an argument that can be set fit=false to not actually run the ´fit!` function, to do it later
benchmarks now available comparing ~7 solver types with and without GPU, some with and without Float32. Improvements of up to 10x of the SOA-GPU solver were achieved - fit a large model with 128 channels in wopping 8 seconds! Further tests are needed though to describe the accuracy drop that is achieved by solving X'Xb = X'y instead of Xb = y directly.

Merged pull requests:

predict tutorial (#219) (@behinger)
better fit docs (#221) (@behinger)
docs: add maxvanmigem as a contributor for bug (#222) (@allcontributors[bot])
update to JLD 0.5 (#224) (@behinger)
docs: add ReneSkukies as a contributor for code, and ideas (#226) (@allcontributors[bot])
new solver system (#227) (@behinger)
better benchmark structure (#228) (@behinger)

Closed issues: