Releases: bacpop/mandrake
Releases · bacpop/mandrake
Mandrake v1.2.3: GPU compile
Fixes a flag exporting the GPU code, used on conda
What's Changed
- Update the citation and logo by @johnlees in #32
- Vendor boost lib by @richfitz in #33
- Allow one to set CUDAARCHS. by @ccoulombe in #36
- Simplify and fix CI by @johnlees in #38
- Add readthedocs config file by @johnlees in #42
- Add GPU_AVAILABLE macro to makefile for conda by @absternator in #45
New Contributors
- @richfitz made their first contribution in #33
- @ccoulombe made their first contribution in #36
- @absternator made their first contribution in #45
Full Changelog: v1.2.2...v1.2.3
v1.2.2: Update to sketchlib v2
This version supports pp-sketchlib v1.7.x and v2.0.0, and enables GPU support if these dists are used.
What's Changed
Full Changelog: v1.2.1...v1.2.2
v1.2.1: Package fixes
These are currently a patch in the draft recipe (conda-forge/staged-recipes#16376), but this will pull the fixes into the package.
There are two further changes:
- The SGD was broken on the GPU as eta was not being decreased due to a stupid pointer error
- Sound for the animation no longer uses the squaring transform (so is higher pitched)
v1.2.0: Animation version
This version primarily tweaks the animation output for better videos. There are also some package changes to work better with conda-forge. Changes are as in #27:
- More robust when labels provided are wrong
- Add Makefile for conda-forge issues
- seed also sets label colours
- Scale plot before running HDBSCAN for more consistent results across datasets
- Start embedding on a unit circle rather than square
- Remove template on sce_results (always use doubles, cast in the destride if fp32 used with CUDA)
- Add ability to add sound to the animation
- Reduce the number of H<->D memcpy and remove host callback from CUDA graph. As this is only called occasionally it speeds inter-kernel time up a lot.
- Make the animation screen shaped, change the colour scheme, add a legend.
v1.1.0: Paper version
This is the version we used for the analysis in the accompanying manuscript.
Changes from v1:
- Fix a memory error in the seeding. Doesn't affect function, but can cause a crash #24
- Fix to discrete draws. May cause some issues with SCE, but was evened out by stochastic nature as most draws correct #24
- Fix use of accessory matrix in python code #24
- Various changes to plots #24 #25
- Small bug fixes in pairsnp (returning a few too many distances) #25
- Fix pairsnp progress meter #25
- Allow Ctrl-C during pairsnp run #25
- Deterministic label colours in plot with
--labels
#25
v1.0.0: Initial release
First version of mandrake, used for the analysis in the manuscript