FALCONN
FALCONN copied to clipboard
FALCONN-LIB
Refactor the Makefile
The build targets for all the tests should be generated fully automatically by a script. The list of headers can also be generated by a script.
The script could also figure out which version of swig to use (swig on Mac vs swig3.0 on Linux), and which g++ (g++-5 on Mac with the homebrew g++ install vs g++ on Linux).
Differences between Mac and Linux we noticed so far:
- We want to use swig version 3. On Linux, this binary seems to be called
swig3.0most of the time, while a homebrew install on Mac calls the binary `swig'. - On Mac,
-pthreadgives you warnings because it is not used, while on Linux we need it. - On Mac, the GCC version of g++ (as opposed to the clang wrapper) is usually available as
g++-5(at least when installed via homebrew). On Linux, we should probably just stick withg++.
Currently the NumPy include folder is also hard-coded to the Python 2.7 site-packages directory, which prevented me from installing FALCONN for Python 3 using pip. Other than that it seems to work correctly for Python 3.
Using pip I got some seemingly unrelated compiling errors (but I'm far from an expert). After fixing this I was able to install it manually without any further issues. I haven't tried writing some own code yet, but the tests were successful at least.
I'm excited by the performance you're able to achieve compared to Annoy. For now I'll stick to Annoy though because of its thread-safety which enables me to do multiple queries in an embarrassingly parallel fashion. I'll definitely keep track of your efforts to add this to FALCONN as well. Also, the native support for sparse vectors in Python could be an awesome feature. Nice work!
