CRF-suite on AVX2 cpus
using sklearn-crfsuite or python-crfsuite on an AMD system can be very slow due to optimizations in the precompiled wheel files that are specific to intel processors.
I have a branch of python-crfsuite that has flags for avx2 instructions. To see if your cpu supports avx2 instructions, examine the output of lscpu | grep avx2
. If anything is returned, then your cpu supports the avx2 instruction set.
To use this fork:
Create a virtual environment for this version of python-crfsuite
virtualenv ~/venvs/crfsuite
source ~/venvs/crfsuite/bin/activate
Clone my fork of python-crfsuite
git clone --recurse-submodules git@github.com:ksteimel/python-crfsuite.git
Build python-crfsuite
python setup.py build
python setup.py install
Install additional dependencies
pip install sklearn-crfsuite
pip install scikit-learn
How much does this help?
Even on intel cpus that support avx2 instructions, the time taken to complete a grid search is reduced. For example, a 5 fold grid search with 10 parameter combinations (e.g. 50 total runs) takes 4 minutes to complete using the version in pypi (on a POS tagging problem in a Turkic language). The avx2 compiled version completes in 3 minutes. This becomes more pronounced as the size of the tagset increases.
Using avx only (e.g. changing -mavx2 to -mavx in the setup.py scipt) still results in improvements to performance. On a pair of intel e5-2680, here are the runtimes for this same benchmark script in minutes. Note that the avx on setting for 16 is having trouble even keeping the cpus loaded because the grid search runs are finishing too fast. This is an issue when the number of tags is small as each task in grid search finishes very quickly and most of the time is spent on task overhead. with longer running tasks, the difference is more noticable.
n_jobs | avx off | avx on |
---|---|---|
8 | 3.5 | 2.7 |
16 | 2.8 | 2.5 |
No Comments