For information on liblinear, please refer to README.liblinear
The following table shows the comparison between liblinear and liblinear_gpu. liblinear_gpu achieves 36x speedup compared to the original cpu implementation on the largest training set with 20k samples with dense 4k features.
Liblinear | Liblinear_GPU | ||||||
Data Points | Reading Input | Total Time | Reading Input | Total Time | Computation Speed Up | Total Speed Up | Data Matrix Size in GB |
1000 | 0.83 | 6.194 | 0.81 | 1.316 | 10.60079051 | 4.70668693 | 0.030517578 |
5000 | 4.11 | 73.964 | 4.05 | 7.075 | 23.0922314 | 10.45427562 | 0.152587891 |
10000 | 8.14 | 240.847 | 8.16 | 15.325 | 32.47829728 | 15.71595432 | 0.305175781 |
15000 | 12.24 | 345.174 | 12.16 | 26.416 | 23.35395623 | 13.06685342 | 0.457763672 |
20000 | 16.68 | 602.134 | 16.139999 | 32.096 | 36.69177509 | 18.76040628 | 0.610351563 |
Test datasets are available at http://vision.cornell.edu/~mmoghimi/liblinear_gpu/
Contact: mmoghimi at_sign cs dot cornell dot edu