GPU svm

Meskipun jumlah kelas salah, log berikut menunjukkan beberapa perbedaan kecepatan olah SVM antara GPU-GPU yang sudah saya coba. Hubungan yg nyata dg kecepatan adalah spek RAM, apakah DDR2 atau DDR3.

Using device 0: GeForce GTX 260
Input Train File Name: ../../data/shuttle/shuttle.scale
Input Test File Name: ../../data/shuttle/shuttle.scale.t
Code: All Vs All

# of training samples: 43500
# of testing samples: 14500
# of features: 9
# of tasks: 1
# of classes: 5
Beta: 0.500000
Parsing input data…
Parsing input data done!
Class 0:  1,
Class 1:  -1,
Class 2:  0,
Class 3:  0,
Class 4:  0,
Training classifier…
938803200 bytes of memory found on device, 815972352 bytes currently free
4220 rows of kernel matrix will be cached (174016 bytes per row)
All tasks converged! Iterations 422
Training classifier done!
Training time Launch =0.0 usec, finished=122.3 msec
Task 0, svnum, 34, b 13.155989
Testing classifier…
938803200 bytes of memory found on device, 818315264 bytes currently free
14500 rows of kernel matrix will be cached (192 bytes per row)
numThreads 32, numBlocksRed 1
Testing classifier done
Testing time Launch =0.0 usec, finished=519.5 msec
# of testing samples 14500, # errors 3015, Rate 79.206897

surya@FC6KUNIMI:/media/HOME/surya-ntfs/project/research/multisvm_1.0/bin/release$ ./multisvm_1.0
Using device 0: GeForce 9600 GT
Input Train File Name: ../../data/shuttle/shuttle.scale
Input Test File Name: ../../data/shuttle/shuttle.scale.t
Code: All Vs All
# of training samples: 43500
# of testing samples: 14500
# of features: 9
# of tasks: 1
# of classes: 5
Beta: 0.500000
Parsing input data…
Parsing input data done!
Class 0:  1,
Class 1:  -1,
Class 2:  0,
Class 3:  0,
Class 4:  0,
Training classifier…
536150016 bytes of memory found on device, 342269952 bytes currently free
1769 rows of kernel matrix will be cached (174016 bytes per row)
All tasks converged! Iterations 422
Training classifier done!
Training time Launch =1.0 usec, finished=300.4 msec
Task 0, svnum, 34, b 13.155989
Testing classifier…
536150016 bytes of memory found on device, 344612864 bytes currently free
14500 rows of kernel matrix will be cached (192 bytes per row)
numThreads 32, numBlocksRed 1
Testing classifier done
Testing time Launch =0.0 usec, finished=547.7 msec
# of testing samples 14500, # errors 3015, Rate 79.206897

surya@thesisbox:~/multisvm_1.0/bin/release$ ./multisvm_1.0
Using device 0: GeForce 9600 GT
Input Train File Name: ../../data/shuttle/shuttle.scale
Input Test File Name: ../../data/shuttle/shuttle.scale.t
Code: All Vs All
# of training samples: 43500
# of testing samples: 14500
# of features: 9
# of tasks: 1
# of classes: 5
Beta: 0.500000
Parsing input data…
Parsing input data done!
Class 0:  1,
Class 1:  -1,
Class 2:  0,
Class 3:  0,
Class 4:  0,
Training classifier…
1073414144 bytes of memory found on device, 933208064 bytes currently free
4825 rows of kernel matrix will be cached (174016 bytes per row)
All tasks converged! Iterations 422
Training classifier done!
Training time Launch =1.0 usec, finished=787.6 msec
Task 0, svnum, 34, b 13.155989
Testing classifier…
1073414144 bytes of memory found on device, 935550976 bytes currently free
14500 rows of kernel matrix will be cached (192 bytes per row)
numThreads 32, numBlocksRed 1
Testing classifier done
Testing time Launch =1.0 usec, finished=497.0 msec
# of testing samples 14500, # errors 3015, Rate 79.206897

surya@surya-desktop:/media/XP-BARU/home/surya/project/research/multisvm_1.0/bin/release$ ./multisvm_1.0
Using device 0: GeForce 9800 GT
Input Train File Name: ../../data/shuttle/shuttle.scale
Input Test File Name: ../../data/shuttle/shuttle.scale.t
Code: All Vs All
# of training samples: 43500
# of testing samples: 14500
# of features: 9
# of tasks: 1
# of classes: 3
Beta: 0.500000
Parsing input data…
Parsing input data done!
Class 0:  1,
Class 1:  -1,
Class 2:  0,
Training classifier…
536543232 bytes of memory found on device, 364249088 bytes currently free
1883 rows of kernel matrix will be cached (174016 bytes per row)
All tasks converged! Iterations 422
Training classifier done!
Training time Launch =0.0 usec, finished=265.7 msec
Task 0, svnum, 34, b 13.155989
Testing classifier…
536543232 bytes of memory found on device, 366723072 bytes currently free
14500 rows of kernel matrix will be cached (192 bytes per row)
numThreads 32, numBlocksRed 1
Testing classifier done
Testing time Launch =0.0 usec, finished=416.9 msec
# of testing samples 14500, # errors 3015, Rate 79.206897

Advertisements
Posted in Programming/Notes
9 comments on “GPU svm
  1. nedi says:

    Pak kalau boleh tau, itu melakukan komputasi SVM dengan GPU ya? kemudian itu ada multisvm maksutnya dalam sekali running bisa membedakan multiclass secara langsung?

    terima kasih.

  2. surya says:

    Ya, komputasi svm dg gpu. Kalo tidak salah, sekali running program bs mbedain multiclass langsung. Saya belum baca source-nya secara detil, tp sepertinya pake satu-satu juga kok di dalamnya, ada papernya, yg membahas dg teknik OAO dan OAA.
    Gimana kabar?

  3. baihi says:

    Test1 with GTX260 is very fast.
    Test2 and test both with 9600GT, why speed is different?
    I find you have written an article about gpusvm.
    Do you have compare gpusvm speed with mutlisvm?
    It’s seem that gpusvm don’t use cublas and rbf kernel require memory access only once. Maybe this will cause gpusvm fast.

  4. Rickylucky says:

    Do you execute multisvm and gpusvm in linux?
    I try them in MAC OSX, but I cannot compile them.

  5. surya says:

    to Rickylucky, yes, on linux. Never try on MacOSX.

  6. surya says:

    to baihi, they are two different machines. 9600GT spec is different, one with the lower speed of GPU RAM.

  7. rera says:

    wah, sangat menarik sekali..

  8. bapak, bolehkah non mahasiswa diskusi disini?

  9. surya says:

    to busana muslim, boleh saja

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s

%d bloggers like this: