xfbingshan / multisvm Goto Github PK

What steps will reproduce the problem?
1. Sync source code using command on Source page
2. Open MultiSVM.sln

What is the expected output? What do you see instead?
Visual Studio 2008 has an error opening the multisvm_2.0/MultiSVM++ 
project.
"The project consists entirely of configurations that require support for 
platforms which are not installed on this machine.  The project cannot be 
loaded."

What version of the product are you using? On what operating system?
R51, VS2008 on Windows 7 64bit.  8800GTX. Tried installing Microsoft 
Robotics Studio since one of the missing dlls was in there 
(Microsoft.Ccr.Core), but it didn't help.

Please provide any additional information below.
Please help.  I am very interesting in running this beautiful code :-)

Tesla is so expensive. Is geforce good enought for svm?
Is multisvm sensitive for shader clock, or sensitive for memory bandwidth?
Which card has most high performance-price ratio?
Any suggestions or help will be appreciated.

What steps will reproduce the problem?
1. Upon running usps (10-class) dataset I get an accuracy of 17%
2. I tried printing out all the SVs and they are all same and equal to C (i am 
running an RBF kernel with C=10 and g=1/256)
3.

What is the expected output? What do you see instead?
Expected accuracy = 97 --> get 17
Incorrect SVs being generated

What version of the product are you using? On what operating system?
I am using multisvm_1.1

Please provide any additional information below.
Additionally it runs out of memory for Adult and Shuttle. I have a 2GB GTX 660.

What steps will reproduce the problem?
1. Call the trainclassifier() function twice within the same program (for 
hyperparameter validation for instance)
2. The program will crash with the following output
multisvm_1.0(61852) malloc: *** error for object 0x7fe793a36760: pointer being 
freed was not allocated
*** set a breakpoint in malloc_error_break to debug
Abort trap: 6

What is the expected output? What do you see instead?
No crash!!!

What version of the product are you using? On what operating system?
multisvm_1.1_11-14-09

Please provide any additional information below.
I solved the problem in function Cache::Cache(int nPointsIn, int cacheSizeIn) 
from Cache.cpp of directory.reserve(nPointsIn) by changing the following line

    directory.reserve(nPointsIn);

to 

   directory.resize(nPointsIn);

This change has the effect of causing a explicit initialization of 
Cache::directory[i].status to "0", where i belongs to [0 nPointsIn].

Before making the change, this variable was not explicitly initialized.  On the 
first run of the trainclassifier(), the directory[i].status value of class 
Cache is correctly initialized to 0 after running  directory.reserve(nPointsIn) 
in the constructor Cache::Cache(), but on the second run, directory[i].status = 
1 after running the directory.reserve(nPointsIn) which is obviously not the 
expected behaviour.

Sorry if my explanations are a little unclear.  The attached file contains the 
correction.

Check other kernels

What steps will reproduce the problem?
1. make, show no cutil.h
2. Download one copy of cutil, /usr/bin/ld: cannot find -lcutil

GTX 400 don't support Linux now. So I have to run it on windows.
Does anyone have convert the project to windows?

In testing.cu, at the line #246 "if(kernelcode==0)", his block ends at the
line 477. The kernel evaluation for other kernel types are in this block.
But with "if(kernelcode==0)", it will never compute any other kernel types.
So it won't work on other kernels(linear, poly, sigmoid) except the RBF kernel.

What steps will reproduce the problem?
1.When I compile the your program, it showed these errors as: identifier 
"FLT_MAX" "FLT_MIN" are undefined in the "reduction_kernel.cu"
2.
3.

What is the expected output? What do you see instead?
Could I know the value for FLT_MAX and FLT_MIN

What version of the product are you using? On what operating system?
I compile in the linux system

Please provide any additional information below.

What steps will reproduce the problem?
Compile the source code with latest cuda driver and sdk 3.2.
Run it with a tesla c2050/2070

What is the expected output? What do you see instead?
For adult dataset, the accuracy is 82.67% on tesla c1060 with cuda 2.3.
However the accuracy drops to 48.20 on tesla c2050/2070.

What version of the product are you using? On what operating system?
CUDA 3.2 Fedora Core 13

Please provide any additional information below.

Nothing is modified in the source code. But the number of support vectors as 
well as the bias value are always not the same on c1060 and c2050 cards. Please
give me some suggestions on what could go wrong with the new card.

Thanks!