The classification process takes long time (11 hours for a picture) but gives excellent results. i have already shrunk my training bank to the minimum. the File of ClassificanionKNN stores many types of information cells and arrays, uploading this bank to parallel processing or GPU processing causes errors. (Matlab preferred).
My background is more agricultural than computer science.
Thanks for the help
Hi Omri,
It isn't very trivial to take an algorithm and convert it so that it would run efficiently in the GPU. A great place to start, in my opinion, would be this book https://developer.nvidia.com/content/cuda-example-introduction-general-purpose-gpu-programming-0
One of the main things you need to identify is, what computations are independent from previous computations such that you can perform it in parallel? Of course, there could be some dependence and you can work around that, but the more dependence you have, the less improvement you'll get.
I would suggest you go through that book (or another one if you want) because you are going to have to structure things in memory, and the way you store and access your data will be critical for your performance. Depending on your GPU, If you can structure your problem and data in a nice way you might be able to get your running time in 1 hour, but if you define your problem poorly you might get a much worse running time than in the CPU.
Nvidia also has a suggested reading page, https://developer.nvidia.com/suggested-reading , go over some of the material and implement some of the things they describe to understand the kind of problem you are going to have to deal with.
Let me know if I can assist you in any other way, I'll be glad to do so.
If your background is more agriculture than computer science, then I do not recommend you to look into the nitty-gritty details of GPU programming. Programming the GPU is generally hard, and we found that interfacing efficient GPU code with MATLAB is not a trivial task either. Rather, use a high-level machine learning library that is already accelerated by GPUs.
Fortunately there is no shortage of such libraries, albeit not so much for MATLAB. As you are using kNN, a quick search brought up this result:
https://vincentfpgarcia.github.io/kNN-CUDA/
It has a MATLAB interface, but the code seems a bit dated.
If you are willing to go beyond MATLAB, but still keeping it high-level, Theano and Torch are two excellent libraries to help you build GPU-accelerated learning algorithms:
http://deeplearning.net/software/theano/
http://torch.ch/
NVidia also has some recommendations on the topic:
https://www.nvidia.com/object/machine-learning.html
Thank you very much for the help Fransisco and Peter.
as Peter said, my background is more agriculture than computer science. i'm learning this all by myself, no courses or lessons... i'm using arrayfunk and predict which both suitable for GPU but....
making it work in parallel computing was fine but the data gathering was bad in this option.
an example code or similar code will be a good example for me to learn from...
i'm giving a look to the link Peter gave, but after two weeks of GPU.... i think i have passed it around... now computer is rendering... waiting for results.
thanks again
Omri Krikeb
How many input data do you have? GPU could calculate KNN faster than KDTree (http://vincentfpgarcia.github.io/kNN-CUDA/) . If you show a few inputs I could give you some recommendations.
I have three arrays of 20170320 pixels and a bank of 14500 points in 7 classes. if the bank is smaller classification is less good.
thanks for your help.
i currently trying a faster method ... it takes time for results....
I have three arrays of 20170320 pixels and a bank of 14500 points in 7 classes. if the bank is smaller classification is less good.
thanks for your help.
i currently trying a faster method ... it takes time for results....
- Badges
- Science topic
- Similar topics
- Computer Science