Dear all, I finally found some time to do some polishing on the TV denoising code that Ma�l had written a while ago. I added a new module to my branch https://github.com/emmanuelle/scikits.image/tree/tv_denoise (https://github.com/emmanuelle/scikits.image/blob/tv_denoise/scikits/image/fi...). I'd like to have your opinion on how the code could be improved before a pull request. Cheers, Emmanuelle
Great news! have you been able to do some benchmarks to see if the performance hit from Python is reasonnable?
have you been able to do some benchmarks to see if the performance hit from Python is reasonnable?
I had done a rapid benchmark when I translated the C code into Python, and found that the Python code was 3 times slower compared to the C with Cython wrapping. But now I have improved the performance with a stop criterion on the energy, instead of a fixed number of iterations; hence it is difficult to compare now (of course, the stop criterion could be implemented in C too, but I was much more comfortable to do it in Python!). The performance has to be reasonable, because I have a student who is using this code to denoise large 3D images (512^3)! Emmanuelle �
great news :) On Sat, May 28, 2011 at 16:07, Emmanuelle Gouillart < emmanuelle.gouillart@normalesup.org> wrote:
have you been able to do some benchmarks to see if the performance hit from Python is reasonnable?
I had done a rapid benchmark when I translated the C code into Python, and found that the Python code was 3 times slower compared to the C with Cython wrapping. But now I have improved the performance with a stop criterion on the energy, instead of a fixed number of iterations; hence it is difficult to compare now (of course, the stop criterion could be implemented in C too, but I was much more comfortable to do it in Python!).
The performance has to be reasonable, because I have a student who is using this code to denoise large 3D images (512^3)!
Emmanuelle
Hi Emmanuelle On Sat, May 28, 2011 at 3:05 PM, Emmanuelle Gouillart <emmanuelle.gouillart@normalesup.org> wrote:
I finally found some time to do some polishing on the TV denoising code that Maël had written a while ago. I added a new module to my branch https://github.com/emmanuelle/scikits.image/tree/tv_denoise (https://github.com/emmanuelle/scikits.image/blob/tv_denoise/scikits/image/fi...).
I'd like to have your opinion on how the code could be improved before a pull request.
The code looks great. Two minor things: a) let's move the tests into tests/test_tv_denoise.py b) Document the input dtype--currently data needs to be in floating point format--should it be in [0,1]? This is not quite clear. Thanks for taking the time to do this! Cheers Stéfan
The code looks great. Two minor things: a) let's move the tests into tests/test_tv_denoise.py b) Document the input dtype--currently data needs to be in floating point format--should it be in [0,1]? This is not quite clear.
Done; please tell me if it's ok. The input data can have any numerical dtype, actually, it doesn't have to be in [0, 1]. Cheers, Emmanuelle
Dear Emmanuelle and Maël On Mon, May 30, 2011 at 11:15 PM, Emmanuelle Gouillart <emmanuelle.gouillart@normalesup.org> wrote:
The code looks great. Two minor things: a) let's move the tests into tests/test_tv_denoise.py b) Document the input dtype--currently data needs to be in floating point format--should it be in [0,1]? This is not quite clear.
Done; please tell me if it's ok. The input data can have any numerical dtype, actually, it doesn't have to be in [0, 1].
Thanks for all your effort. Your code has been merged into the main repository: https://github.com/stefanv/scikits.image/commit/ceb2e419c3eaf58cdc28062cca2e... Regards Stéfan
Thank you, St�fan! On Tue, May 31, 2011 at 12:49:06PM +0200, St�fan van der Walt wrote:
Dear Emmanuelle and Ma�l
On Mon, May 30, 2011 at 11:15 PM, Emmanuelle Gouillart <emmanuelle.gouillart@normalesup.org> wrote:
The code looks great. �Two minor things: a) let's move the tests into tests/test_tv_denoise.py b) Document the input dtype--currently data needs to be in floating point format--should it be in [0,1]? �This is not quite clear.
Done; please tell me if it's ok. The input data can have any numerical dtype, actually, it doesn't have to be in [0, 1].
Thanks for all your effort. Your code has been merged into the main repository:
https://github.com/stefanv/scikits.image/commit/ceb2e419c3eaf58cdc28062cca2e...
Regards St�fan
participants (3)
-
Emmanuelle Gouillart -
Maël Primet -
Stéfan van der Walt