Re: [scipion-users] increase the iteration of classification run

[ashutosh s. <ash...@gm...>]

Dear Carlos

Thank you for the explanation.

Best Regards
Ashutosh

On Sun, Feb 19, 2017 at 3:20 PM, Carlos Oscar Sorzano <co...@cn...>
wrote:

> Dear Ashutosh,
>
> On 17/02/2017 2:53, ashutosh srivastava wrote:
>
> Hi Everyone
>
> Is it possible to increase the number of iterations in an existing CL2D
> classification run if it does not converge?
>
> Note that CL2D is not an optimization algorithm, the algorithm tends to
> maximize the similarity between the class representative and the class
> images, but it may depart from a strictly monotonic optimization.
>
> I performed a CL2D classification for ~23000 particles for 50 iterations
> and it took quite a lot of time even on a cluster.
>
> Normally, the way of using the algorithms is with multiple levels (with
> different number of classes at each level, 4, 8, 16, 32, ...) and
> relatively few iterations at each level (10 should suffice for practical
> purposes, and even smaller numbers, like 5, usually work although they are
> not recommended).
>
> Visualizing the convergence I realized that the classification does not
> seem to be converged with CL2D similarity of 0.4622 and CL2D changes 8814
> for the 50th iteration. I believe I need to run the classification for
> larger number of iterations. However, there does not seem to a way to
> continue the same run for further 50 or 100 iterations.
>
> This behavior of 1/3 of the particles changing classes is normal in CL2D
> when the number of classes is relatively high. It does not mean that
> classes are bad or that you should remove classes, it means that each class
> has another very close class in such a way that particles have a hard time
> to decide with which one to go (remind that the algorithm is not aimed at
> maximizing a functional, although in general, it tends to maximize the
> similarity between the class representatives and images).
>
> At the level of protocol there is not an easy option to continue from some
> iteration, when using the program there is. A trick could be to use the
> "Additional parameters" field and provide the parameter --ref0 with the
> path to the xmd file of the last iteration of the previous classification.
>
> Kind regards, Carlos Oscar
>
> I am using Scipion v1.0.1.
> Thank you
>
> Best Regards
> Ashutosh
>
>
> ------------------------------------------------------------------------------
> Check out the vibrant tech community on one of the world's most
> engaging tech sites, SlashDot.org! http://sdm.link/slashdot
>
>
>
> _______________________________________________
> scipion-users mailing lis...@li...https://lists.sourceforge.net/lists/listinfo/scipion-users
>
>
> --
> ------------------------------------------------------------------------
> Carlos Oscar Sánchez Sorzano                  e-mail:   co...@cn...
> Biocomputing unit                             http://biocomp.cnb.csic.es
> National Center of Biotechnology (CSIC)
> c/Darwin, 3
> Campus Universidad Autónoma (Cantoblanco)     Tlf: 34-91-585 4510
> 28049 MADRID (SPAIN)                          Fax: 34-91-585 4506
> ------------------------------------------------------------------------
>
>