Hi Reza,
In my experience the 3dvariability maps need to be low-pass filtered before they can be interpreted, so binning your data will not limit you so long as the binned pixel size is not larger than FilterFrequency/2. Say you need to filter to 8A, with a pixel size
of 1.16A your binned by 2 Nyquist frequency will be 4.64 , so binning your data will not limit you at all and will speed up processing.
Binning is accomplished by the --decimate flag and low-pass filtering can
be implemented with the --fl and --aa flags. I use the EMAN/SPARX daily build and a Relion 2.0 refinement data.star file as my input and run the following script on 16x16cpu nodes:
sxrelion2sparx.py Refine3D/job###/run_data.star --output_dir=Sparx_3DVar --create_stack
sxcpy.py sparx_stack.hdf bdb:sparx_stack
sxheader.py bdb:sparx_stack --import=sparx_stack_proj3d.txt --params=xform.projection
sxheader.py bdb:sparx_stack --import=sparx_stack_ctf.txt --params=ctf
mpirun -np 256 sx3dvariability.py bdb:sparx_stack --CTF --VERBOSE --img_per_grp=100 --fl=0.40 --aa=0.1 \
--ave3D=3Dvol_SPX.hdf --var3D=3Dvar_SPX.hdf --decimate=2
I usually have to filter my variability maps even further, but the hdf files do not contain the right pixel size so you will have to adjust that. I usually convert
them to mrc files with the correct pixel size (here because we decimate by 2 the pixel size should be 2.32) then use relion_image_handler to filter to the desired frequency. Or you can adjust the pixel size and use the gaussian filter in Chimera with a stdv
of 2 or 3 usually works for me. Depending on how isotropic your angular distribution is and the number of particles in the data set you may want to play around with the --img_per_group flag,
maybe vary it between 50 and 200 and compare the results. The --ave3D produces a 3D reconstruction from the grouped/averaged images and gives you a good idea of whether
or not you are getting any strange rotational artifacts from grouping.
With 68K particles decimated by 2 and 40 cpus the job should not take too long, maybe a few hours, but its hard to say. with ~150K particles and 256 cpus my jobs do not take more than an hour. You can always run it on a smaller subset to benchmark the speed.
The Scripps Research Institute
Can anyone suggest some starting parameters for using sx3dvariability in Sphire/Sparx? I have about 68,000 particles, 1.16Å pixel size, 256x256 boxsize, and a resolution of 4.3Å. Should I bin the images, how many img_per_grp (default is 10 but example uses
100)... Thanks.
Also, how long will this take if I only have 40 cores for the job?
The rule of thumb is to use a group number as small as possible to get reliable motion correction. When you see many times of message of repeating measurements
on larger patch, this is a sign of unreliable motion correction.
Someone else also suggested to use larger b factor. The default is 100. You can double it.
Hi CCPEM
When performing motion-correction of low-dose data sets what would be the best method to optimize the group parameter ( to group frames in order to increase signal).Is there a rule of thumb based on dose accumulated or a particular metric used to track the
effect of grouping on the actual signal . Any inputs or suggestions on how this is usually done would be really helpful
Sincerely
Joshua Lobo
