Thanks Takanori,
To get bfactor plots I started with 150 particles, 3.5 hours default, 14 minutes without gridding and padding, 30 minutes with either gridding or padding disabled, another 40 minutes for a last Refine3D with padding and gridding.
It’s apoferritin however, 35K particles in a 360 box with 0.52015 Å pixel size going to ~1.9 Å on full dataset. Could it be a local computing issue regarding the M-step?
Best,
Wim Hagen
> On 20. Oct 2019, at 21:35, Takanori Nakane <[log in to unmask]> wrote:
>
> Hi,
>
> Skip padding: as pointed out, this reduces GPU memory
> requirement to 1/8. With 11 GB of GPU memory (1080Ti, 2080Ti),
> you can use up to 1024 px box without padding.
> With padding, the maximum is 512 px.
>
> This should also make the E-step faster, but in reality, the
> difference is small unless the number of particles is very large.
>
> For unknown reasons, skip padding sometimes makes class separation
> worse. So this option is not recommended for Class3D.
>
> Skip padding leads to artifacts at the corners of the box when the
> box size is small. This should be harmless as long as you can mask
> it out. I don't know analytical formula how big a box must be
> relative to the particle diameter to avoid artifacts completely.
>
> You can also use non-integer padding (e.g. "--pad 1.5").
> Note that the GUI always appends --pad 1 or --pad 2.
> In RELION 3.1, you can specify --pad 1.5 as "additional arguments",
> as the command line parser uses the last value when the same option
> is specified multiple times. In earlier versions, you have to
> remove --pad 1 (or 2) produced by the GUI, since the command line
> parser uses the first value.
>
> Skip gridding makes the M step faster.
> The M step is not GPU accelerated and the time depends on the
> box size, (current) resolution and the number of classes,
> but not on the number of particles.
>
> In our tests on several datasets, the results look fine but
> I cannot guarantee it is always fine, especially when severe
> preferred orientation is present.
>
> We welcome feedback if skip gridding leads to significantly worse
> results than usual.
>
> Regarding Wim's original questions:
>
> >> In Relion 3.1 Refine3D, skipping padding and gridding, and once
> >> finished, continue again with padding and gridding, saves time:
> >> - Is this a valid approach?
>
> This should work, but not tested. Please let us know if it fails.
>
> >> - If so, could this be an option? So for skip padding and skip
> >> gridding, have options yes/no/except last iteration(s)?
>
> "--pad 2 only in the last interation" is probably useless because
> (1) --pad 1 does not makes E steps much faster as I wrote above and
> (2) the last iteration uses the highest amount of GPU memory.
> If it can survive --pad 2, then one can use --pad 2 for earlier
> iterations as well.
>
> "do gridding only in the last iteration" is more interesting but
> we need to study whether skip gridding is really harmfull and
> doing it at the last iteration solves the issue.
>
> Best regards,
>
> Takanori Nakane
>
>> On 2019/10/20 13:40, 류범한 wrote:
>> Dear Wim,
>> As long as I know, skip padding is useful when you are facing with some kind of memory-related error, and such situation is mostly due to huge box size that consumes a lot memory. In the case, skip padding is often recommended to finish your refinement job anyway. So it can be said as a valid way.
>> But I am not sure if you use it just for saving the processing time. Since refinement jobs even with exactly same given parameters resulted in more or less different consequences(depends on your data?), I couldn’t prove the option is not totally harmful in making the same result.
>> And I also wonder for skip-gridding option.
>> I think Sjors or Takanori will give you much clear answer :)
>> Best regards,
>> Han
>> *Bumhan Ryu*
>> *Ph.D*
>> Research engineering fellow
>> Protein Communication Group
>> Pioneer Research Center for Biomolecular and Cellular Structure
>> Institute of Basic Science (IBS)
>> (34126) 55, Expo-ro, Yuseong-gu, Daejeon <x-apple-data-detectors://1>, South Korea
>>> 2019. 10. 19. 오후 9:15, Wim Hagen <[log in to unmask]> 작성:
>>>
>>> Hi,
>>>
>>> In Relion 3.1 Refine3D, skipping padding and gridding, and once finished, continue again with padding and gridding, saves time:
>>> - Is this a valid approach?
>>> - If so, could this be an option? So for skip padding and skip gridding, have options yes/no/except last iteration(s)?
>>>
>>> - Is there a rule-of-thumb for skipping padding based on particle vs box size?
>>>
>>> Best,
>>>
>>> Wim Hagen
>>> ########################################################################
>>>
>>> To unsubscribe from the CCPEM list, click the following link:
>>> https://www.jiscmail.ac.uk/cgi-bin/webadmin?SUBED1=CCPEM&A=1
>> ------------------------------------------------------------------------
>> To unsubscribe from the CCPEM list, click the following link:
>> https://www.jiscmail.ac.uk/cgi-bin/webadmin?SUBED1=CCPEM&A=1
>
> ########################################################################
>
> To unsubscribe from the CCPEM list, click the following link:
> https://www.jiscmail.ac.uk/cgi-bin/webadmin?SUBED1=CCPEM&A=1
########################################################################
To unsubscribe from the CCPEM list, click the following link:
https://www.jiscmail.ac.uk/cgi-bin/webadmin?SUBED1=CCPEM&A=1
|
