|
⇤ ← Revision 1 as of 2010-01-14 21:34:09
Size: 1033
Comment:
|
← Revision 2 as of 2010-01-14 21:39:08 ⇥
Size: 970
Comment:
|
| Deletions are marked like this. | Additions are marked like this. |
| Line 1: | Line 1: |
| '''Q''': I have a quetion regarding the reccomended use of "classiter=" in the refine routine. My normal approach to reconstruction involves generating a starting model, then a number of rounds of refine at high angular spacing and starting at classiter=8 to avoid initial model bias, reducing down to classiter=5 then classiter=3. Something like refine 3 ang=10 classiter=8 ..., refine 10 ang=10 classiter=5..., refine 20 ang=10 classiter=3. I would then proceed to reduce ang down until it has reached an appropriate limit based on the target resolution and/or dataset size. My questions are | Q: What value should I use for classiter ? |
| Line 3: | Line 3: |
| 1) Given that higher values of classiter are expected to yield more blurred models, is it important to refine to psuedo convergence at this stage? 2) When i change to a lower value of ang, should i iterate classiter=8, classiter=5, classiter=3 again again at each new value of ang, or just proceed with classiter=3. I.e. is it important to "reheat" the reconstruction once you have locked into something at high values of ang? |
A: Large (4-8) values of classiter do an excellent job of preventing model/noise bias, but slightly degrade resolution and sharpness (different things) of the reconstruction. I suggest using 4-8 for a few rounds until convergence, then do your high resolution refinement with 3 (which actually does only 1 iteration for historical reasons). 3 is still reasonably consevative with respect to inducing model bias. Then at the end, if you want to 'sharpen up' your model again, you can use 0 for at most 2 iterations. In many cases you can continue to improve the resolution as measured by FSC by iterating more with classiter = 0, but this is meaningless, as it's really just inducing noise bias. One or (at most) two iterations of classiter=0 is reasonably safe at the very end of the process, but if you want to have a little extra confidence that you can trust what you see, you might want to stick with 3 or more. |
Q: What value should I use for classiter ?
A: Large (4-8) values of classiter do an excellent job of preventing model/noise bias, but slightly degrade resolution and sharpness (different things) of the reconstruction. I suggest using 4-8 for a few rounds until convergence, then do your high resolution refinement with 3 (which actually does only 1 iteration for historical reasons). 3 is still reasonably consevative with respect to inducing model bias. Then at the end, if you want to 'sharpen up' your model again, you can use 0 for at most 2 iterations. In many cases you can continue to improve the resolution as measured by FSC by iterating more with classiter = 0, but this is meaningless, as it's really just inducing noise bias. One or (at most) two iterations of classiter=0 is reasonably safe at the very end of the process, but if you want to have a little extra confidence that you can trust what you see, you might want to stick with 3 or more.