It came up in the context of rigid bodies that certain optimizers (namely CG) like to have all their parameters vary over a similar scale. As a result, it can be important to rescale non-xyz attributes. One way of doing this would be to add a method to Model: FloatPair get_range(FloatKey) which returns the current range of values for a particular FloatKey. Then CG could use that to rescale all its values internally.
If it is useful to be able to override the range, we could also add set_range which allows you to set the expected range (perhaps even allowed range) for a given attribute type.
Both of these would be relatively easy to implement on the Model side. The CG code is a bit hairy, so tweaking that could take more work (and is probably best done by adding another set of methods to access the values in the Optimizer base class).
On Feb 17, 2009, at 1:23 AM, Daniel Russel wrote:
> It came up in the context of rigid bodies that certain optimizers > (namely CG) like to have all their parameters vary over a similar > scale. As a result, it can be important to rescale non-xyz > attributes. One way of doing this would be to add a method to Model: > FloatPair get_range(FloatKey) which returns the current range of > values for a particular FloatKey. Then CG could use that to rescale > all its values internally. something like that would be useful indeed. the typical changes of the values would be of importance for optimizers - the absolute values are not really important. thus, this function would somehow need to accumulate knowledge on the magnitude of changes for the variables throughout optimization.
cheers
frido > > > If it is useful to be able to override the range, we could also add > set_range which allows you to set the expected range (perhaps even > allowed range) for a given attribute type. > > Both of these would be relatively easy to implement on the Model > side. The CG code is a bit hairy, so tweaking that could take more > work (and is probably best done by adding another set of methods to > access the values in the Optimizer base class). > _______________________________________________ > IMP-dev mailing list > IMP-dev@salilab.org > https://salilab.org/mailman/listinfo/imp-dev >
--
Friedrich Foerster Max-Planck Institut fuer Biochemie Am Klopferspitz 18 D-82152 Martinsried
Tel: +49 89 8578 2651 Fax: +49 89 8578 2641
foerster@biochem.mpg.de
www.tomotronic.org
>> It came up in the context of rigid bodies that certain optimizers >> (namely CG) like to have all their parameters vary over a similar >> scale. As a result, it can be important to rescale non-xyz >> attributes. One way of doing this would be to add a method to >> Model: FloatPair get_range(FloatKey) which returns the current >> range of values for a particular FloatKey. Then CG could use that >> to rescale all its values internally. > something like that would be useful indeed. the typical changes of > the values would be of importance for optimizers - the absolute > values are not really important. thus, this function would somehow > need to accumulate knowledge on the magnitude of changes for the > variables throughout optimization. I was thinking to use the relative magnitudes as a proxy for the relative changes, but that doesn't always make sense. However, for the rigid body case they probably do a good enough job: that is, for computing a structure from a random conformation x,y,z and the angles will go through their whole range, and if you are just refining the structure, and have reasonable sized rigid bodies, they will both go through a small, roughly comparable, fraction of their range. This does not hold if we are using internal coordinates for a protein or similar non-xyz variables.
Tracking changes is problematic as they are highly variable when you use something like CG and past behavior does not predict the future well (since things settle down). I was trying to use the changes over the last few steps to predict how much slack to add in the non-bonded list, and that was just a mess.
So a refined suggestion: Model would have a parameter called scale per optimizeable attribute. This scale is by default the width of the range of values exhibited for that attribute, but can be set to anything by the user. Optimizers can use this scale as a hint to improve performance (that is, it is non-binding).
On Tue, Feb 17, 2009 at 5:44 PM, Daniel Russel drussel@gmail.com wrote: >>> It came up in the context of rigid bodies that certain optimizers (namely >>> CG) like to have all their parameters vary over a similar scale. As a >>> result, it can be important to rescale non-xyz attributes. One way of doing >>> this would be to add a method to Model: FloatPair get_range(FloatKey) which >>> returns the current range of values for a particular FloatKey. Then CG could >>> use that to rescale all its values internally. >> >> something like that would be useful indeed. the typical changes of the >> values would be of importance for optimizers - the absolute values are not >> really important. thus, this function would somehow need to accumulate >> knowledge on the magnitude of changes for the variables throughout >> optimization. > > I was thinking to use the relative magnitudes as a proxy for the relative > changes, but that doesn't always make sense. However, for the rigid body > case they probably do a good enough job: that is, for computing a structure > from a random conformation x,y,z and the angles will go through their whole > range, and if you are just refining the structure, and have reasonable sized > rigid bodies, they will both go through a small, roughly comparable, > fraction of their range. This does not hold if we are using internal > coordinates for a protein or similar non-xyz variables. > > Tracking changes is problematic as they are highly variable when you use > something like CG and past behavior does not predict the future well (since > things settle down). I was trying to use the changes over the last few steps > to predict how much slack to add in the non-bonded list, and that was just a > mess. > > So a refined suggestion: > Model would have a parameter called scale per optimizeable attribute. This > scale is by default the width of the range of values exhibited for that > attribute, but can be set to anything by the user. Optimizers can use this > scale as a hint to improve performance (that is, it is non-binding). fine. in the past everybody could live with such a user-defined solution in modeller. a more advanced automated function would of course always be appreciated ;)
frido > _______________________________________________ > IMP-dev mailing list > IMP-dev@salilab.org > https://salilab.org/mailman/listinfo/imp-dev > >
Daniel Russel wrote: > It came up in the context of rigid bodies that certain optimizers > (namely CG) like to have all their parameters vary over a similar scale. > As a result, it can be important to rescale non-xyz attributes. One way > of doing this would be to add a method to Model: FloatPair > get_range(FloatKey) which returns the current range of values for a > particular FloatKey. Then CG could use that to rescale all its values > internally.
I share others' concerns that it'll be difficult to determine sensible ranges in many cases. But otherwise, as long as it's an "opt-in" method, it sounds reasonable to me (has to be opt-in because other optimizers such as MD and probably DOMINO want to see the original unscaled optimizable attributes).
Ben
-
Ben Webb -
Daniel Russel -
Friedrich Foerster