On Mon, Oct 20, 2008 at 5:12 PM, Tyler Reddy <TREDDY@xxxxxx> wrote:
Hello,
1)
I have been using the multi model and model selection scripts in relax
1.3.2 but
I have trouble displaying the output in a tabulated format. Both scripts
seem
to produce an xml document with various headers that isn't easy to read. It
looks like format='columnar' isn't supported. I wonder what other options I
have to look at this data? For some reason, I don't recall having this
problem
on a Linux machine a few weeks ago (using a Mac OS 10.4 at the moment), but
anyways it would be nice to get readable model-free output.
The results file is now in XML format and the more readable 'columnar'
format has been removed from the 1.3 line. With the change to the new
XML results file all the contents of a data pipe, irrespective of what
that data is, is packaged. So you can put data into this pipe
yourself and it will save that information (for advanced users,
complex python objects will need the to_xml() and from_xml() methods
to package and unpackage the data). The reason for removing the
'columnar' format was that it was considered too inflexible for the
changes occuring in the 1.3 line, it contained duplicate information,
had numerical precision issues, and there were alternatives to easily
view this data. You can use the value.display() and value.write()
user functions to display and save the results for a single parameter.
If needed, these user functions could be extended to accept a list of
parameters.
Then there is the sample_scripts/latex_mf_table.py sample script which
will generate a LaTeX table of the model-free results. This file can
be copied and modified - this requires learning a bit of python - to
format and display the results any way you wish. And finally if
anyone really wants to, and has the skills to, they can modify this
sample script to recreate a version of the 'columnar' format. This
could be added to the relax sample scripts, and if their skills are
very advanced, then much code from the 1.2 relax versions can be
recycled.
2)
The error input for the relaxation rate parameters is currently my
non-linear
curve fitting standard deviation. I'm not sure if that means subsequent
analysis will be completely incorrect? I guess it depends on the comparison
of
magnitude between these errors and the type of error that is propagated by
relax during its own curve-fitting (which I am unable to do at the moment).
I'm guessing this is the sum of squared error value (SSE) from the
fit. Or is it a regression coefficient or a chi-squared value? Did
the fitting use a technique such as bootstrapping or jackknife
simulations to estimate the parameter errors via propagation? Or did
it use the covariance matrix? If it is the SSE, chi-squared, or
regression coefficient then that value cannot be used. This will be
wildly wrong and cause massive failure in model selection. It will
cause big problems in optimisation, and if you are unlucky and have
spaces with long, curved valleys or flat curved spaces leading to the
minimum (that's model-free models m5 to m8 in most cases and not so
uncommon in model m4) then the minimum can be squeezed and appear in
another completely different region in the space. It will likely also
cause model failure issues, which although removed by the eliminate()
user function, might discount the best solution. I would guess that
all of this will have a measurable affect on the final diffusion
tensor as well and, if so, this will cause the appearance of
artificial motions (my 2007 JBNMR paper at
http://dx.doi.org/10.1039/b702202f explains these problems in detail).
If one is not careful with the errors and they are significantly off,
then the result is that the results may not be real. So I would only
use the error if it comes from an established error propagation
technique (i.e. from data to parameter error propagation).
Regards,
Edward
Re: Model-Free Results Output