help-gsl
[Top][All Lists]
Advanced

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: Checking GSL for Spectroscopy


From: Fritz Sonnichsen
Subject: Re: Checking GSL for Spectroscopy
Date: Thu, 18 Mar 2021 11:47:28 -0400

Mark-Speed is an issue as follows"
  Our device is an underwater microscope to which we are adding a raman
spectrometer.
  Plankton, plastics or other objects pass thru a flow cell and pass a
camera and the spectrometer inputs.--all in a few seconds
  AI on the device quickly decides what it is "seeing" and records this.
  Apparently this must all be done "in situ" --e.g. no post processing on
the shore.
  Meanwhile a spectrum is taken and this checks against 50,000 spectra,
combining them if necessary.

So the camera passes the "shape" that it sees to an AI engine which has to
very quickly tell what plankton type it sees-using opencv and various
morphology and size. It is subtle--I am a physicist but the biologist tells
me that small differences in structure mean a lot in determining species
and state of growth. (To me they all look like YUK!)
  Ocean plastics are easier-the camera does not play into these--it is all
about the spectrum

Amazing thing is that most of this works but we want to "see" things every
5 seconds, decide if an ROI exists--and then go ahead with the speciation
--say every minute.

This is not really a Finite Element problem--the spectra are linear--the
images are largely 2D.
     Yes--Dates should be very simple but at least Octave, Matlab, gnuplots
and others that I can't think of-all seem to make a mess of simple dates in
plots
Those old card punches did indeed have "TAB".  The O29 keypunch from IBM
had a way to set them up and knowing this could rate a few dates. We also
had card duplicators and sorters. I recall the long 2 mile walk in 2 feet
of snow from college, carrying 2 foot-long  boxes of cards.Inevitably some
would get wet. (plastic grocery bags didn't exist yet!). To this day, on
occasion I will pull a 1970 book from my library and a card falls out. They
still make good bookmarkers! And yes--water in real crystals today is every
bit as much a problem to spectroscopy as is wet cards representing a
crystal !!!

SO-I suppose I am ranting off topic here--but in summary I think science
has run off the rails, computers have the tail swinging the dog--and nobody
is in charge to set it right again. As Richard Feyman called it--many
scientists have "got the bug" and put away their math books to play endless
games keeping up with superfluous computer changes. Thankfully a few
products like GSL keep their C and their sanity.

cheers
fritz MA,USA



On Thu, Mar 18, 2021 at 8:52 AM Mike Marchywka <marchywka@hotmail.com>
wrote:

> On Thu, Mar 18, 2021 at 08:33:05AM -0400, Fritz Sonnichsen wrote:
> >    Thanks Mike
> >      as you mention the code that I am starting with (written by a
> colleague) is more "model based" in that it
> >    deliberately does not invoke the underpinning properties of Raman
> optical spectra. This code is designed to run really fast
> >    on GPUs and minimizes any calculations. When I start working with it
> I will figure out if some preprocessing to categorize
> >    the spectra for example would be of value.
>
> Why is speed such an issue? Is this something to run on a distributed
> system? You
> could take a look at the libraries called by FEM code from national labs
> like
> libmesh. This is really complicated but something to look at as an
> alternative-
> I would stick with plain c code if you can :)
>
> >       Your comment on CRAN R is of interest-i looked at it a few days
> ago and it may be an approach--I think I will know more
> >    once I get the original code and see how sophisticated the statistics
> usage is. But at least at this point I think C with a
> >    little GSL will do.
>
> I actually wrote some of my own c++ code for graphing which sounds dumb
> since R
> should do that well but I had issues with dates and wanted to get svg and
> latex
> outputs although R supports both to some extent IIRC.
>
> >      I am interested to find that Python is losing at least some ground
> these days. I started on UNIVACs using assembler and
> >    FORTRAN on punch cards so I was always a little distressed by
> adherence to column position issues necessary back then. So
>
> I'm not sure if cards had the concept of a "tab" lol that was the real
> annoyance.
>
> >    when I had  to write in Python a few years ago I actually had to call
> a computer colleague to make sure I wasn't
> >    missing something when the column requirements came up. Jeeesh. More
> disturbing is the inability to write comments at the
> >    end of lines. I am a big supporter of in-line doc and I expect this.
> Putting documents before the line takes a lot of space
> >    and more words--I have never rated code well on how many pages you
> have to flip. But the real problems with Python started
> >    when I needed to do serial port programming--some concoction of
> shared routines that ran differently  on ver 2 and ver 3
>
>  I seem to recall that the Python module concept may bridge the language-
> subroutine
> idea.
>  I also like to "code for grep" and end of line comments are great
>
> >    were needed and it was a disaster. A lot of other stories but I think
> people here know them. It is sad to see a large body
> >    of scientists here thinking that somehow Python was necessary to
> accommodate its rich set of routines. They have never seen
> >    another language and don't understand that those routines could have
> been written in something more stable and simpler.
> >      Your story about writing a citation program hits home for me. I
> worked on a system for formatting NOAA data that one of
> >    the institutions uses here. It requires a representative from each
> department and tons of ill-formatted data must be made
> >    to comply with a generic format-all this done with unique code for
> each case. And then the code never worked of course. An
>
> I have not even gotten to unstructured data- this is just trying to find
> the structured
> data although that may be next. Parsing formatted citations .... aarrrgh.
>
>
> >    enormous cost of rare research funds that could have easily been
> saved by simply making a few requirements on CSV data. The
>
> There is a simple line oriented key-value format called RIS and medline
> uses
> a similar or identical format. I prefer bibtex because it is easier to read
> , more versatile, and is my target format anyway .
>
> >    computer tower of babble that has been built in recent years shows to
> me at least that science has lost sight of what it is
> >    supposed to do and how it should get here. The tool is no longer a
> tool-it is becoming the prime cost.
> >       The GPL issues have been discussed here and I appreciated that
> input. I have our business person looking at this. The
> >    Spectra library that we call must remain proprietary but the
> operations on it are--to my mind-rather generic (I have
> >    written them more than one in C over the years) and we certainly
> don't need to keep these close. And of course our small
> >    company is willing to pay reasonable amounts for a licence. The big
> problem these days seems to be finding out what you
>
> Kepping code hidden means keeping bugs hidden too so I guess part of the
> model
> depends on the application and how users can verify the code is doing
> something
> useful. In medicine, tests are often ambiguous due to the issues with
> biological
> samples and ultimately many seem like "blackboxes" but it may be highly
> desirable to avoid that with software. If you can keep it hidden and
> useful/testable/verifiable/trustworthy that may be a big accomplisment too.
>
> >    need to do regarding licensing--and knowing it will stick once it is
> done. Sometimes it is worse reading the US tax
> >    code--which grew much Kudzu  as computer languages and architectures
> do these days!
> >    cheers
> >    Fritz
> >
> >    On Thu, Mar 18, 2021 at 6:21 AM Mike Marchywka <[mailto:
> marchywka@hotmail.com]marchywka@hotmail.com> wrote:
> >
> >      Thanks. I've looked briefly at a lot of different kinds of
> "spectra" -
> >      audio, solar, image fft, distributions,  xps, even Raman that may
> evolve
> >      with time -
> >       and
> >      as you suggest you may not be interested so much in some abstract
> >      comparison as in extracting some model information. Comparing
> spectra
> >       may be with the intent of resolving a given one into component
> >      pieces- how much of each basis element is  in the  measured thing.
> >      Generally you have lines with some profile- gauss and lorentz would
> >      be well known - and then a continuum which could be anything
> >      with blackbody and I guess fluorescence as examples.  Then you have
> >      instrument issues to resolve- baseline and maybe broadening could
> >      be factors for a library.
> >      You could imagine developing a language around common things-
> >      consider maybe writing "R" packages that use GSL.  CRAN's R
> >      may be a good open replacement for MATLAB.
> >      I played with python briefly and any language that enforces white
> space,
> >      and IIRC earlier distinguished space and tab lol, is a bit
> >      of a suspect ...
> >      I've also run into various language-vs-library issues and thinking
> about business
> >      issues. I've got one "program" to make downloading citation
> information
> >      less distracting from diverse sources targeted at academics or
> anyone doing internet
> >      research ( this could be companies writing white papers or
> technical reports for their own products
> >      compared to competitors,  political hacks writing position or
> policy papers if the
> >      internet sites supply Bibtex for their works ). The code itself is
> almost the opposite
> >      of science- it is a collection of hacks tried in the order in which
> I discovered
> >      they may be useful to try to download citation information without
> bothering
> >      the user much. After looking at maybe 100's of hacks, some patterns
> emerged
> >      and in the conversion from an awful bash script to c/c++ it looked
> >      like you could come up with a mini-language based on "subroutine"
> >      or method calls.  The dev version uses readline for interaction
> which appears
> >      to have some licensing issues but since I almost always just write
> for
> >      myself I don't usually notice stuff like that.
> >      btw, as their are likely academics here if you have your own horror
> or success
> >      stories getting citation information for your publication efforts
> please share
> >      as appropriate here or on the texhax list . Thanks.
> >      note new address
> >       Mike Marchywka 306 Charles Cox Drive Canton, GA 30115
> >       2295 Collinworth  Drive Marietta GA 30062.  formerly 487 Salem
> Woods Drive Marietta GA 30067 404-788-1216 (C)<- leave
> >      message 989-348-4796 (P)<- emergency
> >      ________________________________________
> >      From: Fritz Sonnichsen <[mailto:sonnichs@gmail.com]
> sonnichs@gmail.com>
> >      Sent: Tuesday, March 16, 2021 9:53 AM
> >      To: Mike Marchywka
> >      Cc: [mailto:help-gsl@gnu.org]help-gsl@gnu.org
> >      Subject: Re: Checking GSL for Spectroscopy
> >      Mark
> >        I am converting someone's MATLAB code so I am not sure what he is
> doing yet--but several years ago I did spectral
> >      analysis in MATLAB and probably very similar. This is for Raman and
> LIBS spectra.
> >      1) "Usually" I apply a high pass filter to the spectrum. This gets
> rid of the noise I need control over this since as
> >      you would expect the signal and noise can get pretty close!
> Intuition comes into play here.
> >      2) Next I baseline the spectra. This removes any constant bias.
> For LIBS I was usually able to further filter "spikes"
> >      and then take a mean of the remaining line, subtracting this from
> the overall spectrum. Raman can get a bit more
> >      difficult-I am, at least,  subtracting the fluorescent line which
> can have a lot of features (e.g. spikes). At times, if
> >      you know this background you can subtract it first but you get all
> types of complications from normalization.
> >      Again--intuition comes into play.
> >      3) The resulting spectrum needs to be compared to a database. For
> LIBS the latter is quite small--mostly
> >      atomic/elemental data such as NIST. I could generally do a discrete
> comparison of the spike locations using a
> >      peak-finder, align them with the known examples and get a pretty
> high hit rate. This was for qualitative data.
> >      Raman is, again, much more complex. The data I was using was
> constrained and simpler but the case in hand here is much
> >      more complex. We are doing mixed plastics at the moment. My
> colleague found the best matches by taking a stats
> >      correlation with 44000 entries and pulling out the values closest
> to "one". It works remarkably well.
> >      I don't think there is much above that cannot be written in C in a
> reasonable amount of time. But we are looking ahead
> >      and would like to draw on the collective experience of the science
> community. This type of analysis is quite common and
> >      there are enough new wheels out there that we don't want to
> re-invent old ones!
> >          Very important is that "intuition" part. I would think a lot of
> this issue has been better solved since I was doing
> >      this. There are a lot of adjustments that could be made-for example
> iterating trial baselines, rejecting noise at varied
> >      levels etc. Processors are faster now and the AI movement has
> brought in PCA and a lot of other techniques that begin to
> >      transcend my current state of knowledge (I work more on the physics
> end of things and would prefer to use routines from
> >      the communities if possible to save time).
> >      Thanks for your interest Mark!
> >      Fritz
> >      On Tue, Mar 16, 2021 at 9:25 AM Mike Marchywka
> >      <[mailto:marchywka@hotmail.com]marchywka@hotmail.com<mailto:
> [mailto:marchywka@hotmail.com]marchywka@hotmail.com>> wrote:
> >      Can you comment on how you compare spectra? Just for my own
> >      personal interest, not sure if will further the thread here
> however..
> >      Not sure a "dot product" in the conventional sense would help much.
> >      You could imagine comparing peak positions and relative heights
> >      or a fit to a continuum for example.  Peaks plus black body in some
> >      vector comparison?
> >      note new address
> >       Mike Marchywka 306 Charles Cox Drive Canton, GA 30115
> >       2295 Collinworth  Drive Marietta GA 30062.  formerly 487 Salem
> Woods Drive Marietta GA 30067 404-788-1216 (C)<- leave
> >      message 989-348-4796 (P)<- emergency
> >      ________________________________________
> >      From: Help-gsl
> >      <help-gsl-bounces+marchywka=[mailto:hotmail.com@gnu.org]
> hotmail.com@gnu.org<mailto:[mailto:hotmail.com@gnu.org]hotmail.c
> >      om@gnu.org>> on behalf of Fritz Sonnichsen
> >      <[mailto:sonnichs@gmail.com]sonnichs@gmail.com<mailto:[mailto:
> sonnichs@gmail.com]sonnichs@gmail.com>>
> >      Sent: Tuesday, March 16, 2021 9:15 AM
> >      To: [mailto:help-gsl@gnu.org]help-gsl@gnu.org<mailto:[mailto:
> help-gsl@gnu.org]help-gsl@gnu.org>
> >      Subject: Checking GSL for Spectroscopy
> >      I am preparing to convert MATLAB code to something more general.
> The new
> >      code will run on LInux and ARM processors.
> >         For a lot of reasons I am not going to use Python. We also want
> to
> >      keep this project "close" to scientists and do not want to turn it
> into a
> >      full time computer programming job. So the final word is that I am
> looking
> >      for something that can be called by (and hopefully is written) in
> C. Worse
> >      case I will just write the code myself but would prefer to start
> >      integrating our systems into something with a lot of pre-written
> and vetted
> >      routines.
> >      GSL looks like a good choice. Maybe R comes next. We have a mix of
> needs
> >      but I will point out a few:
> >      1) Baselining a spectrum
> >      2) Finding peaks in that spectrum
> >      3) using Pearson correlation to compare the spectrum QUICKLY to
> >      about 50,000 recorded examples.
> >      We also have some uses with basic statistics and we do some image
> >      processing.
> >      So my question is--does GSL position itself in these areas? MATLAB
> (with
> >      packages) does them all.
> >           I am not sure how active GSL, if it is keeping up with AI,
> imaging and
> >      spectroscopy--or is it fading or giving way to popular languages for
> >      example. I was surprised that the 600+ page manual did not seem to
> show
> >      anything relating to the simple spectral analysis described above
> for
> >      example. Certainly I can search the web for others' code but at
> some point
> >      if I cannot attach to a well established product I will just write
> it
> >      myself.
> >      Any comments appreciated
> >      thanks
> >      Fritz
>
> --
>
> mike marchywka
> 306 charles cox
> canton GA 30115
> USA, Earth
> marchywka@hotmail.com
> 404-788-1216
> ORCID: 0000-0001-9237-455X
>
>


reply via email to

[Prev in Thread] Current Thread [Next in Thread]