Changeset 82e5fb

Timestamp:

May 25, 2021, 11:04:50 PM (4 years ago)

Author:

Frederik Heber <frederik.heber@…>

Branches:

Candidate_v1.7.0, stable

Children:

91793c

Parents:

889067e

git-author:

Frederik Heber <frederik.heber@…> (05/25/21 21:55:59)

git-committer:

Frederik Heber <frederik.heber@…> (05/25/21 23:04:50)

Message:

Added option error-file to fit potential actions.

• DOCU: Added option to userguide.
• TESTS: Added option to FitCompoundPotential regression test.

Files:

• doc/userguide/userguide.xml (modified) (2 diffs)
• src/Actions/PotentialAction/FitCompoundPotentialAction.cpp (modified) (1 diff)
• src/Actions/PotentialAction/FitCompoundPotentialAction.def (modified) (1 diff)
• src/Actions/PotentialAction/FitPotentialAction.cpp (modified) (1 diff)
• src/Actions/PotentialAction/FitPotentialAction.def (modified) (1 diff)
• src/Potentials/PotentialTrainer.cpp (modified) (2 diffs)
• src/Potentials/PotentialTrainer.hpp (modified) (1 diff)
• tests/Python/AllActions/options.dat (modified) (1 diff)
• tests/regression/Potential/FitCompoundPotential/testsuite-potential-fit-compound-potential.at (modified) (1 diff)

doc/userguide/userguide.xml

@@ -2430 +2430 @@
       --set-threshold 1e-3 \
       --training-file test.dat
+      --error-file error.dat
    </programlisting>
           <para>Now, all empirical potential functions are summed up into a
@@ -2437 +2438 @@
           potential between oxygen and hydrogen and another angular potential
           for the angle between hydrogen, oxygen, and hydrogen atom. Then, we would
-          fit a function consisting of the sum of the two potentials functions in order to approximate the energy of a single
-          water molecule (actually, it&apos;s the sum of three potentials. As mentioned before, a constant potential is always added to compensate non-bonding energies, i.e. not depending on interatomic distances). Here, the threshold criterion takes the place of the
-<emphasis role="bold"> take-best-of</emphasis> option. Here, the minimization is reiterated so often on random (but to some extent chosen from a sensible range) starting parameters until the final L2 error is below 1e-3. Also, all data used
-          for training, i.e. the tuples consisting of the fragments nuclei
-          coordinates and the associated energy value are written to the file
-          <filename>test.dat</filename>. This allows for graphical representation or other
-way of analysis, e.g. for a Morse potential between oxygen and hydrogen the bonding energy can be plotted as a one-dimensional function and compared to the &quot;point cloud&quot; of sample points from the fragment term of Born-Oppenheimer surface. It is this point cloud, i.e. the training data, that is written to the file
-          <filename>test.dat</filename>.</para>
+          fit a function consisting of the sum of the two potentials functions in 
+          order to approximate the energy of a single water molecule (actually, 
+          it&apos;s the sum of three potentials. As mentioned before, a constant 
+          potential is always added to compensate non-bonding energies, i.e. not 
+          depending on interatomic distances). Here, the threshold criterion takes 
+          the place of the<emphasis role="bold"> take-best-of</emphasis> option. 
+          Here, the minimization is reiterated so often on random (but to some 
+          extent chosen from a sensible range) starting parameters until the final 
+          L2 error is below 1e-3. Also, all data used for training, i.e. the tuples 
+          consisting of the fragments nuclei coordinates and the associated energy 
+          value are written to the file <filename>test.dat</filename>. This allows 
+          for graphical representation or other way of analysis, e.g. for a Morse 
+          potential between oxygen and hydrogen the bonding energy can be plotted 
+          as a one-dimensional function and compared to the &quot;point cloud&quot; 
+          of sample points from the fragment term of Born-Oppenheimer surface. 
+          It is this point cloud, i.e. the training data, that is written to the file
+          <filename>test.dat</filename>. Moreover, the error per fragment is written
+          to the file <filename>error.dat</filename></para>
           <para>Note that you can combine the two ways, i.e. start with a
           fit-potential call but give an empty potential file. The resulting

src/Actions/PotentialAction/FitCompoundPotentialAction.cpp

@@ -104 +104 @@
       graph,
       params.training_file.get(),
+      params.error_file.get(),
       params.max_iterations.get(),
       params.threshold.get(),

src/Actions/PotentialAction/FitCompoundPotentialAction.def

@@ -23 +23 @@
 // ValueStorage by the token "Z" -> first column: int, Z, "Z"
 // "undefine" if no parameters are required, use (NOPARAM_DEFAULT) for each (undefined) default value
-#define paramtypes (boost::filesystem::path)(std::vector<const element *>)(unsigned int)(double)(unsigned int)
-#define paramtokens ("training-file")("fragment-charges")("take-best-of")("set-threshold")("set-max-iterations")
-#define paramdescriptions ("optional file to write training data to")("charges specifying the fragment")("take the best among this many approximations")("Require L2 error to be smaller than threshold, overrides number of attempts")("sets the maximum number of iterations spent on the optimization of parameters")
-#define paramdefaults (PARAM_DEFAULT(""))(NOPARAM_DEFAULT)(PARAM_DEFAULT(3))(PARAM_DEFAULT(1.))(PARAM_DEFAULT(100))
-#define paramreferences (training_file)(fragment)(best_of_howmany)(threshold)(max_iterations)
+#define paramtypes (boost::filesystem::path)(boost::filesystem::path)(std::vector<const element *>)(unsigned int)(double)(unsigned int)
+#define paramtokens ("training-file")("error-file")("fragment-charges")("take-best-of")("set-threshold")("set-max-iterations")
+#define paramdescriptions ("optional file to write training data to")("optional file to write error data to")("charges specifying the fragment")("take the best among this many approximations")("Require L2 error to be smaller than threshold, overrides number of attempts")("sets the maximum number of iterations spent on the optimization of parameters")
+#define paramdefaults (PARAM_DEFAULT(""))(PARAM_DEFAULT(""))(NOPARAM_DEFAULT)(PARAM_DEFAULT(3))(PARAM_DEFAULT(1.))(PARAM_DEFAULT(100))
+#define paramreferences (training_file)(error_file)(fragment)(best_of_howmany)(threshold)(max_iterations)
 #define paramvalids \
+(DummyValidator<boost::filesystem::path>()) \
 (DummyValidator<boost::filesystem::path>()) \
 (STLVectorValidator< std::vector<const element *> >(1,99, ElementValidator())) \

src/Actions/PotentialAction/FitPotentialAction.cpp

@@ -141 +141 @@
       graph,
       params.training_file.get(),
+      params.error_file.get(),
       params.max_iterations.get(),
       params.threshold.get(),

src/Actions/PotentialAction/FitPotentialAction.def

@@ -23 +23 @@
 // ValueStorage by the token "Z" -> first column: int, Z, "Z"
 // "undefine" if no parameters are required, use (NOPARAM_DEFAULT) for each (undefined) default value
-#define paramtypes (boost::filesystem::path)(std::string)(std::vector<const element *>)(std::vector<const element *>)(unsigned int)(double)(unsigned int)
-#define paramtokens ("training-file")("potential-type")("potential-charges")("fragment-charges")("take-best-of")("set-threshold")("set-max-iterations")
-#define paramdescriptions ("optional file to write training data to")("potential type to fit")("charges specifying the potential")("charges specifying the fragment")("take the best among this many approximations")("Require L2 error to be smaller than threshold, overrides number of attempts")("sets the maximum number of iterations spent on the optimization of parameters")
-#define paramdefaults (PARAM_DEFAULT(""))(NOPARAM_DEFAULT)(NOPARAM_DEFAULT)(NOPARAM_DEFAULT)(PARAM_DEFAULT(3))(PARAM_DEFAULT(1.))(PARAM_DEFAULT(100))
-#define paramreferences (training_file)(potentialtype)(charges)(fragment)(best_of_howmany)(threshold)(max_iterations)
+#define paramtypes (boost::filesystem::path)(boost::filesystem::path)(std::string)(std::vector<const element *>)(std::vector<const element *>)(unsigned int)(double)(unsigned int)
+#define paramtokens ("training-file")("error-file")("potential-type")("potential-charges")("fragment-charges")("take-best-of")("set-threshold")("set-max-iterations")
+#define paramdescriptions ("optional file to write training data to")("optional file to write error data to")("potential type to fit")("charges specifying the potential")("charges specifying the fragment")("take the best among this many approximations")("Require L2 error to be smaller than threshold, overrides number of attempts")("sets the maximum number of iterations spent on the optimization of parameters")
+#define paramdefaults (PARAM_DEFAULT(""))(PARAM_DEFAULT(""))(NOPARAM_DEFAULT)(NOPARAM_DEFAULT)(NOPARAM_DEFAULT)(PARAM_DEFAULT(3))(PARAM_DEFAULT(1.))(PARAM_DEFAULT(100))
+#define paramreferences (training_file)(error_file)(potentialtype)(charges)(fragment)(best_of_howmany)(threshold)(max_iterations)
 #define paramvalids \
+(DummyValidator<boost::filesystem::path>()) \
 (DummyValidator<boost::filesystem::path>()) \
 (PotentialTypeValidator()) \

src/Potentials/PotentialTrainer.cpp

@@ -71 +71 @@
     const HomologyGraph &_graph,
     const boost::filesystem::path &_trainingfile,
+    const boost::filesystem::path &_errorfile,
     const unsigned int _maxiterations,
     const double _threshold,
@@ -180 +181 @@
     TrainingData::L2ErrorConfigurationIndexMap_t WorseFragmentMap =
         data.getWorstFragmentMap(model, fragmentrange);
-    LOG(0, "RESULT: WorstFragmentMap " << WorseFragmentMap << ".");
-
+    if (_errorfile.string().empty()) {
+      LOG(0, "RESULT: WorstFragmentMap " << WorseFragmentMap << ".");
+    } else {
+      std::ofstream errorstream(_errorfile.string().c_str());
+      if (errorstream.good()) {
+        LOG(3, "DEBUG: Writing error data to file " <<
+            _errorfile.string() << ".");
+        errorstream << "step\terror" << std::endl;
+        // resort into step as key
+        typedef std::map< size_t, double > step_error_t;
+        step_error_t step_error;
+        for (TrainingData::L2ErrorConfigurationIndexMap_t::const_reverse_iterator iter = WorseFragmentMap.rbegin();
+            iter != WorseFragmentMap.rend(); ++iter)
+          step_error.insert( std::make_pair(iter->second, iter->first) );
+        for (step_error_t::const_iterator iter = step_error.begin();
+            iter != step_error.end(); ++iter)
+          errorstream << iter->first << "\t" << iter->second << std::endl;
+      }
+      errorstream.close();
+    }
   }
 

src/Potentials/PotentialTrainer.hpp

@@ -36 +36 @@
         const HomologyGraph &_graph,
         const boost::filesystem::path &_trainingfile,
+        const boost::filesystem::path &_errorfile,
         const unsigned int _maxiterations,
         const double _threshold,

tests/Python/AllActions/options.dat

@@ -78 +78 @@
 end_step        "1"
 enforce_net_zero_charge "0"
+error_file      ""
 ExcludeHydrogen "1"
 fastparsing     "1"

tests/regression/Potential/FitCompoundPotential/testsuite-potential-fit-compound-potential.at

@@ -35 +35 @@
         --parse-potentials ${abs_top_srcdir}/tests/regression/Potential/FitCompoundPotential/pre/water.potentials \
         --fit-compound-potential \
+                --error-file error.csv \
                 --fragment-charges 1 8 1 \
                 --take-best-of 3 \
         --save-potentials water.potentials], 0, [stdout], [ignore])
+# check that error file was written
+AT_CHECK([test -e error.csv], 0, [ignore], [ignore])
 # check that L_2 error
 AT_CHECK([grep "Best parameters with L2 error" stdout | awk '{if ($8 > 0.34) exit 1}'], 0, [ignore], [ignore])