batch-libsvm

#!/usr/bin/env bash
#
# usage: batch-libsvm < exe file > < log file >
#        < exe file > = path of the executable file, default ./SVM_test
#        < log file > = output log file (no log if absent)
#
# the three-way cross-check: a general-purpose :MILPSolver on the abstract
# representation, the ad hoc SMOSolver and LIBSVM, the reference
# implementation, on the physical one. They address the same training problem,
# hence they have to agree on its optimal value, and LIBSVM being an
# independent implementation of the very algorithm SMOSolver implements makes
# it the natural reference to check the latter against.
#
# The sweep is smaller than that of batch-dual because LIBSVM only trains the
# problems it can express: the linear loss with the bias not regularised
# (hence wf == 0 alone) and any kernel but the Laplacian one (hence K == 3 is
# left out, as is the sigmoid one, which is not a Mercer kernel and on which
# no two Solver need agree). It is also run on more, and larger, data sets
# than the other batches: LIBSVM is fast, so what is worth exercising here is
# the agreement on many different problems.

# shared CLI/run helpers - - - - - - - - - - - - - - - - - - - - - - - - - -
source "$(dirname "$0")/../../batch_common.sh"

DEFAULT_EXE=./SVM_test
DEFAULT_PAR=BSPar-LSVM.txt
parse_batch_args "$@"

# 0 = linear, 1 = polynomial, 2 = gaussian
for K in 0 1 2; do

  # classification and regression
  for task in "" "-g"; do
    for C in 0.1 1 10; do

      run_test "$exe" -S "$par" -B SVMCfg-dual.txt -f 0 -K "$K" -C "$C" \
               -N 60 -M 4 -n 3 -e 1 $task

    done
  done
done

# the end - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -