batch-liblinear

#!/usr/bin/env bash
#
# usage: batch-liblinear < 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 on the side of the problem LIBSVM does not cover:
# a general-purpose :MILPSolver on the abstract representation, the ad hoc
# SMOSolver and LIBLINEAR on the physical one. They address the same training
# problem, hence they have to agree on its optimal value.
#
# LIBLINEAR trains a linear model by working in the weights rather than in the
# multipliers, and what it can express is the complement of what LIBSVM can:
# the linear kernel only, hence K == 0 alone, and the regularised bias only,
# hence the bit 1 of wf always on, but both losses, hence wf == 2 and wf == 3,
# where batch-libsvm only has the linear one.
#
# Its dual solvers are capped at 300 iterations by their own sources, whatever
# tolerance they are given, so the runs with the linear loss usually stop
# there: LIBLINEARSolver reports kLowPrecision, and the cross-check then holds
# it to the only thing it claims, i.e., that the optimum is not above the
# value it returns.

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

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

# 2 = regularised bias and linear loss, 3 = regularised bias and squared one
for wf in 2 3; 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 "$wf" -K 0 -C "$C" \
               -N 60 -M 4 -n 3 -e 1 $task

    done
  done
done

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