Using vw hypersearch - martinpopel/vowpal_wabbit GitHub Wiki

Introduction

vw-hypersearch is a vw wrapper to help in finding lowest-loss hyper-parameters (argmin).

For example: to find the lowest average loss for --l1 (L1-norm regularization) on a train-set called train.dat:

    $ vw-hypersearch  1e-10  5e-4  vw  --l1 %  train.dat

vw-hypersearch trains multiple times (but in a efficient way) until it finds the --l1 value resulting in the lowest average training loss.

Explanation of the example:

  • the % character is a placeholder for the (argmin) parameter we are looking for.
  • 1e-10 is the lower-bound for the search range
  • 5e-4 is the upper-bound of the search range

The lower & upper bounds are arguments to vw-hypersearch. Anything after vw is a normal vw argument, exactly as you would use without vw-hypersearch. The only difference in the vw training command is the use of a % placeholder instead of the value of the parameter we're trying to optimize on.

Calling vw-hypersearch without any arguments prints a usage message.

More advanced options

Additional arguments which can be passed to vw-hypersearch before vw itself:

  • -L do a log-space golden-section search instead of a simple golden-section search.
  • -t test.dat (note: this must come before the vw argument): search for the training parameter that results in a minimum loss on test.dat rather than train.dat (ignores the training loss)
  • -c test.dat.cache: use test.dat.cache as test-cache file + evaluate goodness on it (this implies -t except the cache test.dat.cache will be used instead of test.dat)
  • -e <script_name> or even -e '<script_name> <script_args>...': use an external program <script_name> and try to minimize its last numeric output. This allows you to plug-in any external tool for argmin. For example, you could write a script that calls vw with the generated model, in prediction mode (-p somefile), and then use the KDD perf utility to calculate accuracy, various AUCs etc. on the predictions. The only interface between vw-hypersearch and <script_name> is the last numeric value printed to stdout by <script_name> which vw-hypersearch will try to minimize. If you'd rather maximize the argmin (default behavior is to minimize), you can simply negate the value inside <script_name> before printing it.
  • An optional, 3rd numeric parameter to vw-hypersearch, will be interpreted as a tolerance parameter directing vw-hypersearch to stop only when the loss (or other metric) difference in two consecutive run errors is less than tolerance

Examples

# Find the learning-rate resulting in the lowest average-loss
# for a logistic loss train-set:
vw-hypersearch 0.1 100 vw --loss_function logistic --learning_rate % train.dat

# Find the bootstrap value resulting in the lowest average-loss
# vw-hypersearch will automatically search in integer-space
# since --bootstrap expects an integer
vw-hypersearch 2 16 vw --bootstrap % train.dat

Implementation notes

vw-hypersearch conducts a golden-section search search by default. This search method strikes a good balance between safety and efficiency.

Caveats

  • Lowest average loss is not necessarily optimal
  • Your real goal should always be to find a minimal generalization error, not training error.
  • Some parameters do not have a convex loss, for these vw-hypersearch will converge on some local-minimum instead of a global one

More questions?

vw-hypersearch is written in perl and is included with vowpal wabbit (in the utl subdirectory). In case of doubt, look at the source.

Credits

#   * Paul Mineiro ('vowpalwabbit/demo') wrote the original script
#   * Ariel Faigon
#       - Generalize, rewrite, refactor
#       - Add usage message
#       - Add tolerance as optional parameter
#       - Add documentation
#       - Better error messages when underlying command fails for any reason
#       - Add '-t test-set' to optimize on test error rather than train-set error
#       - Add integral-value option support
#       - Add external plugin support via -e ...
#       - More reliable/informative progress indication
#       Bug fixes:
#           - Golden-section search bug-fix
#           - Loss value capture bug-fix (can be in scientific notation)
#           - Handle special cases where certain options don't make sense
#   * Alex Hudek:
#       - Support log-space search which seems to work better
#         with --l1 (very small values) and/or hinge-loss
#   * Alex Trufanov:
#       - A bunch of very useful bug reports:
#           https://github.com/JohnLangford/vowpal_wabbit/issues/406