apaam/mlpack
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

I've concluded that this test isn't useful.

Browse Source 
Basically, the test is just checking that the implementation is exactly the same
as in the test, and that's not actually a great test, because the implementation
may change without breaking anything.
This commit is contained in:
Ryan Curtin
2015-03-02 20:24:04 -05:00
parent 2b76e6d0ac
commit b7769dcffb
1 changed files with 0 additions and 72 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/mlpack/tests/allkrann_search_test.cpp
-72
View File
@@ -23,78 +23,6 @@ using namespace mlpack::bound;
BOOST_AUTO_TEST_SUITE(AllkRANNTest);
// Test AllkRANN in naive mode for exact results when the random seeds are set
// the same. This may not be the best test; if the implementation of RANN-RS
// gets random numbers in a different way, then this test might fail.
BOOST_AUTO_TEST_CASE(NaiveSearchExact)
{
// First test on a small set.
arma::mat rdata(2, 10);
rdata << 3 << 2 << 4 << 3 << 5 << 6 << 0 << 8 << 3 << 1 << arma::endr <<
0 << 3 << 4 << 7 << 8 << 4 << 1 << 0 << 4 << 3 << arma::endr;
arma::mat qdata(2, 3);
qdata << 3 << 2 << 0 << arma::endr
<< 5 << 3 << 4 << arma::endr;
metric::SquaredEuclideanDistance dMetric;
double rankApproximation = 30;
double successProb = 0.95;
// Search for 1 rank-approximate nearest-neighbors in the top 30% of the point
// (rank error of 3).
arma::Mat<size_t> neighbors;
arma::mat distances;
// Test naive rank-approximate search.
// Predict what the actual RANN-RS result would be.
math::RandomSeed(0);
size_t numSamples = (size_t) ceil(log(1.0 / (1.0 - successProb)) /
log(1.0 / (1.0 - (rankApproximation / 100.0))));
arma::Mat<size_t> samples(qdata.n_cols, numSamples);
for (size_t j = 0; j < qdata.n_cols; j++)
for (size_t i = 0; i < numSamples; i++)
samples(j, i) = (size_t) math::RandInt(10);
arma::Col<size_t> rann(qdata.n_cols);
arma::vec rannDistances(qdata.n_cols);
rannDistances.fill(DBL_MAX);
for (size_t j = 0; j < qdata.n_cols; j++)
{
for (size_t i = 0; i < numSamples; i++)
{
double dist = dMetric.Evaluate(qdata.unsafe_col(j),
rdata.unsafe_col(samples(j, i)));
if (dist < rannDistances[j])
{
rann[j] = samples(j, i);
rannDistances[j] = dist;
}
}
}
// Use RANN-RS implementation.
math::RandomSeed(0);
RASearch<> naive(rdata, qdata, true);
naive.Search(1, neighbors, distances, rankApproximation);
// Things to check:
//
// 1. (implicitly) The minimum number of required samples for guaranteed
// approximation.
// 2. (implicitly) Check the samples obtained.
// 3. Check the neighbor returned.
for (size_t i = 0; i < qdata.n_cols; i++)
{
BOOST_REQUIRE_EQUAL(neighbors(0, i), rann[i]);
BOOST_REQUIRE_CLOSE(distances(0, i), rannDistances[i], 1e-5);
}
}
// Test the correctness and guarantees of AllkRANN when in naive mode.
BOOST_AUTO_TEST_CASE(NaiveGuaranteeTest)
{