Skip to content

KNN-regressor n-d output #214

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 14 commits into from
Feb 21, 2023
Merged

KNN-regressor n-d output #214

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
14 commits
Select commit Hold shift + click to select a range
8bc4e97
in progress - stalled with eigen bridging
tremblap Jan 26, 2023
fe499ce
corrections - compiling and almost working
tremblap Jan 31, 2023
266672a
algo - post owen review 1
tremblap Feb 2, 2023
3addb52
handrolled application of weights = expected results
tremblap Feb 2, 2023
84ae85a
fixed the crash in predict and zero the output vector in algo
tremblap Feb 2, 2023
caaa929
corrected the ratio in case of no-weighting
tremblap Feb 2, 2023
a0ea138
added the RT allocator to the RT query
tremblap Feb 2, 2023
8dd32c1
sneaky KR size check was wrong...
tremblap Feb 2, 2023
139d259
Merge remote-tracking branch 'origin/dev' into KNNregressor-multiout
tremblap Feb 14, 2023
36e2d10
Merge branch 'dev' into KNNregressor-multiout
tremblap Feb 14, 2023
1dea3e9
Merge branch 'dev' into KNNregressor-multiout
tremblap Feb 14, 2023
8af6399
Merge branch 'main' into KNNregressor-multiout
tremblap Feb 21, 2023
a4d74c2
made use of auto to avoid allocation in the loop
tremblap Feb 21, 2023
e0bb9da
now thanks to @weefuzzy a proper vectorised process
tremblap Feb 21, 2023
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
58 changes: 30 additions & 28 deletions include/algorithms/public/KNNRegressor.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -29,46 +29,48 @@ class KNNRegressor
public:
using DataSet = FluidDataSet<std::string, double, 1>;

double predict(KDTree const& tree, DataSet const& targets,
RealVectorView point, index k, bool weighted,
void predict(KDTree const& tree, DataSet const& targets,
RealVectorView input, RealVectorView output,
index k, bool weighted,
Allocator& alloc = FluidDefaultAllocator()) const
{
using namespace std;
double prediction = 0;
auto [distances, ids] = tree.kNearest(point, k, 0, alloc);
double uniformWeight = 1.0 / k;
rt::vector<double> weights(asUnsigned(k), weighted ? 0 : uniformWeight,
alloc);
double sum = 0;
using _impl::asEigen;
using Eigen::Array;

auto [distances, ids] = tree.kNearest(input, k, 0, alloc);

ScopedEigenMap<Eigen::VectorXd> weights(k, alloc);
weights.setConstant(weighted ? 0 : (1.0 / k));

if (weighted)
{
bool binaryWeights = false;
for (size_t i = 0; i < asUnsigned(k); i++)
auto distanceArray =
Eigen::Map<Eigen::ArrayXd>(distances.data(), distances.size());

if ((distanceArray < epsilon).any())
{
if (distances[i] < epsilon)
{
binaryWeights = true;
weights[i] = 1;
}
else
sum += (1.0 / distances[i]);
weights = (distanceArray < epsilon).select(1.0, weights);
}
if (!binaryWeights)
else
{
for (size_t i = 0; i < asUnsigned(k); i++)
{
weights[i] = (1.0 / distances[i]) / sum;
}
double sum = (1.0 / distanceArray).sum();
weights = (1.0 / distanceArray) / sum;
}
}

output.fill(0);

for (size_t i = 0; i < asUnsigned(k); i++)
{
auto point = targets.get(*ids[i]);
prediction += (weights[i] * point(0));
}
return prediction;
rt::vector<index> indices(ids.size(), alloc);

transform(ids.cbegin(), ids.cend(), indices.begin(),
[&targets](const string* id) { return targets.getIndex(*id); });

auto targetPoints = asEigen<Array>(targets.getData())(indices, Eigen::all);

asEigen<Array>(output) = (targetPoints.colwise() * weights.array()).colwise().sum().transpose();
}
};
} // namespace algorithm
} // namespace fluid

57 changes: 32 additions & 25 deletions include/clients/nrt/KNNRegressorClient.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -119,23 +119,29 @@ class KNNRegressorClient : public FluidBaseClient,
return {};
}

MessageResult<double> predictPoint(InputBufferPtr data) const
MessageResult<void> predictPoint(InputBufferPtr in, BufferPtr out) const
{
index k = get<kNumNeighbors>();
bool weight = get<kWeight>() != 0;
if (k == 0) return Error<double>(SmallK);
if (mAlgorithm.tree.size() == 0) return Error<double>(NoDataFitted);
if (mAlgorithm.tree.size() < k) return Error<double>(NotEnoughData);
if (k == 0) return Error(SmallK);
if (mAlgorithm.tree.size() == 0) return Error(NoDataFitted);
if (mAlgorithm.tree.size() < k) return Error(NotEnoughData);

InBufferCheck bufCheck(mAlgorithm.tree.dims());
if (!bufCheck.checkInputs(data.get()))
return Error<double>(bufCheck.error());
if (!bufCheck.checkInputs(in.get()))
return Error(bufCheck.error());
BufferAdaptor::ReadAccess inBuf(in.get());
BufferAdaptor::Access outBuf(out.get());
if (!outBuf.exists()) return Error(InvalidBuffer);
Result resizeResult = outBuf.resize(mAlgorithm.target.dims(), 1, inBuf.sampleRate());
if (!resizeResult.ok()) return Error(BufferAlloc);
algorithm::KNNRegressor regressor;
RealVector point(mAlgorithm.tree.dims());
point <<= BufferAdaptor::ReadAccess(data.get())
.samps(0, mAlgorithm.tree.dims(), 0);
double result =
regressor.predict(mAlgorithm.tree, mAlgorithm.target, point, k, weight);
return result;
RealVector input(mAlgorithm.tree.dims());
RealVector output(mAlgorithm.target.dims());
input <<= inBuf.samps(0, mAlgorithm.tree.dims(), 0);
regressor.predict(mAlgorithm.tree, mAlgorithm.target, input, output, k, weight);
outBuf.samps(0) <<= output;
return OK();
}

MessageResult<void> predict(InputDataSetClientRef source,
Expand All @@ -158,12 +164,12 @@ class KNNRegressorClient : public FluidBaseClient,
algorithm::KNNRegressor regressor;
auto ids = dataSet.getIds();
auto data = dataSet.getData();
DataSet result(1);
DataSet result(mAlgorithm.target.dims());
RealVector prediction(mAlgorithm.target.dims());
for (index i = 0; i < dataSet.size(); i++)
{
RealVectorView point = data.row(i);
RealVector prediction = {regressor.predict(
mAlgorithm.tree, mAlgorithm.target, point, k, weight)};
regressor.predict(mAlgorithm.tree, mAlgorithm.target, point, prediction, k, weight);
result.add(ids(i), prediction);
}
destPtr->setDataSet(result);
Expand Down Expand Up @@ -226,11 +232,11 @@ class KNNRegressorQuery : public FluidBaseClient, ControlIn, ControlOut


template <typename T>
void process(std::vector<FluidTensorView<T, 1>>& input,
std::vector<FluidTensorView<T, 1>>& output, FluidContext&)
void process(std::vector<FluidTensorView<T, 1>>& in,
std::vector<FluidTensorView<T, 1>>& out, FluidContext& c)
{
output[0] <<= input[0];
if (input[0](0) > 0)
out[0] <<= in[0];
if (in[0](0) > 0)
{
auto knnPtr = get<kModel>().get().lock();
if (!knnPtr)
Expand All @@ -248,17 +254,18 @@ class KNNRegressorQuery : public FluidBaseClient, ControlIn, ControlOut
get<kOutputBuffer>().get()))
return;
auto outBuf = BufferAdaptor::Access(get<kOutputBuffer>().get());
if (outBuf.samps(0).size() != 1) return;
if (outBuf.samps(0).size() != algorithm.target.dims()) return;

algorithm::KNNRegressor regressor;

RealVector point(algorithm.tree.dims());
point <<= BufferAdaptor::ReadAccess(get<kInputBuffer>().get())
RealVector input(algorithm.tree.dims(), c.allocator());
RealVector output(algorithm.target.dims(), c.allocator());

Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

These, on the other hand, probably would benefit from being allocated elsewhere, and using an allocator

input <<= BufferAdaptor::ReadAccess(get<kInputBuffer>().get())
.samps(0, algorithm.tree.dims(), 0);

double result =
regressor.predict(algorithm.tree, algorithm.target, point, k, weight);
outBuf.samps(0)[0] = result;
regressor.predict(algorithm.tree, algorithm.target, input, output, k, weight, c.allocator());
outBuf.samps(0) <<= output;
}
}

Expand Down