public static DataSparse Dense2Sparse(DataDense data)
{
return((data == null) ? null : new DataSparse()
{
Features = DatasetTest.Dense2Sparse(data.Features),
Labels = data.Labels,
Weights = data.Weights,
Groups = data.Groups
});
}
public void BenchmarkEval()
{
var rand = new Random(Seed);
int numColumns = 100;
var pms = new Parameters();
pms.Objective.Objective = ObjectiveType.Binary;
pms.Dataset.MaxBin = 63;
pms.Learning.LearningRate = 1e-3;
pms.Learning.NumIterations = 1000;
pms.Common.DeviceType = DeviceType.CPU;
var categorical = new Dictionary <int, int>(); // i.e., no cat
var trainData = CreateRandomDenseClassifyData(rand, 2, ref categorical, pms.Dataset.UseMissing, numColumns);
DataDense validData = null;
pms.Dataset.CategoricalFeature = categorical.Keys.ToArray();
using (var datasets = new Datasets(pms.Common, pms.Dataset, trainData, validData))
using (var trainer = new BinaryTrainer(pms.Learning, pms.Objective))
{
var model = trainer.Train(datasets);
output.WriteLine($"MaxNumTrees={model.Managed.MaxNumTrees}");
var timer = System.Diagnostics.Stopwatch.StartNew();
model.Native.GetOutputs(trainData.Features);
var elapsed1 = timer.Elapsed;
output.WriteLine($"EvalNativeMulti={elapsed1.TotalMilliseconds}");
timer.Restart();
foreach (var row in trainData.Features)
{
trainer.Evaluate(Booster.PredictType.Normal, row);
}
var elapsed2 = timer.Elapsed;
output.WriteLine($"EvalNativeSingle={elapsed2.TotalMilliseconds}");
foreach (var maxThreads in new int[] { 1, 2, 4, 8, 16, 32, Environment.ProcessorCount }) //
{
model.Managed.MaxThreads = maxThreads;
timer.Restart();
foreach (var row in trainData.Features)
{
double output = 0;
var input = new VBuffer <float>(row.Length, row);
model.Managed.GetOutput(ref input, ref output);
}
var elapsed3 = timer.Elapsed;
output.WriteLine($"MaxThreads={maxThreads} EvalManaged={elapsed3.TotalMilliseconds}");
}
}
}
//[Fact]
public void BenchmarkBinary()
{
var rand = new Random(Seed);
for (int test = 0; test < 3; ++test)
{
for (int gpu = 0; gpu < 2; gpu++)
{
int numColumns = 50 * (test + 1);
var pms = new Parameters();
pms.Objective.Objective = ObjectiveType.Binary;
pms.Dataset.MaxBin = 63;
pms.Learning.BaggingFraction = 1;
pms.Learning.BaggingFreq = 1;
pms.Learning.LearningRate = 1e-3;
pms.Learning.NumIterations = 10;
pms.Common.DeviceType = (gpu > 0) ? DeviceType.GPU : DeviceType.CPU;
var categorical = new Dictionary <int, int>(); // i.e., no cat
var trainData = CreateRandomDenseClassifyData(rand, 2, ref categorical, pms.Dataset.UseMissing, numColumns);
DataDense validData = null;
pms.Dataset.CategoricalFeature = categorical.Keys.ToArray();
try
{
using (var datasets = new Datasets(pms.Common, pms.Dataset, trainData, validData))
using (var trainer = new BinaryTrainer(pms.Learning, pms.Objective))
{
var timer = System.Diagnostics.Stopwatch.StartNew();
var model = trainer.Train(datasets);
var elapsed = timer.Elapsed;
output.WriteLine($"{pms.Common.DeviceType}: NumRows={trainData.NumRows} NumCols={numColumns} MaxNumTrees={model.Managed.MaxNumTrees} TrainTimeSecs={elapsed.TotalSeconds}");
}
}
catch (Exception e)
{
throw new Exception($"Failed: {Seed} #{test} {pms}", e);
}
}
}
}
public static DataDense CreateRandomDenseData(
Random rand,
ref Dictionary <int, int> categorical,
bool useMissing,
int numColumns
)
{
var numRows = rand.Next(100, 500);
// from column index to number of classes
if (categorical == null)
{
categorical = new Dictionary <int, int>();
if (rand.Next(2) == 0)
{
for (int j = 0; j < numColumns; j++)
{
if (rand.Next(10) == 0)
{
categorical.Add(j, rand.Next(3, 100));
}
}
}
}
var scales = Enumerable.Range(0, numColumns).Select(x => Math.Pow(10.0, rand.Next(-10, 10))).ToArray();
var rows = new float[numRows][];
var weights = (rand.Next(2) == 0) ? new float[numRows] : null;
for (int i = 0; i < numRows; ++i)
{
var row = new float[numColumns];
for (int j = 0; j < row.Length; ++j)
{
if (useMissing && rand.Next(50) == 0)
{
row[j] = float.NaN;
}
else
{
if (categorical.TryGetValue(j, out int numClass))
{
row[j] = rand.Next(numClass);
}
else
{
row[j] = (rand.Next(100) == 0) ? 0.0f : (float)(scales[j] * (rand.NextDouble() - 0.5));
}
}
}
rows[i] = row;
if (weights != null)
{
weights[i] = (float)rand.NextDouble();
}
}
var rslt = new DataDense
{
Features = rows,
Weights = weights,
Groups = null
};
return(rslt);
}
