TEST(MatrixBatchTransTest, test_batch_matrix_transpose) {
const int nx = 100;
const int ny = 50;
const int numSamples = 50;
MatrixPtr cMat = Matrix::create(numSamples, nx * ny, false, false);
MatrixPtr gMat = Matrix::create(numSamples, nx * ny, false, true);
MatrixPtr cBatchTransMat = Matrix::create(numSamples, nx * ny, false, false);
MatrixPtr gBatchTransMat = Matrix::create(numSamples, nx * ny, false, true);
MatrixPtr cMat_d2h = Matrix::create(numSamples, nx * ny, false, false);
real* cData = cMat->getData();
real* gold = cBatchTransMat->getData();
// host
for (int sample_id = 0; sample_id < numSamples; ++sample_id)
for (int j = 0; j < ny; j++)
for (int i = 0; i < nx; i++)
cData[sample_id * nx * ny + j * nx + i] = j * nx + i;
// correct result for error checking
for (int sample_id = 0; sample_id < numSamples; ++sample_id)
for (int j = 0; j < ny; j++)
for (int i = 0; i < nx; i++)
gold[sample_id * nx * ny + i * ny + j] =
cData[sample_id * nx * ny + j * nx + i];
// device
gMat->copyFrom(*cMat, HPPL_STREAM_DEFAULT);
batchTranspose(
gMat->getData(), gBatchTransMat->getData(), nx, ny, numSamples);
cMat_d2h->copyFrom(*gBatchTransMat, HPPL_STREAM_DEFAULT);
checkMatrixEqual(cBatchTransMat, cMat_d2h);
}
TEST(Matrix, CopySparseMatrixToGpuSparseMatrix) {
const size_t HEIGHT = 20;
const size_t WIDTH = 10;
const size_t WIDTH_TEST = 15;
MatrixPtr testMatrix(
new CpuSparseMatrix(HEIGHT, WIDTH, HEIGHT * 2, FLOAT_VALUE, SPARSE_CSR));
MatrixPtr testCpuMatrix(new CpuMatrix(HEIGHT, WIDTH));
testCpuMatrix->randomizeUniform();
testMatrix->copyFrom(*testCpuMatrix, HPPL_STREAM_DEFAULT);
MatrixPtr testGpuMatrix = testMatrix->clone(HEIGHT, WIDTH, true);
hl_stream_t gpuStream(HPPL_STREAM_3);
testGpuMatrix->copyFrom(*testMatrix, gpuStream);
hl_stream_synchronize(gpuStream);
MatrixPtr mulCpuMatrix(new CpuMatrix(WIDTH, WIDTH_TEST));
mulCpuMatrix->randomizeUniform();
MatrixPtr mulGpuMatrix(new GpuMatrix(WIDTH, WIDTH_TEST));
mulGpuMatrix->copyFrom(*mulCpuMatrix);
MatrixPtr ret1(new CpuMatrix(HEIGHT, WIDTH_TEST));
MatrixPtr ret2(new GpuMatrix(HEIGHT, WIDTH_TEST));
ret1->zeroMem();
ret2->zeroMem();
ret1->mul(*testMatrix, *mulCpuMatrix, 1.0, 1.0);
ret2->mul(*testGpuMatrix, *mulGpuMatrix, 1.0, 1.0);
checkMatrixEqual(ret1, ret2);
}
const real* getData(const Matrix& matrix) {
if (matrix.useGpu()) {
MatrixPtr cpuMatrix = Matrix::create(
matrix.getHeight(), matrix.getWidth(), matrix.isTransposed(), false);
cpuMatrix->copyFrom(matrix);
return cpuMatrix->getData();
} else {
return matrix.getData();
}
}
void KmaxSeqScoreLayer::forward(PassType passType) {
Layer::forward(passType);
const Argument& input = getInput(0);
const MatrixPtr inputScore = getInputValue(0);
CHECK(input.hasSeq() || input.hasSubseq())
<< "input of " << getName()
<< " must be a sequence or a nested sequence.";
CHECK_EQ(input.value->getWidth(), 1UL)
<< "input of " << getName() << " are scores over a sequence or "
<< "a nested sequence, so its width must be 1.";
if (useGpu_) {
/*
* currently, this Layer only runs in CPU, if the other part of the model is
* runing on GPU, then copy the input to this layer from GPU to CPU.
*/
Matrix::resizeOrCreate(scores_,
inputScore->getHeight(),
1,
false /* trans */,
false /* useGpu */);
scores_->copyFrom(*inputScore);
} else {
scores_ = inputScore;
}
/*
* TODO(caoying)
* In PaddePaddle, currently all matrices are real number types,
* but output of this layer which is some selected indices of the give
* sequence are actually filled with int types so that storing int types
* information in a real number matrix is dangerous, since real numbers will
* be convered to int types.
*/
Matrix::resizeOrCreate(
output_.value,
input.hasSubseq() ? input.getNumSubSequences() : input.getNumSequences(),
beamSize_,
false,
false);
output_.value->one();
output_.value->mulScalar(-1.);
kmaxScorePerSeq(scores_->getData(),
output_.value->getData(),
input.hasSubseq() ? input.subSequenceStartPositions
: input.sequenceStartPositions);
}
void testBilinearFwdBwd(int numSamples,
int imgSizeH,
int imgSizeW,
int channels) {
int inWidth = imgSizeH * imgSizeW * channels;
int outWidth = 2 * imgSizeH * 2 * imgSizeW * channels;
real ratioH = 0.5;
real ratioW = 0.5;
// forward
MatrixPtr input = CpuMatrix::create(numSamples, inWidth, false, false);
MatrixPtr inputGpu = GpuMatrix::create(numSamples, inWidth, false, true);
MatrixPtr target = CpuMatrix::create(numSamples, outWidth, false, false);
MatrixPtr targetGpu = GpuMatrix::create(numSamples, outWidth, false, true);
MatrixPtr targetCheck = CpuMatrix::create(numSamples, outWidth, false, false);
input->randomizeUniform();
inputGpu->copyFrom(*input);
{
// nvprof: GPU Proflier
REGISTER_GPU_PROFILER("testBilinearFwdBwd");
target->bilinearForward(*input,
imgSizeH,
imgSizeW,
2 * imgSizeH,
2 * imgSizeW,
channels,
ratioH,
ratioW);
targetGpu->bilinearForward(*inputGpu,
imgSizeH,
imgSizeW,
2 * imgSizeH,
2 * imgSizeW,
channels,
ratioH,
ratioW);
}
// check
targetCheck->copyFrom(*targetGpu);
MatrixCheckErr(*target, *targetCheck);
// backward
MatrixPtr inputGrad = CpuMatrix::create(numSamples, inWidth, false, false);
MatrixPtr inputGpuGrad = GpuMatrix::create(numSamples, inWidth, false, true);
MatrixPtr targetGrad = CpuMatrix::create(numSamples, outWidth, false, false);
MatrixPtr targetGpuGrad =
GpuMatrix::create(numSamples, outWidth, false, true);
MatrixPtr targetCheckGrad =
CpuMatrix::create(numSamples, inWidth, false, false);
inputGrad->randomizeUniform();
targetGrad->randomizeUniform();
inputGpuGrad->copyFrom(*inputGrad);
targetGpuGrad->copyFrom(*targetGrad);
inputGrad->bilinearBackward(*targetGrad,
2 * imgSizeH,
2 * imgSizeW,
imgSizeH,
imgSizeW,
channels,
ratioH,
ratioW);
inputGpuGrad->bilinearBackward(*targetGpuGrad,
2 * imgSizeH,
2 * imgSizeW,
imgSizeH,
imgSizeW,
channels,
ratioH,
ratioW);
// check
targetCheckGrad->copyFrom(*inputGpuGrad);
MatrixCheckErr(*inputGrad, *targetCheckGrad);
}
TEST(Layer, priorBoxLayerFwd) {
vector<int> minSize;
vector<int> maxSize;
vector<real> aspectRatio;
vector<real> variance;
bool useGpu = false;
minSize.push_back(276);
maxSize.push_back(330);
variance.push_back(0.1);
variance.push_back(0.1);
variance.push_back(0.2);
variance.push_back(0.2);
// CPU case 1.
MatrixPtr result;
real resultData[] = {0.04,
0.04,
0.96,
0.96,
0.1,
0.1,
0.2,
0.2,
0,
0,
1,
1,
0.1,
0.1,
0.2,
0.2};
result = Matrix::create(1, 2 * 8, false, useGpu);
result->setData(resultData);
doOnePriorBoxTest(/* feature_map_width */ 1,
/* feature_map_height */ 1,
/* image_width */ 300,
/* image_height */ 300,
minSize,
maxSize,
aspectRatio,
variance,
useGpu,
result);
// CPU case 2.
variance[1] = 0.2;
variance[3] = 0.1;
maxSize.pop_back();
real resultData2[] = {0, 0, 0.595, 0.595, 0.1, 0.2, 0.2, 0.1,
0.405, 0, 1, 0.595, 0.1, 0.2, 0.2, 0.1,
0, 0.405, 0.595, 1, 0.1, 0.2, 0.2, 0.1,
0.405, 0.405, 1, 1, 0.1, 0.2, 0.2, 0.1};
Matrix::resizeOrCreate(result, 1, 4 * 8, false, useGpu);
result->setData(resultData2);
doOnePriorBoxTest(/* feature_map_width */ 2,
/* feature_map_height */ 2,
/* image_width */ 400,
/* image_height */ 400,
minSize,
maxSize,
aspectRatio,
variance,
useGpu,
result);
// CPU case 3.
aspectRatio.push_back(2);
real resultData3[] = {0.04, 0.04, 0.96, 0.96, 0.1, 0.2,
0.2, 0.1, 0, 0.17473088, 1, 0.825269,
0.1, 0.2, 0.2, 0.1, 0.17473088, 0,
0.825269, 1, 0.1, 0.2, 0.2, 0.1};
Matrix::resizeOrCreate(result, 1, 3 * 8, false, useGpu);
result->setData(resultData3);
doOnePriorBoxTest(/* feature_map_width */ 1,
/* feature_map_height */ 1,
/* image_width */ 300,
/* image_height */ 300,
minSize,
maxSize,
aspectRatio,
variance,
useGpu,
result);
#ifndef PADDLE_ONLY_CPU
// reset the input parameters
variance[1] = 0.1;
variance[3] = 0.2;
maxSize.push_back(330);
aspectRatio.pop_back();
MatrixPtr resultGpu;
useGpu = true;
// GPU case 1.
resultGpu = Matrix::create(1, 2 * 8, false, useGpu);
resultGpu->copyFrom(resultData, 2 * 8);
doOnePriorBoxTest(/* feature_map_width */ 1,
/* feature_map_height */ 1,
/* image_width */ 300,
/* image_height */ 300,
minSize,
maxSize,
aspectRatio,
variance,
useGpu,
resultGpu);
// GPU case 2.
variance[1] = 0.2;
variance[3] = 0.1;
maxSize.pop_back();
Matrix::resizeOrCreate(resultGpu, 1, 4 * 8, false, useGpu);
resultGpu->copyFrom(resultData2, 4 * 8);
doOnePriorBoxTest(/* feature_map_width */ 2,
/* feature_map_height */ 2,
/* image_width */ 400,
/* image_height */ 400,
minSize,
maxSize,
aspectRatio,
variance,
useGpu,
resultGpu);
// GPU case 3.
aspectRatio.push_back(2);
Matrix::resizeOrCreate(resultGpu, 1, 3 * 8, false, useGpu);
resultGpu->copyFrom(resultData3, 3 * 8);
doOnePriorBoxTest(/* feature_map_width */ 1,
/* feature_map_height */ 1,
/* image_width */ 300,
/* image_height */ 300,
minSize,
maxSize,
aspectRatio,
variance,
useGpu,
resultGpu);
#endif
}
