void CosSimVecMatLayer::forward(PassType passType) {
Layer::forward(passType);
CHECK_EQ(forward_.size(), 1UL) << "Only one forward function needed";
MatrixPtr inV0 = getInputValue(0);
MatrixPtr inV1 = getInputValue(1);
size_t batchSize = inV0->getHeight();
size_t numKeys = getSize();
CHECK_EQ(batchSize, inV1->getHeight());
{
REGISTER_TIMER_INFO("FwResetTimer", getName().c_str());
reserveOutput(batchSize, numKeys);
}
MatrixPtr outV = getOutputValue();
CHECK(outV && inV0 && inV1);
REGISTER_TIMER_INFO("FwCosVMTimer", getName().c_str());
for (size_t i = 0; i < batchSize; i++) {
tmpRow0->setData(inV0->rowBuf(i));
tmpMtx0->setData(inV1->rowBuf(i));
tmpRow2->setData(outV->rowBuf(i));
BufferArgs inputs;
BufferArgs outputs;
inputs.addArg(*tmpMtx0);
inputs.addArg(*tmpRow0);
outputs.addArg(*tmpRow2, ASSIGN_TO);
forward_[0]->calc(inputs, outputs);
}
}
TEST(Layer, convTransLayerFwd2) {
MatrixPtr result;
result = Matrix::create(1, 5 * 5, false, false);
result->zeroMem();
result->add(1.0);
doOneConvtTest(/* imgSize */ 5,
/* output_x */ 1,
/* stride */ 1,
/* padding */ 0,
/* filter_size */ 5,
result);
float resultData[] = {1, 2, 2, 2, 1,
2, 4, 4, 4, 2,
2, 4, 4, 4, 2,
2, 4, 4, 4, 2,
1, 2, 2, 2, 1};
result->setData(resultData);
doOneConvtTest(/* imgSize */ 5,
/* output_x */ 2,
/* stride */ 1,
/* padding */ 0,
/* filter_size */ 4,
result);
float resultData2[] = {1, 2, 2, 2, 1,
2, 4, 4, 4, 2,
2, 4, 4, 4, 2,
2, 4, 4, 4, 2,
1, 2, 2, 2, 1};
result->setData(resultData2);
doOneConvtTest(/* imgSize */ 5,
/* output_x */ 2,
/* stride */ 2,
/* padding */ 1,
/* filter_size */ 5,
result);
float resultData3[] = {1, 1, 2, 1, 1,
1, 1, 2, 1, 1,
2, 2, 4, 2, 2,
1, 1, 2, 1, 1,
1, 1, 2, 1, 1};
result->setData(resultData3);
doOneConvtTest(/* imgSize */ 5,
/* output_x */ 2,
/* stride */ 2,
/* padding */ 0,
/* filter_size */ 3,
result);}
void CosSimVecMatLayer::backward(const UpdateCallback& callback) {
CHECK_EQ(backward_.size(), 1UL) << "Only one forward function needed";
MatrixPtr inV0 = getInputValue(0);
MatrixPtr inV1 = getInputValue(1);
MatrixPtr inG0 = getInputGrad(0);
MatrixPtr inG1 = getInputGrad(1);
MatrixPtr outV = getOutputValue();
MatrixPtr outG = getOutputGrad();
size_t batchSize = inV0->getHeight();
CHECK(inV0 && inV1 && inG0 && inG1 && outV && outG);
REGISTER_TIMER_INFO("BwCosVMTimer", getName().c_str());
for (size_t i = 0; i < batchSize; i++) {
tmpRow0->setData(inV0->rowBuf(i));
tmpRow1->setData(inG0->rowBuf(i));
tmpMtx0->setData(inV1->rowBuf(i));
tmpMtx1->setData(inG1->rowBuf(i));
tmpRow2->setData(outV->rowBuf(i));
tmpRow3->setData(outG->rowBuf(i));
BufferArgs inputs;
BufferArgs outputs;
inputs.addArg(*tmpRow3);
inputs.addArg(*tmpRow2);
inputs.addArg(*tmpMtx0);
inputs.addArg(*tmpRow0);
outputs.addArg(*tmpMtx1, ADD_TO);
outputs.addArg(*tmpRow1, ADD_TO);
backward_[0]->calc(inputs, outputs);
}
}
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
}
