void MultiplexLayer::forward(PassType passType) {
Layer::forward(passType);
IVectorPtr copyIds = getInput(0).ids;
MatrixPtr inV1 = getInputValue(1);
CHECK_EQ(copyIds->getSize(), inV1->getHeight());
for (size_t i = 2; i < inputLayers_.size(); i++) {
CHECK_EQ(inV1->getHeight(), getInputValue(i)->getHeight());
CHECK_EQ(inV1->getWidth(), getInputValue(i)->getWidth());
}
calculateCopySchedule(copyIds, inputLayers_.size() - 1);
{
REGISTER_TIMER_INFO("FwResetTimer", getName().c_str());
reserveOutput(inV1->getHeight(), inV1->getWidth());
}
MatrixPtr outV = getOutputValue();
{
REGISTER_TIMER_INFO("FwLMultplexingTimer", getName().c_str());
AsyncGpuBlock block;
for (const CopyInfo& info : copySchedule_) {
outV->subMatrix(info.startIdx, info.length, tmpDest_)
->copyFrom(*getInputValue(info.copyIdx + 1)
->subMatrix(info.startIdx, info.length, tmpSrc_));
}
}
/* activation */ {
REGISTER_TIMER_INFO("FwAtvTimer", getName().c_str());
forwardActivation();
}
}
TEST(Argument, poolSequenceWithStride) {
Argument input, output;
ICpuGpuVector::resizeOrCreate(input.sequenceStartPositions, 5, false);
int* inStart = input.sequenceStartPositions->getMutableData(false);
inStart[0] = 0;
inStart[1] = 9;
inStart[2] = 14;
inStart[3] = 17;
inStart[4] = 30;
int strideResult[] = {0, 5, 9, 14, 17, 22, 27, 30};
int strideResultReversed[] = {0, 4, 9, 14, 17, 20, 25, 30};
for (auto reversed : {false, true}) {
IVectorPtr stridePositions;
output.poolSequenceWithStride(
input, 5 /* stride */, &stridePositions, reversed);
const int* outStart = output.sequenceStartPositions->getData(false);
CHECK_EQ(outStart[0], 0);
CHECK_EQ(outStart[1], 2);
CHECK_EQ(outStart[2], 3);
CHECK_EQ(outStart[3], 4);
CHECK_EQ(outStart[4], 7);
CHECK_EQ(stridePositions->getSize(), 8);
auto result = reversed ? strideResultReversed : strideResult;
for (int i = 0; i < 8; i++) {
CHECK_EQ(stridePositions->getData()[i], result[i]);
}
}
}
void zadatakprvi() {
IVectorPtr a = Vector::parseSimple("2 3 -4");
IVectorPtr b = Vector::parseSimple("-1 4 -3");
IVectorPtr c = Vector::parseSimple("2 2 4");
IMatrixPtr d = Matrix::parseSimple("1 2 3 | 2 1 3 | 4 5 1");
IMatrixPtr e = Matrix::parseSimple("-1 2 -3 | 5 -2 7 | -4 -1 3");
IMatrixPtr f = Matrix::parseSimple("-24 18 5 | 20 -15 -4 | -5 4 1");
IMatrixPtr g = Matrix::parseSimple("1 2 3 | 0 1 4 | 5 6 0");
IVectorPtr v1 = a->nAdd(b);
cout << "v1: " << endl << v1->toString() << endl;
double s = v1->scalarProduct(b);
cout << "s: " << endl << s << endl << endl;
IVectorPtr v2 = v1->nVectorProduct(a);
cout << "v2: " << endl << v2->toString() << endl;
IVectorPtr v3 = v2->copy()->normalize();
cout << "v3: " << endl << v3->toString() << endl;
IVectorPtr v4 = v2->copy()->nScalarMultiply(-1);
cout << "v4: " << endl << v4->toString() << endl;
IMatrixPtr m1 = d->nAdd(e);
cout << "m1: " << endl << m1->toString() << endl;
IMatrixPtr m2 = d->nMultiply(e->nTranspose(true));
cout << "m2: " << endl << m2->toString() << endl;
IMatrixPtr m3 = f->nInvert()->nMultiply(g->nInvert());
cout << "m3: " << endl << m3->toString() << endl;
}
void SparsePrefetchRowCpuMatrix::addRows(IVectorPtr ids) {
std::vector<unsigned int>& localIndices = indexDictHandle_->localIndices;
size_t numSamples = ids->getSize();
int* index = ids->getData();
for (size_t i = 0; i < numSamples; ++i) {
if (index[i] == -1) continue;
unsigned int id = (unsigned int)index[i];
CHECK_LT(id, this->getHeight())
<< "id:" << id << "Height:" << this->getHeight()
<< "sparse id value exceeds the max input dimension, "
<< "it could be caused invalid input data samples";
localIndices.push_back(id);
}
}
void prepareSamples() {
CHECK(!useGpu_) << "GPU is not supported";
int batchSize = getInput(*labelLayer_).getBatchSize();
IVectorPtr label = getInput(*labelLayer_).ids;
CpuSparseMatrixPtr multiLabel = std::dynamic_pointer_cast<CpuSparseMatrix>(
getInput(*labelLayer_).value);
CHECK(label || multiLabel)
<< "The label layer must have ids or NonValueSparseMatrix value";
auto& randEngine = ThreadLocalRandomEngine::get();
samples_.clear();
samples_.reserve(batchSize * (1 + config_.num_neg_samples()));
real* weight =
weightLayer_ ? getInputValue(*weightLayer_)->getData() : nullptr;
for (int i = 0; i < batchSize; ++i) {
real w = weight ? weight[i] : 1;
if (label) {
int* ids = label->getData();
samples_.push_back({i, ids[i], true, w});
} else {
const int* cols = multiLabel->getRowCols(i);
int n = multiLabel->getColNum(i);
for (int j = 0; j < n; ++j) {
samples_.push_back({i, cols[j], true, w});
}
}
for (int j = 0; j < config_.num_neg_samples(); ++j) {
int id = sampler_ ? sampler_->gen(randEngine) : rand_(randEngine);
samples_.push_back({i, id, false, w});
}
}
prepared_ = true;
}
void generateMDimSequenceData(const IVectorPtr& sequenceStartPositions,
IVectorPtr& cpuSequenceDims) {
/* generate sequences with 2 dims */
int numSeqs = sequenceStartPositions->getSize() - 1;
int numDims = 2;
cpuSequenceDims = IVector::create(numSeqs * numDims, /* useGpu= */ false);
int* bufStarts = sequenceStartPositions->getData();
int* bufDims = cpuSequenceDims->getData();
for (int i = 0; i < numSeqs; i++) {
int len = bufStarts[i + 1] - bufStarts[i];
/* get width and height randomly */
std::vector<int> dimVec;
for (int j = 0; j < len; j++) {
if (len % (j + 1) == 0) {
dimVec.push_back(1);
}
}
int idx = rand() % dimVec.size(); // NOLINT use rand_r
bufDims[i * numDims] = dimVec[idx];
bufDims[i * numDims + 1] = len / dimVec[idx];
}
}
void MultiplexLayer::calculateCopySchedule(const IVectorPtr& copyIds,
size_t numIns) {
copySchedule_.clear();
CopyInfo prevCopyInfo(0, 0, -1);
for (size_t i = 0; i < copyIds->getSize(); i++) {
int copyId = copyIds->getElement(i);
CHECK_GE(copyId, 0);
CHECK_LT(copyId, int(numIns));
// copy same input layer with prevous and will copy consecutive.
if (copyId == prevCopyInfo.copyIdx) {
++prevCopyInfo.length;
} else {
if (prevCopyInfo.copyIdx != -1) {
copySchedule_.emplace_back(prevCopyInfo);
}
prevCopyInfo.startIdx = i;
prevCopyInfo.length = 1;
prevCopyInfo.copyIdx = copyId;
}
}
if (prevCopyInfo.copyIdx != -1) {
copySchedule_.emplace_back(prevCopyInfo);
}
}
void prefetch() override {
prepareSamples();
IVector::resizeOrCreate(labelIds_, samples_.size(), useGpu_);
int* ids = labelIds_->getData();
for (size_t i = 0; i < samples_.size(); ++i) {
ids[i] = samples_[i].labelId;
}
for (int i = 0; i < numInputs_; ++i) {
auto sparseParam =
dynamic_cast<SparsePrefetchRowCpuMatrix*>(weights_[i]->getW().get());
if (sparseParam) {
sparseParam->addRows(labelIds_);
}
}
}
void checkVectorEqual(const IVectorPtr& a, const IVectorPtr& b) {
EXPECT_EQ(a->getSize(), b->getSize());
for (size_t r = 0; r < a->getSize(); ++r) {
EXPECT_FLOAT_EQ(a->get(r), b->get(r));
}
}
void prepareData(DataBatch* batch, const int* numPerSlotType, bool iid,
bool useGpu) {
batch->clear();
int64_t size = uniformRandom(100) + 10;
batch->setSize(size);
ICpuGpuVectorPtr sequenceStartPositions;
ICpuGpuVectorPtr subSequenceStartPositions;
if (!iid) {
int numSeqs = uniformRandom(10) + 1;
sequenceStartPositions =
ICpuGpuVector::create(numSeqs + 1, /* useGpu= */ false);
int* buf = sequenceStartPositions->getMutableData(false);
subSequenceStartPositions =
ICpuGpuVector::create(numSeqs + 1, /* useGpu= */ false);
int* subBuf = subSequenceStartPositions->getMutableData(false);
int64_t pos = 0;
int maxLen = 2 * size / numSeqs;
for (int i = 0; i < numSeqs; ++i) {
int len =
uniformRandom(min<int64_t>(maxLen, size - pos - numSeqs + i)) + 1;
buf[i] = pos;
subBuf[i] = pos;
pos += len;
VLOG(1) << " len=" << len;
}
buf[numSeqs] = size;
subBuf[numSeqs] = size;
}
vector<Argument>& arguments = batch->getStreams();
for (int i = 0; i < numPerSlotType[SlotDef::VECTOR_DENSE]; ++i) {
int64_t dim = rand() % 10 + 4; // NOLINT rand_r
MatrixPtr mat = Matrix::create(size, dim, /* trans= */ false, false);
mat->randomizeUniform();
Argument arg;
arg.value = mat;
arg.sequenceStartPositions = sequenceStartPositions;
arguments.push_back(arg);
}
for (int i = 0; i < numPerSlotType[SlotDef::VECTOR_SPARSE_NON_VALUE]; ++i) {
MatrixPtr mat =
makeRandomSparseMatrix(size, kSpraseMatrixDim, false, useGpu);
Argument arg;
arg.value = mat;
arg.sequenceStartPositions = sequenceStartPositions;
arg.subSequenceStartPositions = subSequenceStartPositions;
arguments.push_back(arg);
}
for (int i = 0; i < numPerSlotType[SlotDef::VECTOR_SPARSE_VALUE]; ++i) {
MatrixPtr mat =
makeRandomSparseMatrix(size, kSpraseMatrixDim, true, useGpu);
Argument arg;
arg.value = mat;
arg.sequenceStartPositions = sequenceStartPositions;
arguments.push_back(arg);
}
for (int i = 0; i < numPerSlotType[SlotDef::STRING]; ++i) {
int64_t dim = rand() % 10 + 4; // NOLINT rand_r
SVectorPtr vec = std::make_shared<std::vector<std::string>>();
for (int j = 0; j < size; ++j) {
vec->push_back(randStr(dim));
}
Argument arg;
arg.strs = vec;
arg.sequenceStartPositions = sequenceStartPositions;
arguments.push_back(arg);
}
for (int i = 0; i < numPerSlotType[SlotDef::INDEX]; ++i) {
int64_t dim = rand() % 10 + 4; // NOLINT rand_r
IVectorPtr vec = IVector::create(size, /* useGpu= */ false);
int* buf = vec->getData();
for (int j = 0; j < size; ++j) {
buf[j] = uniformRandom(dim);
}
Argument arg;
arg.ids = vec;
arg.sequenceStartPositions = sequenceStartPositions;
arguments.push_back(arg);
}
}
