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 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 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 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 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 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);
}
}