void ExpandLayer::backward(const UpdateCallback& callback) {
if (biases_ && biases_->getWGrad()) {
biases_->getWGrad()->collectBias(*getOutputGrad(), 1);
/* Increasing the number of gradient */
biases_->getParameterPtr()->incUpdate(callback);
}
if (!getInputGrad(0)) return;
MatrixPtr inputGrad = getInputGrad(0);
MatrixPtr outputGrad = getOutputGrad();
auto cpuSeqStartPos = type_ ? getInput(1).subSequenceStartPositions
: getInput(1).sequenceStartPositions;
size_t numSequences = cpuSeqStartPos->getSize() - 1;
const int* starts = cpuSeqStartPos->getData(false);
CHECK_EQ(inputGrad->getWidth(), outputGrad->getWidth());
CHECK_EQ(outputGrad->getHeight(), (size_t)starts[numSequences]);
AsyncGpuBlock asyncGpuBlock;
// sum to get the grad
real scale = 1;
for (size_t sequenceId = 0; sequenceId < numSequences; sequenceId++) {
// TODO(Dangqingqing) optimization for GPU
int sequenceLength = starts[sequenceId + 1] - starts[sequenceId];
if (sequenceLength == 0) {
// empty sequence
continue;
}
MatrixPtr copyData = inputGrad->subMatrix(sequenceId, 1);
copyData->collectBias(
*outputGrad->subMatrix(starts[sequenceId], sequenceLength), scale);
}
}
void SequenceConcatLayer::backward(const UpdateCallback& callback) {
/* activation */
backwardActivation();
if (biases_ && biases_->getWGrad()) {
biases_->getWGrad()->collectBias(*getOutputGrad(), 1);
// Increasing the number of gradient
biases_->getParameterPtr()->incUpdate(callback);
}
MatrixPtr inputGrad1 = getInputGrad(0);
MatrixPtr inputGrad2 = getInputGrad(1);
MatrixPtr outputGrad = getOutputGrad();
auto startPositions1 = getInput(0).sequenceStartPositions->getVector(false);
auto startPositions2 = getInput(1).sequenceStartPositions->getVector(false);
size_t numSequences1 = startPositions1->getSize() - 1;
size_t numSequences2 = startPositions2->getSize() - 1;
CHECK_EQ(numSequences1, numSequences2);
const int* starts1 = startPositions1->getData();
const int* starts2 = startPositions2->getData();
{
AsyncGpuBlock asyncGpuBlock;
REGISTER_TIMER_INFO("SequenceConcatLayerBackward", getName().c_str());
size_t offset = 0;
size_t leftNumIns = 0;
size_t rightNumIns = 0;
for (size_t seqId = 0; seqId < numSequences1; ++seqId) {
leftNumIns = starts1[seqId + 1] - starts1[seqId];
if (inputGrad1) {
inputGrad1->subMatrix(starts1[seqId], leftNumIns)
->add(*(outputGrad->subMatrix(offset, leftNumIns)));
}
offset += leftNumIns;
rightNumIns = starts2[seqId + 1] - starts2[seqId];
if (inputGrad2) {
inputGrad2->subMatrix(starts2[seqId], rightNumIns)
->add(*(outputGrad->subMatrix(offset, rightNumIns)));
}
offset += rightNumIns;
}
}
}
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();
}
}
void DeConv3DLayer::forward(PassType passType) {
Layer::forward(passType);
int batchSize = inputLayers_[0]->getOutputValue()->getHeight();
int outWidth = getSize();
resetOutput(batchSize, outWidth);
const MatrixPtr outMat = getOutputValue();
REGISTER_TIMER_INFO("FwdDeConv3D", getName().c_str());
for (size_t i = 0; i != inputLayers_.size(); ++i) {
const MatrixPtr &inMat = getInputValue(i);
int M = M_[i];
int N = N_[i];
int K = K_[i];
MatrixPtr wMat = weights_[i]->getW();
Matrix::resizeOrCreate(colBuf_, K * groups_[i], N, false, useGpu_);
for (int n = 0; n < batchSize; ++n) {
real *inData = inMat->getData() + n * inMat->getStride();
for (int g = 0; g < groups_[i]; ++g) {
MatrixPtr inMatSub = Matrix::create(inData, M, N, false, useGpu_);
MatrixPtr wMatSub = wMat->subMatrix(g * K, K);
MatrixPtr colBufDataSub = colBuf_->subMatrix(g * K, K);
colBufDataSub->mul(*wMatSub, *inMatSub, 1.0, 0.0);
inData += M * N;
}
colBuf_->col2Vol(outMat->getData() + n * outMat->getStride(),
numFilters_,
imgSizeD_[i],
imgSizeH_[i],
imgSizeW_[i],
filterSizeZ_[i],
filterSizeY_[i],
filterSize_[i],
strideZ_[i],
strideY_[i],
stride_[i],
paddingZ_[i],
paddingY_[i],
padding_[i],
1.0,
1.0);
}
}
if (nullptr != this->biasParameter_) {
this->addBias();
}
forwardActivation();
}
void MultiplexLayer::backward(const UpdateCallback& callback) {
/* Do derivation */ {
REGISTER_TIMER_INFO("BpAvtTimer", getName().c_str());
backwardActivation();
}
MatrixPtr outG = getOutputGrad();
{
REGISTER_TIMER_INFO("BwLMultiplexTimer", getName().c_str());
AsyncGpuBlock block;
for (const CopyInfo& info : copySchedule_) {
if (getInputGrad(info.copyIdx + 1)) {
getInputGrad(info.copyIdx + 1)
->subMatrix(info.startIdx, info.length, tmpDest_)
->add(*outG->subMatrix(info.startIdx, info.length, tmpSrc_));
}
}
}
}
void SequenceConcatLayer::forward(PassType passType) {
Layer::forward(passType);
size_t dim = getSize();
const Argument& input1 = getInput(0);
size_t numSequences1 = input1.getNumSequences();
auto startPositions1 = input1.sequenceStartPositions->getVector(false);
const Argument& input2 = getInput(1);
size_t numSequences2 = input2.getNumSequences();
auto startPositions2 = input2.sequenceStartPositions->getVector(false);
CHECK_EQ(dim, input1.value->getWidth());
CHECK_EQ(startPositions1->getData()[numSequences1], input1.getBatchSize());
CHECK_EQ(numSequences1, startPositions1->getSize() - 1);
CHECK_EQ(dim, input2.value->getWidth());
CHECK_EQ(startPositions2->getData()[numSequences2], input2.getBatchSize());
CHECK_EQ(numSequences2, startPositions2->getSize() - 1);
CHECK_EQ(numSequences1, numSequences2);
MatrixPtr inputValue1 = getInputValue(0);
MatrixPtr inputValue2 = getInputValue(1);
// reset output
reserveOutput(inputValue1->getHeight() + inputValue2->getHeight(), dim);
MatrixPtr outputValue = getOutputValue();
const int* starts1 = startPositions1->getData();
const int* starts2 = startPositions2->getData();
{
AsyncGpuBlock asyncGpuBlock;
REGISTER_TIMER_INFO("SequenceConcatLayerForward", getName().c_str());
size_t offset = 0;
size_t leftNumIns = 0;
size_t rightNumIns = 0;
for (size_t seqId = 0; seqId < numSequences1; ++seqId) {
leftNumIns = starts1[seqId + 1] - starts1[seqId];
outputValue->subMatrix(offset, leftNumIns)
->assign(*(inputValue1->subMatrix(starts1[seqId], leftNumIns)));
offset += leftNumIns;
rightNumIns = starts2[seqId + 1] - starts2[seqId];
outputValue->subMatrix(offset, rightNumIns)
->assign(*(inputValue2->subMatrix(starts2[seqId], rightNumIns)));
offset += rightNumIns;
}
// modify the sequenceStartPositions
ICpuGpuVector::resizeOrCreate(
output_.sequenceStartPositions, numSequences1 + 1, false);
int* tgtBuf = output_.sequenceStartPositions->getMutableData(false);
for (size_t seqId = 0; seqId < numSequences1 + 1; ++seqId) {
tgtBuf[seqId] = starts1[seqId] + starts2[seqId];
}
}
if (biases_.get() != NULL) {
MatrixPtr outV = getOutputValue();
outV->addBias(*(biases_->getW()), 1);
}
/* activation */
forwardActivation();
}
