void DummyNodeTest<ElemType>::SetMinibatch(size_t minibatchSize, SmallVector<size_t> sampleDimensions,
std::vector<ElemType>& data)
{
// Set given shape and allocate matrices.
TensorShape shape(sampleDimensions);
if (shape.GetNumElements() * minibatchSize != data.size())
LogicError("Data size is incompatible with specified dimensions.");
MBLayoutPtr mbLayout = make_shared<MBLayout>();
mbLayout->InitAsFrameMode(minibatchSize);
this->LinkToMBLayout(mbLayout);
this->SetDims(shape, true);
this->CreateValueMatrixIfNull();
this->Value().Resize(shape.GetNumElements(), minibatchSize);
this->CreateGradientMatrixIfNull();
this->Gradient().Resize(shape.GetNumElements(), minibatchSize);
this->Value().SetValue(minibatchSize, shape.GetNumElements(), this->m_deviceId, data.data());
}
MBLayoutPtr SequencePacker::CreateMBLayout(const StreamBatch& batch)
{
vector<MBLayout::SequenceInfo> infos;
for (size_t index = 0; index < batch.size(); ++index)
{
MBLayout::SequenceInfo info;
info.seqId = index;
info.tBegin = 0;
info.tEnd = batch[index]->m_numberOfSamples;
infos.push_back(info);
}
vector<pair<size_t, size_t>> placement;
vector<size_t> rowAllocations;
// Creating the minibatch layout.
MBLayoutPtr pMBLayout = make_shared<MBLayout>();
pMBLayout->InitAsPackedSequences(infos, placement, rowAllocations);
return pMBLayout;
}
void ComputationNode<ElemType>::WriteMinibatchWithFormatting(FILE* f, const FrameRange& fr,
size_t onlyUpToRow, size_t onlyUpToT, bool transpose, bool isCategoryLabel, bool isSparse,
const vector<string>& labelMapping, const string& sequenceSeparator,
const string& sequencePrologue, const string& sequenceEpilogue,
const string& elementSeparator, const string& sampleSeparator,
string valueFormatString,
bool outputGradient) const
{
// get minibatch matrix -> matData, matRows, matStride
const Matrix<ElemType>& outputValues = outputGradient ? Gradient() : Value();
let matRows = outputValues.GetNumRows();
let matStride = matRows; // how to get from one column to the next
unique_ptr<ElemType[]> matDataPtr(outputValues.CopyToArray());
ElemType* matData = matDataPtr.get();
let sampleLayout = GetSampleLayout(); // this is currently only used for sparse; dense tensors are linearized
// process all sequences one by one
MBLayoutPtr pMBLayout = GetMBLayout();
if (!pMBLayout) // no MBLayout: We are printing aggregates (or LearnableParameters?)
{
pMBLayout = make_shared<MBLayout>();
pMBLayout->Init(1, outputValues.GetNumCols()); // treat this as if we have one single sequence consisting of the columns
pMBLayout->AddSequence(0, 0, 0, outputValues.GetNumCols());
}
let& sequences = pMBLayout->GetAllSequences();
let width = pMBLayout->GetNumTimeSteps();
TensorShape tensorShape = GetSampleLayout();
stringstream str;
let dims = tensorShape.GetDims();
for (auto dim : dims)
str << dim << ' ';
let shape = str.str(); // BUGBUG: change to string(tensorShape) to make sure we always use the same format
bool sequencePrologueHasShape = sequencePrologue.find("%x") != sequencePrologue.npos;
bool sampleSeparatorHasShape = sampleSeparator.find("%x") != sampleSeparator.npos;
bool sequencePrologueHasSeqId = sequencePrologue.find("%d") != sequencePrologue.npos;
bool sampleSeparatorHasSeqId = sampleSeparator.find("%d") != sampleSeparator.npos;
for (size_t s = 0; s < sequences.size(); s++)
{
const auto& seqInfo = sequences[s];
if (seqInfo.seqId == GAP_SEQUENCE_ID) // nothing in gaps to print
continue;
let tBegin = seqInfo.tBegin >= 0 ? seqInfo.tBegin : 0;
let tEnd = seqInfo.tEnd <= width ? seqInfo.tEnd : width;
// [tBegin,tEnd) is where the sequence resides.
// fr is also referencing where a sequence resides.
// narrow to FrameRange if needed
auto t0 = fr.IsAllFrames() ? tBegin : fr.m_timeOffset + (ptrdiff_t)fr.timeIdxInSeq;
auto t1 = fr.IsAllFrames() ? tEnd : fr.m_timeOffset + (ptrdiff_t)fr.timeIdxInSeq + (ptrdiff_t)fr.m_timeRange;
if (t0 < tBegin)
t0 = tBegin;
if (t1 > tEnd)
t1 = tEnd;
// [t0,t1) is the range we want to print
if (t0 > (ptrdiff_t)t1)
continue; // skip this sequence
// get sequence matrix -> seqData, seqRows, seqCols, seqStride
let seqData = matData + pMBLayout->GetColumnIndex(seqInfo, t0 - tBegin) * matStride;
auto seqRows = matRows;
let seqCols = t1 - t0;
let seqStride = pMBLayout->GetNumParallelSequences() * matStride;
auto seqProl = sequencePrologue;
auto sampleSep = sampleSeparator;
if (sequencePrologueHasShape || sampleSeparatorHasShape)
{
auto sh = msra::strfun::_strprintf<char>("%s%ld", shape.c_str(), (unsigned long long)seqInfo.GetNumTimeSteps());
if (sequencePrologueHasShape)
seqProl = msra::strfun::ReplaceAll<std::string>(seqProl, "%x", sh);
if (sampleSeparatorHasShape)
sampleSep = msra::strfun::ReplaceAll<std::string>(sampleSep, "%x", sh);
}
if (sequencePrologueHasSeqId || sampleSeparatorHasSeqId)
{
auto sh = msra::strfun::_strprintf<char>("%ld", (unsigned long long)seqInfo.seqId);
if (sequencePrologueHasSeqId)
seqProl = msra::strfun::ReplaceAll<std::string>(seqProl, "%d", sh);
if (sampleSeparatorHasSeqId)
sampleSep = msra::strfun::ReplaceAll<std::string>(sampleSep, "%d", sh);
}
if (s > 0)
fprintfOrDie(f, "%s", sequenceSeparator.c_str());
fprintfOrDie(f, "%s", seqProl.c_str());
// output it according to our format specification
auto formatChar = valueFormatString.back();
if (isCategoryLabel) // if is category then find the max value and output its index (possibly mapped to a string)
{
if (formatChar == 's') // verify label dimension
{
if (outputValues.GetNumRows() != labelMapping.size() &&
sampleLayout[0] != labelMapping.size()) // if we match the first dim then use that
{
static size_t warnings = 0;
if (warnings++ < 5)
fprintf(stderr, "write: Row dimension %d does not match number of entries %d in labelMappingFile, not using mapping\n", (int)seqRows, (int)labelMapping.size());
valueFormatString.back() = 'u'; // this is a fallback
formatChar = valueFormatString.back();
}
}
// update the matrix in-place from one-hot (or max) to index
// find the max in each column
for (size_t j = 0; j < seqCols; j++) // loop over all time steps of the sequence
{
double maxLoc = -1;
double maxVal = 0;
for (size_t i = 0; i < seqRows; i++) // loop over rows
{
let val = seqData[i + j * seqStride];
if (maxLoc < 0 || val >= maxVal)
{
maxLoc = (double)i;
maxVal = val;
}
}
seqData[0 + j * seqStride] = (ElemType)maxLoc; // overwrite first element in-place
}
seqRows = 1; // ignore remaining dimensions
}
// function to print a value
auto print = [&](double dval)
{
if (formatChar == 'f') // print as real number
{
if (dval == 0) dval = fabs(dval); // clear the sign of a negative 0, which are produced inconsistently between CPU and GPU
fprintfOrDie(f, valueFormatString.c_str(), dval);
}
else if (formatChar == 'u') // print category as integer index
{
fprintfOrDie(f, valueFormatString.c_str(), (unsigned int)dval);
}
else if (formatChar == 's') // print category as a label string
{
size_t uval = (size_t)dval;
if (!labelMapping.empty())
uval %= labelMapping.size();
assert(uval < labelMapping.size());
const char * sval = labelMapping[uval].c_str();
fprintfOrDie(f, valueFormatString.c_str(), sval);
}
};
// bounds for printing
let iend = transpose ? seqRows : seqCols; // true dimension of the data to print
let jend = transpose ? seqCols : seqRows;
let istop = transpose ? onlyUpToRow : onlyUpToT; // we stop at these dimensions (for debugging, one often needs only the first few values of those huge matrices)
let jstop = transpose ? onlyUpToT : onlyUpToRow;
let istride = transpose ? 1 : seqStride;
let jstride = transpose ? seqStride : 1;
if (isSparse)
{
// sparse linearizes the entire matrix into a single vector, and prints that one with coordinates
// TODO: This can be done more nicely. We should keep the block structure.
size_t numPrinted = 0;
for (size_t i = 0; i < iend; i++) // loop over elements --we just flatten them all out
{
for (size_t j = 0; j < jend; j++) // loop over rows
{
double dval = seqData[i * istride + j * jstride];
if (dval == 0) // only print non-0 values
continue;
if (numPrinted++ > 0)
fprintfOrDie(f, "%s", transpose ? sampleSeparator.c_str() : elementSeparator.c_str());
if (dval != 1.0 || formatChar != 'f') // hack: we assume that we are either one-hot or never precisely hitting 1.0
print(dval);
size_t row = transpose ? i : j;
size_t col = transpose ? j : i;
for (size_t k = 0; k < sampleLayout.size(); k++)
{
fprintfOrDie(f, "%c%d", k == 0 ? '[' : ',', row % sampleLayout[k]);
if (sampleLayout[k] == labelMapping.size()) // annotate index with label if dimensions match (which may misfire once in a while)
fprintfOrDie(f, "=%s", labelMapping[row % sampleLayout[k]].c_str());
row /= sampleLayout[k];
}
if (seqInfo.GetNumTimeSteps() > 1)
fprintfOrDie(f, ";%d", col);
fprintfOrDie(f, "]");
}
}
}
else
{
for (size_t j = 0; j < jend; j++) // loop over output rows --BUGBUG: row index is 'i'!! Rename these!!
{
if (j > 0)
fprintfOrDie(f, "%s", sampleSep.c_str());
if (j == jstop && jstop < jend - 1) // if jstop == jend-1 we may as well just print the value instead of '...'
{
fprintfOrDie(f, "...+%d", (int)(jend - jstop)); // 'nuff said
break;
}
// inject sample tensor index if we are printing row-wise and it's a tensor
if (!transpose && sampleLayout.size() > 1 && !isCategoryLabel) // each row is a different sample dimension
{
for (size_t k = 0; k < sampleLayout.size(); k++)
fprintfOrDie(f, "%c%d", k == 0 ? '[' : ',', (int)((j / sampleLayout.GetStrides()[k])) % sampleLayout[k]);
fprintfOrDie(f, "]\t");
}
// print a row of values
for (size_t i = 0; i < iend; i++) // loop over elements
{
if (i > 0)
fprintfOrDie(f, "%s", elementSeparator.c_str());
if (i == istop && istop < iend - 1)
{
fprintfOrDie(f, "...+%d", (int)(iend - istop));
break;
}
double dval = seqData[i * istride + j * jstride];
print(dval);
}
}
}
fprintfOrDie(f, "%s", sequenceEpilogue.c_str());
} // end loop over sequences
fflushOrDie(f);
}
StreamMinibatchPtr SequencePacker::PackStreamMinibatch(const std::vector<SequenceDataPtr>& sequences, size_t streamId)
{
// Create sequence info for each sequences that we have got from the transformer.
std::vector<MBLayout::SequenceInfo> inputSequences;
for (size_t index = 0; index < sequences.size(); ++index)
{
MBLayout::SequenceInfo info;
// In each minibatch sequence ids should be unique.
// They have to match between different input streams in the same minibatch.
// We are using sequence index in the set of received sequences.
// TODO: should we use m_key as sequence id and pass it with data?
info.seqId = index;
info.tBegin = 0;
info.tEnd = sequences[index]->m_numberOfSamples;
inputSequences.push_back(info);
}
std::vector<std::pair<size_t, size_t>> placement;
std::vector<size_t> rowAllocations;
// Creating the minibatch layout.
MBLayoutPtr layout = std::make_shared<MBLayout>();
layout->InitAsPackedSequences(inputSequences, placement, rowAllocations);
// Allocating necessary data buffer for the stream.
size_t sampleSize = GetSampleSize(m_inputStreams[streamId]);
size_t totalNumberOfSamplesInBytes = layout->GetNumCols() * sampleSize;
if (m_streamBufferSizes[streamId] < totalNumberOfSamplesInBytes)
{
m_streamBuffers[streamId] = AllocateBuffer(layout->GetNumCols(), sampleSize);
m_streamBufferSizes[streamId] = totalNumberOfSamplesInBytes;
}
// Packing the actual data.
StorageType storageType = m_inputStreams[streamId]->m_storageType;
size_t elementSize = GetSizeByType(m_inputStreams[streamId]->m_elementType);
const auto& packedSequences = layout->GetAllSequences();
char* streamBuffer = m_streamBuffers[streamId].get();
for (const auto& sequence : packedSequences)
{
if (sequence.seqId == GAP_SEQUENCE_ID)
continue;
const auto& data = sequences[sequence.seqId];
// Packing the sequence
for (size_t sampleIndex = 0; sampleIndex < sequence.GetNumTimeSteps(); ++sampleIndex)
{
char* destination = streamBuffer + layout->GetColumnIndex(sequence, sampleIndex) * sampleSize;
if (storageType == StorageType::dense)
{
PackDenseSample(destination, data, sampleIndex, elementSize, sampleSize);
}
else // sparse
{
assert(storageType == StorageType::sparse_csc);
PackSparseSample(destination, data, sampleIndex, elementSize, sampleSize);
}
}
}
// Ok, minibatch is ready, give it out.
StreamMinibatchPtr result = std::make_shared<StreamMinibatch>();
result->m_data = m_streamBuffers[streamId].get();
result->m_layout = layout;
return result;
}
