void LearnableParameter<ElemType>::InitRandom(const bool uniformInit,
const unsigned long randomSeed,
const ElemType initValueScale,
bool initOnCPUOnly)
{
// fprintf(stderr, "%d x %d: %d %ls\n", (int)GetNumRows(), (int)GetNumCols(), (int)randomSeed, NodeName().c_str());
// the random seed offset is set via the "randomSeedOffset" parameter in config
if (initOnCPUOnly)
Value().TransferToDeviceIfNotThere(CPUDEVICE, true);
#if 1 // this more complex version is needed to repro test cases generated with an older version
auto& value = GetSampleLayout().GetRank() > 2 ? Value() : ValueAsMatrix();
#else
auto& value = Value();
#endif
if (uniformInit)
{
// TODO: move these hidden extra factors out from here and into NDL, and make them visible in BS
ElemType randRange = 0.05f * initValueScale;
value.SetUniformRandomValue(-randRange, randRange, randomSeed);
}
else
{
size_t inputSize = value.GetNumCols();
ElemType randInitstd = 0.2f * initValueScale / sqrt(ElemType(inputSize));
value.SetGaussianRandomValue(0, randInitstd, randomSeed);
}
if (initOnCPUOnly)
Value().TransferToDeviceIfNotThere(m_deviceId, true);
}
void LearnableParameter<ElemType>::InitFromArray(const std::vector<ElemType>& array, size_t numRows, size_t numCols)
{
// infer tensor dimensions from input file if not set
// Note: The mapping of dimensions of the input matrix to tensor dimensions are somewhat confusing.
// The file contains a 2D matrix (one row per text line) that is saved into our column-major representation.
// That representation is then reshaped into a column-major tensor.
if (GetSampleLayout().GetNumElements() == 0) // at least one dimension is 0
{
auto dims = GetSampleLayout().GetDims();
// infer rank
if (dims.size() == 0)
dims.push_back(0);
if (dims.size() == 1 && numCols != 1)
dims.push_back(0);
// infer #rows
if (dims[0] == 0) // infer row dimension as input matrix row dimension
dims[0] = numRows; // (if already set, then mismatch will be caught in VerifyDataSize() below)
// infer #cols: product of all dimensions but the first must match matrix #cols; if there is a single 0 position, we infer it
size_t zeroDim = 0; // 0 means not found
size_t prod = 1;
for (size_t k = 1; k < dims.size(); k++)
{
auto dim = dims[k];
if (dim != 0)
prod *= dim;
else if (zeroDim == 0)
zeroDim = k;
else
InvalidArgument("%ls %ls operation's specified shape [%s] cannot be inferred: Too many unknown dimensions.", NodeName().c_str(), OperationName().c_str(), string(GetSampleLayout()).c_str());
}
if (zeroDim != 0) // we found a zero
{
dims[zeroDim] = numCols / prod;
if (prod * dims[zeroDim] != numCols)
InvalidArgument("%ls %ls operation's specified shape [%s] cannot be inferred: Tensor shape cannot hold a [%d x %d] matrix.", NodeName().c_str(), OperationName().c_str(), string(GetSampleLayout()).c_str(), (int)numRows, (int)numCols);
}
SetDims(TensorShape(dims), false);
}
// BUGBUG: We should allow to read an arbitrary tensor from a single-column file.
// Currently, this would cause a matrix/tensor dimension mismatch. --TODO: Is this comment up-to-date?
Value().SetValue(numRows, numCols, m_deviceId, const_cast<ElemType*>(array.data()), matrixFlagNormal);
// TODO: Get rid of that const_cast, as soon as after Ryan's Matrix-lib refactoring separated out SetValue() from external vs. from deep copy
VerifyDataSize(Value()); // sanity check
}
// form the actual tensor that describes the full object
TensorShape ComputationNodeBase::GetTensorShape(size_t rank) const
{
// If we have an MB layout then add the necessary sequence and time axes. If we have none, then absorb the column dimension.
TensorShape tensorShape = GetSampleLayout(); // TODO: Do we need to expect this tensor to have arbitrary strides? In case it came out of a Slice, Reshape, or Transpose op in-place?
if (HasMBLayout())
{
size_t i = (rank != SIZE_MAX) ? rank : tensorShape.GetRank();
tensorShape.AppendInPlace(i++, GetMBLayout()->GetNumParallelSequences());
tensorShape.AppendInPlace(i++, GetMBLayout()->GetNumTimeSteps());
}
return tensorShape;
}
// determine the sample tensor dimension to use for operations based on output and all inputs
// 'Sample tensor' means we only consider single samples. If we have an MBLayout, that is the sample layout of a single matrix column.
// TODO: Turn rank into a member variable, and call this method once in validation (currently called for every single ForwardProp/BackpropTo()).
size_t ComputationNodeBase::DetermineElementwiseTensorRank() const
{
// determine largest tensor dimension amongst the sample shapes of output and the selected inputs
size_t maxRank = GetSampleLayout().GetRank();
for (size_t i = 0; i < GetNumInputs(); i++)
{
size_t rank = Input(i)->GetSampleLayout().GetRank();
if (maxRank < rank)
maxRank = rank;
}
return maxRank;
}
// helper function for validation
// In complex cases of convolution, dimensions are quite difficult for a user to know/derive.
// This is a feature that allows a node to help resizing its input node to the expected value
// iff that input must be a learnable parameter.
void ComputationNodeBase::ValidateInferBinaryInputDims()
{
// limited inference of children dimensions
// if dimension not specified we assume two operands' dimensions should be the same
// NOTE: The assert is set to check if >= 2 since this is called from nodes which have more than two children.
// The number of children is formally verified elsewhere, so this will not break consistency.
assert(m_inputs.size() >= 2);
for (size_t index = 0; index < 2; index++)
{
auto in = Input( index);
auto other = Input(1 - index);
// borrow any unset dimension on one input from the other input
in->ValidateInferInputDimsFrom(other->GetSampleLayout());
}
}
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);
}
