void Variable::SetValue(const NDArrayViewPtr& value)
{
if (!IsParameter())
LogicError("Variable::SetValue can be only invoked on a Parameter variable!");
else if (GetDataType() != value->GetDataType())
LogicError("Variable::SetValue: 'source' and 'destination' have different data types!");
else if (Shape() != value->Shape() && (AsTensorShape(Shape()) != AsTensorShape(value->Shape())))
LogicError("Variable::SetValue: 'source' and 'destination' have different shapes!");
bool alreadySet = false;
if (m_dataFields->m_initValueFlag)
{
// In the case of lazy initialization, try to avoid the redundant call to the initializer.
std::call_once(*m_dataFields->m_initValueFlag, [=, &value, &alreadySet] {
// If the variable hasn't been initialized yet, clone the content of the supplied value and delete the initializer.
m_dataFields->m_value = value->DeepClone(*m_dataFields->m_valueInitializationDevice, false);
m_dataFields->m_valueInitializer = nullptr;
m_dataFields->m_valueInitializationDevice = nullptr;
alreadySet = true;
});
}
assert(m_dataFields->m_value != nullptr);
if (!alreadySet)
{
// alreadySet is false, the lambda above wasn't called and the variable has been initialized before,
// get a pointer to its value and simply copy the content of the supplied value.
m_dataFields->m_value->CopyFrom(*value);
}
}
void TensorBoardFileWriter::WriteImage(const std::wstring& name, NDArrayViewPtr imageData, uint64_t step)
{
assert(imageData != nullptr);
tensorflow::Event event;
event.set_wall_time(static_cast<double>(std::time(0)));
tensorflow::Summary* summary = event.mutable_summary();
std::vector<size_t> dimensions = imageData->Shape().Dimensions();
const size_t batch_size = dimensions.at(3);
const size_t depth = dimensions.at(2);
const size_t width = dimensions.at(1);
const size_t height = dimensions.at(0);
const DataType dtype = imageData->GetDataType();
std::vector<size_t> start(4, 0);
std::vector<size_t> extent;
extent.push_back(height);
extent.push_back(width);
extent.push_back(depth);
extent.push_back(1);
const int compression = -1;
const std::vector<size_t> imageDim({height, width, depth});
NDShape imageShape(imageDim);
for (size_t i = 0; i < batch_size; i++) {
tensorflow::Summary::Value* summaryValue = summary->add_value();
summaryValue->set_tag(ToString(name) + "/image/" + std::to_string(i));
tensorflow::Summary::Image* summaryImage = summaryValue->mutable_image();
summaryImage->set_height(height);
summaryImage->set_width(width);
summaryImage->set_colorspace(depth);
start.back() = static_cast<size_t>(i);
auto image = imageData->SliceView(start, extent)->AsShape(imageDim);
vector<uchar> buffer;
switch (dtype)
{
case DataType::Float:
WriteImageToBuffer(image->WritableDataBuffer<float>(), height, width, CV_32FC(depth), buffer);
break;
case DataType::Double:
WriteImageToBuffer(image->WritableDataBuffer<double>(), height, width, CV_64FC(depth), buffer);
break;
default:
fprintf(stderr, "TensorBoardFileWriter: Unsupported data type: %d ", static_cast<int>(dtype));
break;
}
string str(buffer.begin(), buffer.end());
summaryImage->set_encoded_image_string(str);
}
WriteRecord(Serialize(event));
}
Value::Value(const NDArrayViewPtr& data, const NDMaskPtr& mask)
: m_data(data), m_mask(mask)
{
if (mask != nullptr)
{
auto dataShape = data->Shape();
auto maskShape = mask->Shape();
if (maskShape.NumAxes() > dataShape.NumAxes())
InvalidArgument("The number of axes of the mask of a Value object cannot exceed the number of axes of the data NDArrayView object");
if (dataShape.SubShape(dataShape.NumAxes() - maskShape.NumAxes()) != maskShape)
InvalidArgument("Invalid Value object; the data and mask are incompatible. The trailing dimensions of the data do not match the dimensions of the mask");
}
}
Value::Value(const NDArrayViewPtr& data, const NDMaskPtr& mask)
: m_data(data), m_mask(mask)
{
if (mask != nullptr)
{
auto dataShape = data->Shape();
auto maskShape = mask->Shape();
if (maskShape.Rank() > dataShape.Rank())
InvalidArgument("The rank (%d) of the mask of a Value object cannot exceed the rank (%d) of the data NDArrayView object", (int)maskShape.Rank(), (int)dataShape.Rank());
if (dataShape.SubShape(dataShape.Rank() - maskShape.Rank()) != maskShape)
InvalidArgument("Invalid Value object; the data and mask are incompatible. The trailing dimensions of the data with shape %S do not match the dimensions of the mask with shape %S", AsStringForErrorReporting(dataShape).c_str(), AsStringForErrorReporting(maskShape).c_str());
}
}
void DistributedLearnerBase::ConvertToOrdered(const std::unordered_map<Parameter, NDArrayViewPtr>& gradientValues, std::vector<std::pair<Parameter, NDArrayViewPtr>>& result, std::unordered_map<Parameter, NDArrayViewPtr>* convertedGradientValues)
{
result.reserve(gradientValues.size());
result.clear();
if (convertedGradientValues)
convertedGradientValues->clear();
for (auto g : gradientValues)
{
NDArrayViewPtr p = g.second;
// convert sparse gradient to dense for accumulation
if (m_convertSparseToDense && p->GetStorageFormat() != StorageFormat::Dense)
{
NDArrayViewPtr pDense = MakeSharedObject<NDArrayView>(0, p->GetDataType(), p->Shape(), p->Device());
pDense->CopyFrom(*p);
p = pDense;
}
auto pair = std::make_pair(g.first, p);
result.push_back(pair);
if (convertedGradientValues)
convertedGradientValues->insert(pair);
}
std::sort(result.begin(), result.end(),
[](const std::pair<Parameter, NDArrayViewPtr>& a, const std::pair<Parameter, NDArrayViewPtr>& b) { return a.first.Uid() < b.first.Uid(); });
}
inline size_t GetBufferSize(const NDArrayViewPtr& viewPtr)
{
return viewPtr->Shape().TotalSize() * DataTypeSize(viewPtr->GetDataType());
}