onert::backend::train::ops Namespace Reference

Data Structures
class	BackPropAccumulator
class	BackPropInitializer
class	BinaryArithmeticLayer
class	ConvolutionLayer
class	DepthwiseConvolutionLayer
class	ElementwiseActivationLayer
class	FullyConnectedLayer
class	GradientApplier
class	LossCategoricalCrossentropyLayer
class	LossLayer
class	LossMeanSquaredErrorLayer
class	MeanLayer
class	PadLayer
class	PoolLayer
class	ReshapeLayer
class	SoftMaxLayer
class	TrainingKernelRegistry

Typedefs
using	OperandType = onert::ir::DataType

Enumerations
enum class	ArithmeticType { kAdd , kSub , kMul , kDiv }
enum class	ElementwiseActivationType { kReLU }
enum class	LossType { kMSE }
enum class	PoolType { kMax , kAvg }

Functions
nnfw::cker::Shape	getShape (const IPortableTensor *tensor)
	Get shape of tensor.
const IPortableTensor *	backpropActivation (const ir::Activation &activation, const IPortableTensor *output, const IPortableTensor *input_backprop, IPortableTensor *output_backprop)
	backpropagate acitvation
void	biasGrad (const IPortableTensor *input_backprop, IPortableTensor *bias_grad)
	backpropagate bias
nnfw::cker::train::LossReductionType	convertLossReductionType (ir::train::LossReductionType type)
	convert loss reduction type

Typedef Documentation

OperandType

using onert::backend::train::ops::OperandType = typedef onert::ir::DataType

Definition at line 27 of file OperationUtils.h.

Enumeration Type Documentation

ArithmeticType

enum class onert::backend::train::ops::ArithmeticType

strong

Enumerator
kAdd
kSub
kMul
kDiv

Definition at line 29 of file BinaryArithmeticLayer.h.

{
  kAdd,
  kSub,
  kMul,
  kDiv,
};

ElementwiseActivationType

enum class onert::backend::train::ops::ElementwiseActivationType

strong

Enumerator
kReLU

Definition at line 28 of file ElementwiseActivationLayer.h.

{
  kReLU,
};

LossType

enum class onert::backend::train::ops::LossType

strong

Enumerator
kMSE

Definition at line 29 of file LossLayer.h.

{
  kMSE,
};

PoolType

enum class onert::backend::train::ops::PoolType

strong

Enumerator
kMax
kAvg

Definition at line 39 of file PoolLayer.h.

{
  kMax,
  kAvg,
};

Function Documentation

backpropActivation()

const IPortableTensor * onert::backend::train::ops::backpropActivation	(	const ir::Activation &	activation,
		const IPortableTensor *	output,
		const IPortableTensor *	input_backprop,
		IPortableTensor *	output_backprop
	)

backpropagate acitvation

        -- forward direction -->

[ current layer ] -— [ next layer ] [ op | act ]

      <-- backward direction --

Parameters

activation	activation of current layer
output	forward direction's output of current layer
input_backprop	backward direction's output of next layer In other words, incoming gradient to current layer
output_backprop	backward direction's output of activation, In other words, outcoming gradient of current layer's acitvation If activation is NONE, this param can be nullptr

Returns

tensor that holds backpropagate result of activation If activation is NONE, just return input_backprop

Definition at line 44 of file OperationUtils.cc.

{
  assert(output != nullptr);
  assert(input_backprop != nullptr);
 
  // handle NONE - just propagate incoming gradient
  if (activation == ir::Activation::NONE)
  {
    return input_backprop;
  }
 
  assert(output_backprop != nullptr);
 
  // handle other activation
  switch (activation)
  {
    case ir::Activation::RELU:
      nnfw::cker::train::ReLUGrad(getShape(output), getBuffer<float>(output),
                                  getShape(input_backprop), getBuffer<float>(input_backprop),
                                  getShape(output_backprop), getBuffer<float>(output_backprop));
      break;
    case ir::Activation::RELU6:
      nnfw::cker::train::ReLU6Grad(getShape(output), getBuffer<float>(output),
                                   getShape(input_backprop), getBuffer<float>(input_backprop),
                                   getShape(output_backprop), getBuffer<float>(output_backprop));
      break;
    // TODO: Add other activation backpropagation here
    default:
      throw std::runtime_error("Unsupported activation type yet");
  }
  return output_backprop;
}

References getShape(), onert::ir::NONE, onert::ir::RELU, onert::ir::RELU6, nnfw::cker::train::ReLU6Grad(), and nnfw::cker::train::ReLUGrad().

Referenced by onert::backend::train::ops::BinaryArithmeticLayer::backward().

biasGrad()

void onert::backend::train::ops::biasGrad	(	const IPortableTensor *	input_backprop,
		IPortableTensor *	bias_grad
	)

backpropagate bias

Parameters

input_backprop	backward direction's output of next layer In other words, incoming gradient to current layer
bias_grad	gradient tensor of bias

Definition at line 80 of file OperationUtils.cc.

{
  assert(bias_grad);
 
  nnfw::cker::Shape input_backprop_shape = getShape(input_backprop);
  float *input_backprop_buffer = reinterpret_cast<float *>(input_backprop->buffer());
 
  nnfw::cker::Shape bias_grad_shape = getShape(bias_grad);
  float *bias_grad_buffer = getBuffer<float>(bias_grad);
 
  nnfw::cker::functor::biasReductionHelper(input_backprop_buffer, input_backprop_shape,
                                           bias_grad_buffer, bias_grad_shape);
}

References nnfw::cker::functor::biasReductionHelper(), onert::backend::ITensor::buffer(), and getShape().

convertLossReductionType()

nnfw::cker::train::LossReductionType onert::backend::train::ops::convertLossReductionType

(

ir::train::LossReductionType

type

)

convert loss reduction type

Parameters

type	loss reduction type defined in ir::train::LossReductionType

Returns

corresponding type defined in cker::train::LossReductionType

Definition at line 94 of file OperationUtils.cc.

{
  switch (type)
  {
    case ir::train::LossReductionType::SumOverBatchSize:
      return nnfw::cker::train::LossReductionType::SUM_OVER_BATCH_SIZE;
    case ir::train::LossReductionType::Sum:
      return nnfw::cker::train::LossReductionType::SUM;
    default:
      throw std::runtime_error("Unsupported LossReductionType");
  }
}

References nnfw::cker::train::SUM, onert::ir::train::Sum, nnfw::cker::train::SUM_OVER_BATCH_SIZE, onert::ir::train::SumOverBatchSize, and type.

Referenced by onert::backend::train::ops::LossCategoricalCrossentropyLayer::backward(), and onert::backend::train::ops::LossMeanSquaredErrorLayer::backward().

getShape()

nnfw::cker::Shape onert::backend::train::ops::getShape

(

const IPortableTensor *

tensor

)

Get shape of tensor.

Parameters

tensor

tensor to get shape

Returns

Shape to be used in cker

Definition at line 26 of file OperationUtils.cc.

{
  if (tensor == nullptr)
    return nnfw::cker::Shape();
 
  assert(!tensor->is_dynamic() && "Dynamic tensor is not supported yet");
 
  const ir::Shape &shape = tensor->get_info().shape();
  auto rank = shape.rank();
  nnfw::cker::Shape ret(rank);
  auto data = ret.DimsData();
  for (int i = 0; i < rank; ++i)
  {
    data[i] = shape.dim(i);
  }
  return ret;
}

References nnfw::cker::Shape::DimsData().

Referenced by backpropActivation(), onert::backend::train::ops::BackPropAccumulator::backward(), onert::backend::train::ops::BinaryArithmeticLayer::backward(), onert::backend::train::ops::LossCategoricalCrossentropyLayer::backward(), onert::backend::train::ops::LossMeanSquaredErrorLayer::backward(), onert::backend::train::ops::MeanLayer::backward(), onert::backend::train::ops::SoftMaxLayer::backward(), biasGrad(), onert::backend::train::ops::ElementwiseActivationLayer::configureBackward(), onert::backend::train::ops::DepthwiseConvolutionLayer::configureBackward(), onert::backend::train::ops::PadLayer::depad(), onert::backend::train::ops::LossCategoricalCrossentropyLayer::forward(), and onert::backend::train::ops::LossMeanSquaredErrorLayer::forward().