DML_LSTM_OPERATOR_DESC structure (directml.h)

Performs a one-layer long short term memory (LSTM) function on the input. This operator uses multiple gates to perform this layer. These gates are performed multiple times in a loop, dictated by the sequence length dimension and the SequenceLengthsTensor.

Equation for the forward direction

Equation for the backward direction

Equation legend

Syntax

struct DML_LSTM_OPERATOR_DESC {
  const DML_TENSOR_DESC           *InputTensor;
  const DML_TENSOR_DESC           *WeightTensor;
  const DML_TENSOR_DESC           *RecurrenceTensor;
  const DML_TENSOR_DESC           *BiasTensor;
  const DML_TENSOR_DESC           *HiddenInitTensor;
  const DML_TENSOR_DESC           *CellMemInitTensor;
  const DML_TENSOR_DESC           *SequenceLengthsTensor;
  const DML_TENSOR_DESC           *PeepholeTensor;
  const DML_TENSOR_DESC           *OutputSequenceTensor;
  const DML_TENSOR_DESC           *OutputSingleTensor;
  const DML_TENSOR_DESC           *OutputCellSingleTensor;
  UINT                            ActivationDescCount;
  const DML_OPERATOR_DESC         *ActivationDescs;
  DML_RECURRENT_NETWORK_DIRECTION Direction;
  float                           ClipThreshold;
  BOOL                            UseClipThreshold;
  BOOL                            CoupleInputForget;
};

Members

InputTensor

Type: const DML_TENSOR_DESC*

A tensor containing the input data, X. Packed (and potentially padded) into one 4-D tensor with the sizes of { 1, seq_length, batch_size, input_size }. seq_length is the dimension that is mapped to the index, t. The tensor doesn't support the DML_TENSOR_FLAG_OWNED_BY_DML flag.

WeightTensor

Type: const DML_TENSOR_DESC*

A tensor containing the weight data, W. Concatenation of W_[iofc] and W_B[iofc] (if bidirectional). The tensor has sizes { 1, num_directions, 4 * hidden_size, input_size }. The tensor doesn't support the DML_TENSOR_FLAG_OWNED_BY_DML flag.

RecurrenceTensor

Type: const DML_TENSOR_DESC*

A tensor containing the recurrence data, R. Concatenation of R_[iofc] and R_B[iofc] (if bidirectional). This tensor has sizes { 1, num_directions, 4 * hidden_size, hidden_size }. The tensor doesn't support the DML_TENSOR_FLAG_OWNED_BY_DML flag.

BiasTensor

Type: _Maybenull_ const DML_TENSOR_DESC*

An optional tensor containing the bias data, B. Concatenation of { W_b[iofc], R_b[iofc] }, and { W_Bb[iofc], R_Bb[iofc] } (if bidirectional). This tensor has sizes { 1, 1, num_directions, 8 * hidden_size }. If not specified, then defaults to 0 bias. The tensor doesn't support the DML_TENSOR_FLAG_OWNED_BY_DML flag.

HiddenInitTensor

Type: _Maybenull_ const DML_TENSOR_DESC*

An optional tensor containing the hidden node initializer data, H_(t-1). Contents of this tensor are only used on the first loop index t. If not specified, then defaults to 0. This tensor has sizes { 1, num_directions, batch_size, hidden_size }. The tensor doesn't support the DML_TENSOR_FLAG_OWNED_BY_DML flag.

CellMemInitTensor

Type: _Maybenull_ const DML_TENSOR_DESC*

An optional tensor containing the cell initializer data, C_(t-1). Contents of this tensor are only used on the first loop index t. If not specified, then defaults to 0. This tensor has sizes { 1, num_directions, batch_size, hidden_size }. The tensor doesn't support the DML_TENSOR_FLAG_OWNED_BY_DML flag.

SequenceLengthsTensor

Type: _Maybenull_ const DML_TENSOR_DESC*

An optional tensor containing an independent seq_length for each element in the batch. If not specified, then all sequences in the batch have length seq_length. This tensor has sizes { 1, 1, 1, batch_size }. The tensor doesn't support the DML_TENSOR_FLAG_OWNED_BY_DML flag.

PeepholeTensor

Type: _Maybenull_ const DML_TENSOR_DESC*

An optional tensor containing the weight data for peepholes, P. If not specified, then defaults to 0. Concatenation of P_[iof] and P_B[iof] (if bidirectional). This tensor has sizes { 1, 1, num_directions, 3 * hidden_size }. The tensor doesn't support the DML_TENSOR_FLAG_OWNED_BY_DML flag.

OutputSequenceTensor

Type: _Maybenull_ const DML_TENSOR_DESC*

An optional tensor with which to write the concatenation of all the intermediate output values of the hidden nodes, H_t. This tensor has sizes { seq_length, num_directions, batch_size, hidden_size }. seq_length is mapped to the loop index t.

OutputSingleTensor

Type: _Maybenull_ const DML_TENSOR_DESC*

An optional tensor with which to write the last output value of the hidden nodes, H_t. This tensor has sizes { 1, num_directions, batch_size, hidden_size }.

OutputCellSingleTensor

Type: _Maybenull_ const DML_TENSOR_DESC*

An optional tensor with which to write the last output value of the cell, C_t. This tensor has sizes { 1, num_directions, batch_size, hidden_size }.

ActivationDescCount

Type: UINT

This field determines the size of the ActivationDescs array.

ActivationDescs

Type: _Field_size_(ActivationDescCount) const DML_OPERATOR_DESC*

An array of DML_OPERATOR_DESC containing the descriptions of the activation operators f(), g(), and h(). f(), g(), and h() are defined independently of direction, meaning that if DML_RECURRENT_NETWORK_DIRECTION_FORWARD or DML_RECURRENT_NETWORK_DIRECTION_BACKWARD are supplied in Direction, then three activations must be provided. If DML_RECURRENT_NETWORK_DIRECTION_BIDIRECTIONAL is defined, then six activations must be provided. For bidirectional, activations must be provided f(), g(), and h() for forward followed by f(), g(), and h() for backwards.

Direction

Type: const DML_RECURRENT_NETWORK_DIRECTION*

The direction of the operator: forward, backward, or bidirectional.

ClipThreshold

Type: float

The cell clip threshold. Clipping bounds the elements of a tensor in the range of [-ClipThreshold, +ClipThreshold], and is applied to the input of activations.

UseClipThreshold

Type: BOOL

TRUE if ClipThreshold should be used. Otherwise, FALSE.

CoupleInputForget

Type: BOOL

TRUE if the input and forget gates should be coupled. Otherwise, FALSE.

Availability

This operator was introduced in DML_FEATURE_LEVEL_1_0.

Tensor constraints

BiasTensor, CellMemInitTensor, HiddenInitTensor, InputTensor, OutputCellSingleTensor, OutputSequenceTensor, OutputSingleTensor, PeepholeTensor, RecurrenceTensor, and WeightTensor must have the same DataType.

Tensor support

Tensor	Kind	Supported dimension counts	Supported data types
InputTensor	Input	4	FLOAT32, FLOAT16
WeightTensor	Input	4	FLOAT32, FLOAT16
RecurrenceTensor	Input	4	FLOAT32, FLOAT16
BiasTensor	Optional input	4	FLOAT32, FLOAT16
HiddenInitTensor	Optional input	4	FLOAT32, FLOAT16
CellMemInitTensor	Optional input	4	FLOAT32, FLOAT16
SequenceLengthsTensor	Optional input	4	UINT32
PeepholeTensor	Optional input	4	FLOAT32, FLOAT16
OutputSequenceTensor	Optional output	4	FLOAT32, FLOAT16
OutputSingleTensor	Optional output	4	FLOAT32, FLOAT16
OutputCellSingleTensor	Optional output	4	FLOAT32, FLOAT16

Requirements

Requirement	Value
Header	directml.h