LSTM Layer

Long short-term memory (LSTM) layer for recurrent neural network (RNN)

Since R2024b

expand all in page
  • LSTM Layer block

Libraries:
Deep Learning Toolbox / Deep Learning Layers / Sequence Layers

Description

The LSTM Layer block represents a recurrent neural network (RNN) layer that learns long-term dependencies between time steps in time-series and sequence data in the CT format (two dimensions corresponding to channels and time steps, in that order).

The exportNetworkToSimulink function generates this block to represent an lstmLayer object.

Limitations

  • The Layer parameter does not accept lstmLayer objects that have the HasStateInputs or HasStateOutputs properties set to 1 (true).

Ports

Input

expand all

Input data. The data must have two dimensions corresponding to channels and time steps, in that order, or one dimension corresponding to channels.

Data Types: single | double | int8 | int16 | int32 | int64 | uint8 | uint16 | uint32 | uint64 | fixed point

Output

expand all

The result of the LSTM operation. The output data has two dimensions corresponding to channels and time steps, in that order.

Data Types: single | double | int8 | int16 | int32 | int64 | uint8 | uint16 | uint32 | uint64 | fixed point

Parameters

expand all

To edit block parameters interactively, use the Property Inspector. From the Simulink® Toolstrip, on the Simulation tab, in the Prepare gallery, select Property Inspector.

Main

Specify the name of a workspace variable that contains an lstmLayer object from a trained network. The LSTM Layer block configures itself by using the properties of the object and calculates the block output by using the learnable parameters of the object.

Programmatic Use

Block Parameter: Layer
Type: workspace variable
Values: lstmLayer object
Default: 'layerObject'

Data format for the input data. The options use the same notation as the fmt argument of the dlarray object, except layer blocks do not support the Batch (B) dimension and instead assume an observation number of 1.

Programmatic Use

Block Parameter: DataFormat
Type: character vector
Values: 'CT'
Default: 'CT'

Whether to use stateful prediction, specified as a boolean. If true, the block maintains the cell state and hidden state between time steps. If false, the block performs stateless prediction by resetting cell states and hidden states to their initial values at the beginning of each time step. Use stateless prediction for frame-based processing where Simulink step time represents frame period and the network processes multiple samples at each time step. For more information, see Sample- and Frame-Based Concepts (DSP System Toolbox).

Programmatic Use

Block Parameter: StatefulPrediction
Type: character vector
Values: 'on' | 'off'
Default: 'on'

Data Types

If the object that you pass as the value of the Layer parameter uses the tanh state activation function or the sigmoid gate activation function, then the block uses the approximation method that you specify to compute the layer output.

Approximation MethodData Types SupportedWhen to Use This Method
None (default)

Floating-point

You are processing only floating-point data.

CORDIC

Floating-point (double and single) and fixed-point with a Bias value of 0 and a Slope value of a power of 2

You are processing fixed-point data and want to deploy to FPGA hardware.

Lookup

Floating-point and fixed-point

You are processing fixed-point data and want to generate C/C++ code.

For more information about the CORDIC approximation method, see cordictanh (Fixed-Point Designer).

Programmatic Use

Block Parameter: ApproximationMethod
Type: character vector
Values: 'None' | 'CORDIC' | 'Lookup'
Default: 'None'

Lower value of the output range that the software checks.

The software uses the minimum to perform:

Tips

Output minimum does not saturate or clip the actual output signal. Use the Saturation (Simulink) block instead.

Dependencies

To enable this parameter, set Output data type to a value other than Inherit: Inherit via internal rule.

Programmatic Use

To set the block parameter value programmatically, use the set_param (Simulink) function.

Parameter: OutMin
Values: '[]' (default) | scalar in quotes

Upper value of the output range that the software checks.

The software uses the maximum value to perform:

Tips

Output maximum does not saturate or clip the actual output signal. Use the Saturation (Simulink) block instead.

Dependencies

To enable this parameter, set Output data type to a value other than Inherit: Inherit via internal rule.

Programmatic Use

To set the block parameter value programmatically, use the set_param (Simulink) function.

Parameter: OutMax
Values: '[]' (default) | scalar in quotes

Choose the data type for the output. The type can be inherited, specified directly, or expressed as a data type object such as Simulink.NumericType. When you choose Inherit: Inherit via internal rule, Simulink sets the output data type to the same type as the hidden state.

Programmatic Use

To set the block parameter value programmatically, use the set_param (Simulink) function.

Parameter: OutDataTypeStr
Values: 'Inherit: Inherit via internal rule' (default) | <data type expression>

Select this parameter to prevent the fixed-point tools from overriding the Output data type you specify on the block. For more information, see Use Lock Output Data Type Setting (Fixed-Point Designer).

Programmatic Use

To set the block parameter value programmatically, use the set_param (Simulink) function.

Parameter: LockScale
Values: 'off' (default) | 'on'

Specify the rounding mode for fixed-point operations. For more information, see Rounding Modes (Fixed-Point Designer).

Block parameters always round to the nearest representable value. To control the rounding of a block parameter, enter an expression using a MATLAB® rounding function in the mask field.

Programmatic Use

To set the block parameter value programmatically, use the set_param (Simulink) function.

Parameter: RndMeth
Values: 'Floor' (default) | 'Ceiling' | 'Convergent' | 'Nearest' | 'Round' | 'Simplest' | 'Zero'

Specify whether integer overflows saturate or wrap.

  • on — Overflows saturate to either the minimum or maximum value that the data type can represent.

  • off — Overflows wrap to the appropriate value that the data type can represent.

For example, the maximum value that the signed 8-bit integer int8 can represent is 127. Any block operation result greater than the maximum value causes overflow of the 8-bit integer.

  • With this parameter selected, the block output saturates at 127. Similarly, the block output saturates at a minimum output value of –128.

  • With this parameter cleared, the software interprets the overflow-causing value as int8, which can produce an unintended result. For example, a block result of 130 (binary 1000 0010) expressed as int8 is –126.

Tips

  • Set this parameter to on when your model has a possible overflow and you want explicit saturation protection in the generated code.

  • To optimize the efficiency of your generated code, keep the default off setting for this parameter. Using the default setting also helps you avoid overspecifying how the block handles out-of-range signals. For more information, see Troubleshoot Signal Range Errors (Simulink).

  • When you select this parameter, saturation applies to every internal operation on the block, not just the output or the result.

  • In general, the code generation process can detect when overflow is not possible. In this case, the code generator does not produce saturation code.

Programmatic Use

To set the block parameter value programmatically, use the set_param (Simulink) function.

Parameter: SaturateOnIntegerOverflow
Values: 'off' (default) | 'on'

The block casts the value of the InputWeights property of the object that you specify with the Layer parameter to this data type. The type can be inherited, specified directly, or expressed as a data type object such as Simulink.NumericType.

Programmatic Use

To set the block parameter value programmatically, use the set_param (Simulink) function.

Parameter: InputWeightsDataTypeStr
Values: 'Inherit: Inherit via back propagation' (default) | 'Inherit: Inherit from 'Constant value'' | <data type expression>

The block casts the value of the RecurrentWeights property of the object that you specify with the Layer parameter to this data type. The type can be inherited, specified directly, or expressed as a data type object such as Simulink.NumericType.

Programmatic Use

To set the block parameter value programmatically, use the set_param (Simulink) function.

Parameter: RecurrentWeightsDataTypeStr
Values: 'Inherit: Inherit via back propagation' (default) | 'Inherit: Inherit from 'Constant value'' | <data type expression>

The block casts the value of the Bias property of the object that you specify with the Layer parameter to this data type. The type can be inherited, specified directly, or expressed as a data type object such as Simulink.NumericType.

Programmatic Use

To set the block parameter value programmatically, use the set_param (Simulink) function.

Parameter: BiasDataTypeStr
Values: 'Inherit: Inherit via back propagation' (default) | 'Inherit: Inherit from 'Constant value'' | <data type expression>

The block casts the value of the CellState property of the object that you specify with the Layer parameter to this data type. The type can be inherited, specified directly, or expressed as a data type object such as Simulink.NumericType.

This parameter affects only the initial cell state, c0. To cast later cell state values, use the Cell state parameter. For more information, see Long Short-Term Memory Layer.

Programmatic Use

To set the block parameter value programmatically, use the set_param (Simulink) function.

Parameter: InitialCellStateDataTypeStr
Values: 'Inherit: Inherit via back propagation' (default) | 'Inherit: Inherit from 'Constant value'' | <data type expression>

Choose the data type for the output of the Sum block ForIteratorSubsystem/LSTMCore/CellAdd inside the LSTM Layer block. The type can be inherited, specified directly, or expressed as a data type object such as Simulink.NumericType. When you select Inherit: Inherit via internal rule, Simulink chooses a data type to balance numerical accuracy, performance, and generated code size, while taking into account the properties of the embedded target hardware.

For a time step t, the Sum block computes the cell state ct as ct=ft⊙ct-1+it⊙gt, where i, f, and g denote the input gate, forget gate, and cell candidate gate, respectively, and ⊙ denotes the Hadamard product (element-wise multiplication of vectors). For more information, see Long Short-Term Memory Layer.

Programmatic Use

To set the block parameter value programmatically, use the set_param (Simulink) function.

Parameter: CellStateDataTypeStr
Values: 'Inherit: Inherit via internal rule' (default) | 'Inherit: Keep MSB' | 'Inherit: Keep LSB' | 'Inherit: Inherit via back propagation' | 'Inherit: Same as first input' | 'Inherit: Same as accumulator' | <data type expression>

The block casts the value of the HiddenState property of the object that you specify with the Layer parameter to this data type. The type can be inherited, specified directly, or expressed as a data type object such as Simulink.NumericType.

This parameter affects only the initial hidden state, h0. To cast later hidden state values, use the Hidden state parameter. For more information, see Long Short-Term Memory Layer.

Programmatic Use

To set the block parameter value programmatically, use the set_param (Simulink) function.

Parameter: InitialHiddenStateDataTypeStr
Values: 'Inherit: Inherit via back propagation' (default) | 'Inherit: Inherit from 'Constant value'' | <data type expression>

Choose the data type for the output of the Product block ForIteratorSubsystem/LSTMCore/HiddenStateProduct inside the LSTM Layer block. The type can be inherited, specified directly, or expressed as a data type object such as Simulink.NumericType. When you select Inherit: Inherit via internal rule, Simulink chooses a data type to balance numerical accuracy, performance, and generated code size, while taking into account the properties of the embedded target hardware.

For a time step t, the Product block computes the hidden state ht as ht=ot⊙σc(ct), where o denotes the output gate, σc denotes the state activation function, ⊙ denotes the Hadamard product (element-wise multiplication of vectors), and ct denotes the cell state at the time step. For more information, see Long Short-Term Memory Layer.

Programmatic Use

To set the block parameter value programmatically, use the set_param (Simulink) function.

Parameter: HiddenStateDataTypeStr
Values: 'Inherit: Inherit via internal rule' (default) | 'Inherit: Keep MSB' | 'Inherit: Keep LSB' | 'Inherit: Inherit via back propagation' | 'Inherit: Same as first input' | 'Inherit: Same as accumulator' | <data type expression>

Choose the data type for the output of the Product block InputWeightsMatrixMultiply/W*x inside the LSTM Layer block. The type can be inherited, specified directly, or expressed as a data type object such as Simulink.NumericType. When you select Inherit: Inherit via internal rule, Simulink chooses a data type to balance numerical accuracy, performance, and generated code size, while taking into account the properties of the embedded target hardware.

For a time step t, the Product block computes the product of the input weights (W) and the input at the time step (xt). For more information, see Long Short-Term Memory Layer.

Programmatic Use

To set the block parameter value programmatically, use the set_param (Simulink) function.

Parameter: InputWeightsMatrixMulitplyOutDataTypeStr
Values: 'Inherit: Inherit via internal rule' (default) | 'Inherit: Keep MSB' | 'Inherit: Match scaling' | 'Inherit: Inherit via back propagation' | 'Inherit: Same as first input' | <data type expression>

Choose the data type for the output of the Product block ForIteratorSubsystem/LSTMCore/RecurrentWeightsMatrixMultiply/R*h_t-1 inside the LSTM Layer block. The type can be inherited, specified directly, or expressed as a data type object such as Simulink.NumericType. When you select Inherit: Inherit via internal rule, Simulink chooses a data type to balance numerical accuracy, performance, and generated code size, while taking into account the properties of the embedded target hardware.

For a time step t, the Product block computes the product of the recurrent weights (R) and the hidden state at the previous time step (ht-1). For more information, see Long Short-Term Memory Layer.

Programmatic Use

To set the block parameter value programmatically, use the set_param (Simulink) function.

Parameter: RecurrentWeightsMatrixMulitplyOutDataTypeStr
Values: 'Inherit: Inherit via internal rule' (default) | 'Inherit: Keep MSB' | 'Inherit: Match scaling' | 'Inherit: Inherit via back propagation' | 'Inherit: Same as first input' | <data type expression>

Choose the data type for the accumulator of the Sum block ForIteratorSubsystem/LSTMCore/LinearGateAdd/Wx+Rh+b inside the LSTM Layer block. The type can be inherited, specified directly, or expressed as a data type object such as Simulink.NumericType. When you select Inherit: Inherit via internal rule, Simulink chooses a data type to balance numerical accuracy, performance, and generated code size, while taking into account the properties of the embedded target hardware.

For a time step t, Sum block computes the sum Wxt+Rht-1+b, where W denotes the input weights, xt denotes the input at the time step, ht-1 denotes the hidden state at the previous time step, and b denotes the bias. For more information, see Long Short-Term Memory Layer.

Programmatic Use

To set the block parameter value programmatically, use the set_param (Simulink) function.

Parameter: LinearGateAccumDataTypeStr
Values: 'Inherit: Inherit via internal rule' (default) | 'Inherit: Same as first input' | <data type expression>

Choose the data type for the output of the Sum block ForIteratorSubsystem/LSTMCore/LinearGateAdd/Wx+Rh+b inside the LSTM Layer block. The type can be inherited, specified directly, or expressed as a data type object such as Simulink.NumericType. When you select Inherit: Inherit via internal rule, Simulink chooses a data type to balance numerical accuracy, performance, and generated code size, while taking into account the properties of the embedded target hardware.

For a time step t, Sum block computes the sum Wxt+Rht-1+b, where W denotes the input weights, xt denotes the input at the time step, ht-1 denotes the hidden state at the previous time step, and b denotes the bias. For more information, see Long Short-Term Memory Layer.

Programmatic Use

To set the block parameter value programmatically, use the set_param (Simulink) function.

Parameter: LinearGateAddOutDataTypeStr
Values: 'Inherit: Inherit via internal rule' (default) | 'Inherit: Keep MSB' | 'Inherit: Keep LSB' | 'Inherit: Inherit via back propagation' | 'Inherit: Same as first input' | 'Inherit: Same as accumulator' | <data type expression>

Choose the data type for the output of the gate activation function. For more information about gate activation functions, see GateActivationFunction. For more information on how an LSTM layer uses activation functions, see Long Short-Term Memory Layer.

The type can be inherited, specified directly, or expressed as a data type object such as Simulink.NumericType. When you select Inherit: Inherit via internal rule, Simulink chooses a data type to balance numerical accuracy, performance, and generated code size, while taking into account the properties of the embedded target hardware.

Programmatic Use

To set the block parameter value programmatically, use the set_param (Simulink) function.

Parameter: GateActivationoutDataTypeStr
Values: Inherit: Inherit via internal rule (default) | Inherit: Same as first input | <data type expression>

Choose the data type for the output of the state activation function. For more information about state activation functions, see StateActivationFunction. For more information on how an LSTM layer uses activation functions, see Long Short-Term Memory Layer.

The type can be inherited, specified directly, or expressed as a data type object such as Simulink.NumericType. When you select Inherit: Inherit via internal rule, Simulink chooses a data type to balance numerical accuracy, performance, and generated code size, while taking into account the properties of the embedded target hardware.

Programmatic Use

To set the block parameter value programmatically, use the set_param (Simulink) function.

Parameter: StateActivationOutDataTypeStr
Values: 'Inherit: Inherit via internal rule' (default) | 'Inherit: Same as first input' | <data type expression>

Choose the data type for the output of the Product blocks ForIteratorSubsystem/LSTMCore/f*c_t-1 and ForIteratorSubsystem/LSTMCore/i*g inside the LSTM Layer block. The type can be inherited, specified directly, or expressed as a data type object such as Simulink.NumericType. When you select Inherit: Inherit via internal rule, Simulink chooses a data type to balance numerical accuracy, performance, and generated code size, while taking into account the properties of the embedded target hardware.

For a time step t, the Product blocks compute the Hadamard products (element-wise multiplication of vectors, denoted as ⊙) ft⊙ct-1 and it⊙gt, where f denotes the forget gate, c denotes the cell state, i denotes the input gate, and g denotes the cell candidate. For more information, see Long Short-Term Memory Layer.

Programmatic Use

To set the block parameter value programmatically, use the set_param (Simulink) function.

Parameter: CellStateProductOutDataTypeStr
Values: 'Inherit: Inherit via internal rule' (default) | 'Inherit: Keep MSB' | 'Inherit: Inherit via back propagation' | 'Inherit: Same as first input' | <data type expression>

Choose the data type for the accumulator of the Sum block ForIteratorSubsystem/LSTMCore/CellAdd inside the LSTM Layer block. The type can be inherited, specified directly, or expressed as a data type object such as Simulink.NumericType. When you select Inherit: Inherit via internal rule, Simulink chooses a data type to balance numerical accuracy, performance, and generated code size, while taking into account the properties of the embedded target hardware.

For a time step t, the Sum block computes the cell state ct asct=ft⊙ct-1+it⊙gt,

where i, f, and g denote the input gate, forget gate, and cell candidate gate, respectively, and ⊙ denotes the Hadamard product (element-wise multiplication of vectors). For more information, see Long Short-Term Memory Layer.

Programmatic Use

To set the block parameter value programmatically, use the set_param (Simulink) function.

Parameter: CellStateProductOutDataTypeStr
Values: 'Inherit: Inherit via internal rule' (default) | 'Inherit: Same as first input' | <data type expression>

Execution

Specify the discrete interval between sample time hits or specify another type of sample time, such as continuous (0) or inherited (-1). For more options, see Types of Sample Time (Simulink).

By default, the block inherits its sample time based on the context of the block within the model.

Programmatic Use

To set the block parameter value programmatically, use the set_param (Simulink) function.

Parameter: SampleTime
Data Types: char
Values: '-1' (default) | scalar

Extended Capabilities

expand all

C/C++ Code Generation
Generate C and C++ code using Simulink® Coder™.

Version History

Introduced in R2024b

expand all

See Also

| | | | | | |

Topics