Block Module

Abstract Block

Base block class

class lava.lib.dl.slayer.block.base.AbstractAffine(neuron_params, in_neurons, out_neurons, weight_scale=1, weight_norm=False, pre_hook_fx=None, dynamics=True, mask=None, count_log=False)

Bases: Module

Abstract affine transform class. This should never be instantiated on its own.

Parameters:

neuron_params (dict, optional) – a dictionary of neuron parameter. Defaults to None.
in_neurons (int) – number of input neurons.
out_neurons (int) – number of output neurons.
weight_scale (int, optional) – weight initialization scaling. Defaults to 1.
weight_norm (bool, optional) – flag to enable weight normalization. Defaults to False.
pre_hook_fx (optional) – a function pointer or lambda that is applied to synaptic weights before synaptic operation. None means no transformation. Defaults to None.
dynamics (bool, optional) – flag to enable neuron dynamics. If False, only the dendrite current is returned. Defaults to True.
mask (bool array, optional) – boolean synapse mask that only enables relevant synapses. None means no masking is applied. Defaults to None.
count_log (bool, optional) – flag to return event count log. If True, an additional value of average event rate is returned. Defaults to False.

export_hdf5(handle)

Hdf5 export method for the block.

Parameters:: handle (file handle) – hdf5 handle to export block description.

forward(x): Forward computation method. The input can be either of NCT or NCHWT format.

property shape: Shape of the block.

class lava.lib.dl.slayer.block.base.AbstractAverage(num_outputs, count_log=False)

Bases: Module

Abstract average block class. This should never be instantiated on its own.

Parameters:

num_outputs (int) – number of output population groups.
count_log (bool, optional) – flag to return event count log. If True, an additional value of average event rate is returned. Defaults to False.

export_hdf5(handle)

Hdf5 export method for the block.

Parameters:: handle (file handle) – hdf5 handle to export block description.

forward(x): Forward computation method. The input can be either of NCT or NCHWT format.

property shape: Shape of the block.

class lava.lib.dl.slayer.block.base.AbstractConv(neuron_params, in_features, out_features, kernel_size, stride=1, padding=0, dilation=1, groups=1, weight_scale=1, weight_norm=False, pre_hook_fx=None, delay=False, delay_shift=True, count_log=False)

Bases: Module

Abstract convolution block class. This should never be instantiated on its own.

Parameters:

neuron_params (dict, optional) – a dictionary of neuron parameter. Defaults to None.
in_features (int) – number of input features.
out_features (int) – number of output features.
kernel_size (int) – kernel size.
stride (int or tuple of two ints, optional) – convolution stride. Defaults to 1.
padding (int or tuple of two ints, optional) – convolution padding. Defaults to 0.
dilation (int or tuple of two ints, optional) – convolution dilation. Defaults to 1.
groups (int, optional) – number of blocked connections. Defaults to 1.
weight_scale (int, optional) – weight initialization scaling. Defaults to 1.
weight_norm (bool, optional) – flag to enable weight normalization. Defaults to False.
pre_hook_fx (optional) – a function pointer or lambda that is applied to synaptic weights before synaptic operation. None means no transformation. Defaults to None.
delay (bool, optional) – flag to enable axonal delay. Defaults to False.
delay_shift (bool, optional) – flag to simulate spike propagation delay from one layer to next. Defaults to True.
count_log (bool, optional) – flag to return event count log. If True, an additional value of average event rate is returned. Defaults to False.

export_hdf5(handle)

Hdf5 export method for the block.

Parameters:: handle (file handle) – hdf5 handle to export block description.

forward(x): Forward computation method. The input must be in NCHWT format.

property shape: Shape of the block.

class lava.lib.dl.slayer.block.base.AbstractConvT(neuron_params, in_features, out_features, kernel_size, stride=1, padding=0, dilation=1, groups=1, weight_scale=1, weight_norm=False, pre_hook_fx=None, delay=False, delay_shift=True, count_log=False)

Bases: Module

Abstract convolution Traspose block class. This should never be instantiated on its own.

Parameters:

neuron_params (dict, optional) – a dictionary of neuron parameter. Defaults to None.
in_features (int) – number of input features.
out_features (int) – number of output features.
kernel_size (int) – kernel size.
stride (int or tuple of two ints, optional) – convolutionT stride. Defaults to 1.
padding (int or tuple of two ints, optional) – convolutionT padding. Defaults to 0.
dilation (int or tuple of two ints, optional) – convolutionT dilation. Defaults to 1.
groups (int, optional) – number of blocked connections. Defaults to 1.
weight_scale (int, optional) – weight initialization scaling. Defaults to 1.
weight_norm (bool, optional) – flag to enable weight normalization. Defaults to False.
pre_hook_fx (optional) – a function pointer or lambda that is applied to synaptic weights before synaptic operation. None means no transformation. Defaults to None.
delay (bool, optional) – flag to enable axonal delay. Defaults to False.
delay_shift (bool, optional) – flag to simulate spike propagation delay from one layer to next. Defaults to True.
count_log (bool, optional) – flag to return event count log. If True, an additional value of average event rate is returned. Defaults to False.

export_hdf5(handle)

Hdf5 export method for the block.

Parameters:: handle (file handle) – hdf5 handle to export block description.

forward(x)

property shape: Shape of the block.

class lava.lib.dl.slayer.block.base.AbstractDense(neuron_params, in_neurons, out_neurons, weight_scale=1, weight_norm=False, pre_hook_fx=None, delay=False, delay_shift=True, mask=None, count_log=False)

Bases: Module

Abstract dense block class. This should never be instantiated on its own.

Parameters:

neuron_params (dict, optional) – a dictionary of neuron parameter. Defaults to None.
in_neurons (int) – number of input neurons.
out_neurons (int) – number of output neurons.
weight_scale (int, optional) – weight initialization scaling. Defaults to 1.
weight_norm (bool, optional) – flag to enable weight normalization. Defaults to False.
pre_hook_fx (optional) – a function pointer or lambda that is applied to synaptic weights before synaptic operation. None means no transformation. Defaults to None.
delay (bool, optional) – flag to enable axonal delay. Defaults to False.
delay_shift (bool, optional) – flag to simulate spike propagation delay from one layer to next. Defaults to True.
mask (bool array, optional) – boolean synapse mask that only enables relevant synapses. None means no masking is applied. Defaults to None.
count_log (bool, optional) – flag to return event count log. If True, an additional value of average event rate is returned. Defaults to False.

export_hdf5(handle)

Hdf5 export method for the block.

Parameters:: handle (file handle) – hdf5 handle to export block description.

forward(x): Forward computation method. The input can be either of NCT or NCHWT format.

property shape: Shape of the block.

class lava.lib.dl.slayer.block.base.AbstractFlatten(count_log=False)

Bases: Module

Abstract flatten block class. This should never be instantiated on its own.

Parameters:: count_log (bool, optional) – flag to return event count log. If True, an additional value of average event rate is returned. Defaults to False.

export_hdf5(handle)

Hdf5 export method for the block.

Parameters:: handle (file handle) – hdf5 handle to export block description.

forward(x): Forward computation method. The input can be either of NCT or NCHWT format.

class lava.lib.dl.slayer.block.base.AbstractInput(neuron_params=None, weight=None, bias=None, delay_shift=True, count_log=False)

Bases: Module

Abstract input block class. This should never be instantiated on its own.

Parameters:

neuron_params (dict, optional) – a dictionary of neuron parameter. Defaults to None.
weight (float, optional) – weight for affine transform of input. None means no weight scaling. Defaults to None.
bias (float, optional) – bias for affine transform of input. None means no bias shift. Defaults to None.
delay_shift (bool, optional) – flag to simulate spike propagation delay from one layer to next. Defaults to True.
count_log (bool, optional) – flag to return event count log. If True, an additional value of average event rate is returned. Defaults to False.

export_hdf5(handle)

Hdf5 export method for the block.

Parameters:: handle (file handle) – hdf5 handle to export block description.

forward(x): Forward computation method. The input can be either of NCT or NCHWT format.

property shape: Shape of the block.

class lava.lib.dl.slayer.block.base.AbstractKWTA(neuron_params, in_neurons, out_neurons, num_winners, self_excitation=0.5, weight_scale=1, weight_norm=False, pre_hook_fx=None, delay_shift=True, requires_grad=True, count_log=False)

Bases: Module

Abstract K-Winner-Takes-All block class. This should never be instantiated on its own. The formulation is described as below:

s_\text{out}[t] = f_s\left(\mathbf{W}\,s_\text{in}[t] + \mathbf{R}\,s_{out}[t-1] + \alpha\,(N-2K)\right)\\ \mathbf{R} = \begin{bmatrix} a &-1 &\cdots &-1\\ -1 & a &\cdots &-1\\ \vdots &\vdots &\ddots &\vdots\\ -1 &-1 &\cdots & a \end{bmatrix},\qquad |a| < 1

Parameters:

neuron_params (dict, optional) – a dictionary of neuron parameter. Defaults to None.
in_neurons (int) – number of input neurons.
out_neurons (int) – number of output neurons.
num_winners ([type]) – number of winners.
self_excitation (float, optional) – self excitation factor. Defaults to 0.5.
weight_scale (int, optional) – weight initialization scaling. Defaults to 1.
weight_norm (bool, optional) – flag to enable weight normalization. Defaults to False.
pre_hook_fx (optional) – a function pointer or lambda that is applied to synaptic weights before synaptic operation. None means no transformation. Defaults to None.
delay_shift (bool, optional) – flag to simulate spike propagation delay from one layer to next. Defaults to True.
requires_grad (bool, optional) – flag for learnable recurrent synapse. Defaults to True.
count_log (bool, optional) – flag to return event count log. If True, an additional value of average event rate is returned. Defaults to False.

clamp()

export_hdf5(handle)

Hdf5 export method for the block.

Parameters:: handle (file handle) – hdf5 handle to export block description.

forward(x): Forward computation method. The input can be either of NCT or NCHWT format.

property shape: Shape of the block.

class lava.lib.dl.slayer.block.base.AbstractPool(neuron_params, kernel_size, stride=None, padding=0, dilation=1, weight_scale=1, weight_norm=False, pre_hook_fx=None, delay=False, delay_shift=True, count_log=False)

Bases: Module

Abstract input block class. This should never be instantiated on its own.

Parameters:

neuron_params (dict, optional) – a dictionary of neuron parameter. Defaults to None.
kernel_size (int or tuple of two ints) – size of pooling kernel.
stride (int or tuple of two ints, optional) – stride of pooling operation. Defaults to None.
padding (int or tuple of two ints, optional) – padding of pooling operation. Defaults to 0.
dilation (int or tuple of two ints, optional) – dilation of pooling kernel. Defaults to 1.
weight_scale (int, optional) – weight initialization scaling. Defaults to 1.
weight_norm (bool, optional) – flag to enable weight normalization. Defaults to False.
pre_hook_fx (optional) – a function pointer or lambda that is applied to synaptic weights before synaptic operation. None means no transformation. Defaults to None.
delay (bool, optional) – flag to enable axonal delay. Defaults to False.
delay_shift (bool, optional) – flag to simulate spike propagation delay from one layer to next. Defaults to True.
count_log (bool, optional) – flag to return event count log. If True, an additional value of average event rate is returned. Defaults to False.

export_hdf5(handle)

Hdf5 export method for the block.

Parameters:: handle (file handle) – hdf5 handle to export block description.

forward(x): Forward computation method. The input must be in NCHWT format.

property shape: Shape of the block.

class lava.lib.dl.slayer.block.base.AbstractRecurrent(neuron_params, in_neurons, out_neurons, weight_scale=1, weight_norm=False, pre_hook_fx=None, requires_grad=True, delay=False, delay_shift=True, count_log=False)

Bases: Module

Abstract recurrent block class. This should never be instantiated on its own. The recurrent formulation is described below:

s_\text{out}[t] = f_s\left(\mathbf{W}\,s_\text{in}[t] + \mathbf{R}\,s_{out}[t-1]\right)

Parameters:

neuron_params (dict, optional) – a dictionary of neuron parameter.Defaults to None.
in_neurons (int) – number of input neurons.
out_neurons (int) – number of output neurons.
weight_scale (int, optional) – weight initialization scaling. Defaults to 1.
weight_norm (bool, optional) – flag to enable weight normalization. Defaults to False.
pre_hook_fx (optional) – a function pointer or lambda that is applied to synaptic weights before synaptic operation. None means no transformation. Defaults to None.
requires_grad (bool, optional) – flag for learnable recurrent synapse. Defaults to True.
delay (bool, optional) – flag to enable axonal delay. Defaults to False.
delay_shift (bool, optional) – flag to simulate spike propagation delay from one layer to next. Defaults to True.
count_log (bool, optional) – flag to return event count log. If True, an additional value of average event rate is returned. Defaults to False.

export_hdf5(handle)

Hdf5 export method for the block.

Parameters:: handle (file handle) – hdf5 handle to export block description.

forward(x): Forward computation method. The input can be either of NCT or NCHWT format.

property shape: Shape of the block.

class lava.lib.dl.slayer.block.base.AbstractResidual(*args, **kwargs): Bases: Module

class lava.lib.dl.slayer.block.base.AbstractTimeDecimation(factor, count_log=False)

Bases: Module

Abstract time decimation block class. This should never be instantiated on its own.

Parameters:

factor (int, optional) – number of time units to decimate in a single bin. Must be in powers of 2.
count_log (bool, optional) – flag to return event count log. If True, an additional value of average event rate is returned. Defaults to False.

export_hdf5(handle)

Hdf5 export method for the block.

Parameters:: handle (file handle) – hdf5 handle to export block description.

forward(x): Forward computation method. The input can be either of NCT or NCHWT format.

property shape: Shape of the block.

class lava.lib.dl.slayer.block.base.AbstractUnpool(neuron_params, kernel_size, stride=None, padding=0, dilation=1, weight_scale=1, weight_norm=False, pre_hook_fx=None, delay=False, delay_shift=True, count_log=False)

Bases: Module

Abstract Unpool block class. This should never be instantiated on its own.

Parameters:

neuron_params (dict, optional) – a dictionary of neuron parameter. Defaults to None.
kernel_size (int or tuple of two ints) – size of unpooling kernel.
stride (int or tuple of two ints, optional) – stride of unpooling operation. Defaults to None.
padding (int or tuple of two ints, optional) – padding of unpooling operation. Defaults to 0.
dilation (int or tuple of two ints, optional) – dilation of unpooling kernel. Defaults to 1.
weight_scale (int, optional) – weight initialization scaling. Defaults to 1.
weight_norm (bool, optional) – flag to enable weight normalization. Defaults to False.
pre_hook_fx (optional) – a function pointer or lambda that is applied to synaptic weights before synaptic operation. None means no transformation. Defaults to None.
delay (bool, optional) – flag to enable axonal delay. Defaults to False.
delay_shift (bool, optional) – flag to simulate spike propagation delay from one layer to next. Defaults to True.
count_log (bool, optional) – flag to return event count log. If True, an additional value of average event rate is returned. Defaults to False.

export_hdf5(handle)

Hdf5 export method for the block.

Parameters:: handle (file handle) – hdf5 handle to export block description.

forward(x)

property shape: Shape of the block.

lava.lib.dl.slayer.block.base.step_delay(module, x)

Step delay computation. This simulates the 1 timestep delay needed for communication between layers.

Parameters:

module (module) – python module instance
x (torch.tensor) – Tensor data to be delayed.

CUrrent BAsed Leaky Integrate and Fire (CUBA) Block

CUBA-LIF layer blocks

class lava.lib.dl.slayer.block.cuba.AbstractCuba(*args, **kwargs)

Bases: Module

Abstract block class for Current Based Leaky Integrator neuron. This should never be instantiated on it’s own.

class lava.lib.dl.slayer.block.cuba.Affine(*args, **kwargs)

Bases: AbstractCuba, AbstractAffine

CUBA LIF affine transform class. This should never be instantiated on its own.

Parameters:

neuron_params (dict, optional) – a dictionary of CUBA LIF neuron parameter. Defaults to None.
in_neurons (int) – number of input neurons.
out_neurons (int) – number of output neurons.
weight_scale (int, optional) – weight initialization scaling. Defaults to 1.
weight_norm (bool, optional) – flag to enable weight normalization. Defaults to False.
pre_hook_fx (optional) – a function pointer or lambda that is applied to synaptic weights before synaptic operation. None means no transformation. Defaults to None.
dynamics (bool, optional) – flag to enable neuron dynamics. If False, only the dendrite current is returned. Defaults to True.
mask (bool array, optional) – boolean synapse mask that only enables relevant synapses. None means no masking is applied. Defaults to None.
count_log (bool, optional) – flag to return event count log. If True, an additional value of average event rate is returned. Defaults to False.

class lava.lib.dl.slayer.block.cuba.Average(*args, **kwargs)

Bases: AbstractAverage

CUBA LIF average block class. This should never be instantiated on its own.

Parameters:

num_outputs (int) – number of output population groups.
count_log (bool, optional) – flag to return event count log. If True, an additional value of average event rate is returned. Defaults to False.

class lava.lib.dl.slayer.block.cuba.Conv(*args, **kwargs)

Bases: AbstractCuba, AbstractConv

CUBA LIF convolution block class. This should never be instantiated on its own.

Parameters:

neuron_params (dict, optional) – a dictionary of CUBA LIF neuron parameter. Defaults to None.
in_features (int) – number of input features.
out_features (int) – number of output features.
kernel_size (int) – kernel size.
stride (int or tuple of two ints, optional) – convolution stride. Defaults to 1.
padding (int or tuple of two ints, optional) – convolution padding. Defaults to 0.
dilation (int or tuple of two ints, optional) – convolution dilation. Defaults to 1.
groups (int, optional) – number of blocked connections. Defaults to 1.
weight_scale (int, optional) – weight initialization scaling. Defaults to 1.
weight_norm (bool, optional) – flag to enable weight normalization. Defaults to False.
pre_hook_fx (optional) – a function pointer or lambda that is applied to synaptic weights before synaptic operation. None means no transformation. Defaults to None.
delay (bool, optional) – flag to enable axonal delay. Defaults to False.
delay_shift (bool, optional) – flag to simulate spike propagation delay from one layer to next. Defaults to True.
count_log (bool, optional) – flag to return event count log. If True, an additional value of average event rate is returned. Defaults to False.

class lava.lib.dl.slayer.block.cuba.ConvT(*args, **kwargs)

Bases: AbstractCuba, AbstractConvT

CUBA LIF convolution Traspose block class. This should never be instantiated on its own.

Parameters:

neuron_params (dict, optional) – a dictionary of CUBA LIF neuron parameter. Defaults to None.
in_features (int) – number of input features.
out_features (int) – number of output features.
kernel_size (int) – kernel size.
stride (int or tuple of two ints, optional) – convolutionT stride. Defaults to 1.
padding (int or tuple of two ints, optional) – convolutionT padding. Defaults to 0.
dilation (int or tuple of two ints, optional) – convolutionT dilation. Defaults to 1.
groups (int, optional) – number of blocked connections. Defaults to 1.
weight_scale (int, optional) – weight initialization scaling. Defaults to 1.
weight_norm (bool, optional) – flag to enable weight normalization. Defaults to False.
pre_hook_fx (optional) – a function pointer or lambda that is applied to synaptic weights before synaptic operation. None means no transformation. Defaults to None.
delay (bool, optional) – flag to enable axonal delay. Defaults to False.
delay_shift (bool, optional) – flag to simulate spike propagation delay from one layer to next. Defaults to True.
count_log (bool, optional) – flag to return event count log. If True, an additional value of average event rate is returned. Defaults to False.

class lava.lib.dl.slayer.block.cuba.Dense(*args, **kwargs)

Bases: AbstractCuba, AbstractDense

CUBA LIF dense block class. The block is 8 bit quantization ready.

Parameters:

neuron_params (dict, optional) – a dictionary of CUBA LIF neuron parameter. Defaults to None.
in_neurons (int) – number of input neurons.
out_neurons (int) – number of output neurons.
weight_scale (int, optional) – weight initialization scaling. Defaults to 1.
weight_norm (bool, optional) – flag to enable weight normalization. Defaults to False.
pre_hook_fx (optional) – a function pointer or lambda that is applied to synaptic weights before synaptic operation. None means no transformation. Defaults to None.
delay (bool, optional) – flag to enable axonal delay. Defaults to False.
delay_shift (bool, optional) – flag to simulate spike propagation delay from one layer to next. Defaults to True.
mask (bool array, optional) – boolean synapse mask that only enables relevant synapses. None means no masking is applied. Defaults to None.
count_log (bool, optional) – flag to return event count log. If True, an additional value of average event rate is returned. Defaults to False.

class lava.lib.dl.slayer.block.cuba.Flatten(*args, **kwargs)

Bases: AbstractFlatten

CUBA LIF flatten block class. This should never be instantiated on its own.

Parameters:: count_log (bool, optional) – flag to return event count log. If True, an additional value of average event rate is returned. Defaults to False.

class lava.lib.dl.slayer.block.cuba.Input(*args, **kwargs)

Bases: AbstractCuba, AbstractInput

CUBA LIF input block class. The block is 8 bit quantization ready.

Parameters:

neuron_params (dict, optional) – a dictionary of CUBA LIF neuron parameter. Defaults to None.
weight (float, optional) – weight for affine transform of input. None means no weight scaling. Defaults to None.
bias (float, optional) – bias for affine transform of input. None means no bias shift. Defaults to None.
delay_shift (bool, optional) – flag to simulate spike propagation delay from one layer to next. Defaults to True.
count_log (bool, optional) – flag to return event count log. If True, an additional value of average event rate is returned. Defaults to False.

class lava.lib.dl.slayer.block.cuba.KWTA(*args, **kwargs)

Bases: AbstractCuba, AbstractKWTA

CUBA LIF K-Winner-Takes-All block class. This should never be instantiated on its own. The formulation is described as below:

s_\text{out}[t] = f_s\left(\mathbf{W}\,s_\text{in}[t] + \mathbf{R}\,s_{out}[t-1] + \alpha\,(N-2K)\right)\\ \mathbf{R} = \begin{bmatrix} a &-1 &\cdots &-1\\ -1 & a &\cdots &-1\\ \vdots &\vdots &\ddots &\vdots\\ -1 &-1 &\cdots & a \end{bmatrix},\qquad |a| < 1

Parameters:

neuron_params (dict, optional) – a dictionary of CUBA LIF neuron parameter. Defaults to None.
in_neurons (int) – number of input neurons.
out_neurons (int) – number of output neurons.
num_winners ([type]) – number of winners.
self_excitation (float, optional) – self excitation factor. Defaults to 0.5.
weight_scale (int, optional) – weight initialization scaling. Defaults to 1.
weight_norm (bool, optional) – flag to enable weight normalization. Defaults to False.
pre_hook_fx (optional) – a function pointer or lambda that is applied to synaptic weights before synaptic operation. None means no transformation. Defaults to None.
delay_shift (bool, optional) – flag to simulate spike propagation delay from one layer to next. Defaults to True.
requires_grad (bool, optional) – flag for learnable recurrent synapse. Defaults to True.
count_log (bool, optional) – flag to return event count log. If True, an additional value of average event rate is returned. Defaults to False.

class lava.lib.dl.slayer.block.cuba.Pool(*args, **kwargs)

Bases: AbstractCuba, AbstractPool

CUBA LIF input block class. The block is 8 bit quantization ready.

Parameters:

neuron_params (dict, optional) – a dictionary of CUBA LIF neuron parameter. Defaults to None.
kernel_size (int or tuple of two ints) – size of pooling kernel.
stride (int or tuple of two ints, optional) – stride of pooling operation. Defaults to None.
padding (int or tuple of two ints, optional) – padding of pooling operation. Defaults to 0.
dilation (int or tuple of two ints, optional) – dilation of pooling kernel. Defaults to 1.
weight_scale (int, optional) – weight initialization scaling. Defaults to 1.
weight_norm (bool, optional) – flag to enable weight normalization. Defaults to False.
pre_hook_fx (optional) – a function pointer or lambda that is applied to synaptic weights before synaptic operation. None means no transformation. Defaults to None.
delay (bool, optional) – flag to enable axonal delay. Defaults to False.
delay_shift (bool, optional) – flag to simulate spike propagation delay from one layer to next. Defaults to True.
count_log (bool, optional) – flag to return event count log. If True, an additional value of average event rate is returned. Defaults to False.

class lava.lib.dl.slayer.block.cuba.Recurrent(*args, **kwargs)

Bases: AbstractCuba, AbstractRecurrent

CUBA LIF recurrent block class. This should never be instantiated on its own. The recurrent formulation is described below:

s_\text{out}[t] = f_s\left(\mathbf{W}\,s_\text{in}[t] + \mathbf{R}\,s_{out}[t-1]\right)

Parameters:

neuron_params (dict, optional) – a dictionary of CUBA LIF neuron parameter.Defaults to None.
in_neurons (int) – number of input neurons.
out_neurons (int) – number of output neurons.
weight_scale (int, optional) – weight initialization scaling. Defaults to 1.
weight_norm (bool, optional) – flag to enable weight normalization. Defaults to False.
pre_hook_fx (optional) – a function pointer or lambda that is applied to synaptic weights before synaptic operation. None means no transformation. Defaults to None.
requires_grad (bool, optional) – flag for learnable recurrent synapse. Defaults to True.
delay (bool, optional) – flag to enable axonal delay. Defaults to False.
delay_shift (bool, optional) – flag to simulate spike propagation delay from one layer to next. Defaults to True.
count_log (bool, optional) – flag to return event count log. If True, an additional value of average event rate is returned. Defaults to False.

class lava.lib.dl.slayer.block.cuba.TimeDecimation(*args, **kwargs)

Bases: AbstractTimeDecimation

CUBA LIF time decimation block class. This should never be instantiated on its own.

Parameters:

factor (int, optional) – number of time units to decimate in a single bin. Must be in powers of 2.
count_log (bool, optional) – flag to return event count log. If True, an additional value of average event rate is returned. Defaults to False.

class lava.lib.dl.slayer.block.cuba.Unpool(*args, **kwargs)

Bases: AbstractCuba, AbstractUnpool

CUBA LIF Unpool block class. The block is 8 bit quantization ready.

Parameters:

neuron_params (dict, optional) – a dictionary of CUBA LIF neuron parameter. Defaults to None.
kernel_size (int or tuple of two ints) – size of unpooling kernel.
stride (int or tuple of two ints, optional) – stride of unpooling operation. Defaults to None.
padding (int or tuple of two ints, optional) – padding of unpooling operation. Defaults to 0.
dilation (int or tuple of two ints, optional) – dilation of unpooling kernel. Defaults to 1.
weight_scale (int, optional) – weight initialization scaling. Defaults to 1.
weight_norm (bool, optional) – flag to enable weight normalization. Defaults to False.
pre_hook_fx (optional) – a function pointer or lambda that is applied to synaptic weights before synaptic operation. None means no transformation. Defaults to None.
delay (bool, optional) – flag to enable axonal delay. Defaults to False.
delay_shift (bool, optional) – flag to simulate spike propagation delay from one layer to next. Defaults to True.
count_log (bool, optional) – flag to return event count log. If True, an additional value of average event rate is returned. Defaults to False.

Resonate and Fire (R&F) Block

Resonate and Fire layer blocks.

class lava.lib.dl.slayer.block.rf.AbstractRF(*args, **kwargs)

Bases: Module

Abstract Resonate and Fire block class. This should never be instantiated on it’s own.

class lava.lib.dl.slayer.block.rf.Affine(*args, **kwargs)

Bases: AbstractRF, AbstractAffine

Resonate & Fire affine transform class. This should never be instantiated on its own.

Parameters:

neuron_params (dict, optional) – a dictionary of RF parameter. Defaults to None.
in_neurons (int) – number of input neurons.
out_neurons (int) – number of output neurons.
weight_scale (int, optional) – weight initialization scaling. Defaults to 1.
weight_norm (bool, optional) – flag to enable weight normalization. Defaults to False.
pre_hook_fx (optional) – a function pointer or lambda that is applied to synaptic weights before synaptic operation. None means no transformation. Defaults to None.
dynamics (bool, optional) – flag to enable neuron dynamics. If False, only the dendrite current is returned. Defaults to True.
mask (bool array, optional) – boolean synapse mask that only enables relevant synapses. None means no masking is applied. Defaults to None.
count_log (bool, optional) – flag to return event count log. If True, an additional value of average event rate is returned. Defaults to False.

class lava.lib.dl.slayer.block.rf.Average(*args, **kwargs)

Bases: AbstractAverage

Resonate & Fire average block class. This should never be instantiated on its own.

Parameters:

num_outputs (int) – number of output population groups.
count_log (bool, optional) – flag to return event count log. If True, an additional value of average event rate is returned. Defaults to False.

class lava.lib.dl.slayer.block.rf.Conv(*args, **kwargs)

Bases: AbstractRF, AbstractConv

Resonate & Fire convolution block class. This should never be instantiated on its own.

Parameters:

neuron_params (dict, optional) – a dictionary of RF parameter. Defaults to None.
in_features (int) – number of input features.
out_features (int) – number of output features.
kernel_size (int) – kernel size.
stride (int or tuple of two ints, optional) – convolution stride. Defaults to 1.
padding (int or tuple of two ints, optional) – convolution padding. Defaults to 0.
dilation (int or tuple of two ints, optional) – convolution dilation. Defaults to 1.
groups (int, optional) – number of blocked connections. Defaults to 1.
weight_scale (int, optional) – weight initialization scaling. Defaults to 1.
weight_norm (bool, optional) – flag to enable weight normalization. Defaults to False.
pre_hook_fx (optional) – a function pointer or lambda that is applied to synaptic weights before synaptic operation. None means no transformation. Defaults to None.
delay (bool, optional) – flag to enable axonal delay. Defaults to False.
delay_shift (bool, optional) – flag to simulate spike propagation delay from one layer to next. Defaults to True.
count_log (bool, optional) – flag to return event count log. If True, an additional value of average event rate is returned. Defaults to False.

class lava.lib.dl.slayer.block.rf.ConvT(*args, **kwargs)

Bases: AbstractRF, AbstractConvT

Resonate & Fire convolution Traspose block class. This should never be instantiated on its own.

Parameters:

neuron_params (dict, optional) – a dictionary of RF parameter. Defaults to None.
in_features (int) – number of input features.
out_features (int) – number of output features.
kernel_size (int) – kernel size.
stride (int or tuple of two ints, optional) – convolutionT stride. Defaults to 1.
padding (int or tuple of two ints, optional) – convolutionT padding. Defaults to 0.
dilation (int or tuple of two ints, optional) – convolutionT dilation. Defaults to 1.
groups (int, optional) – number of blocked connections. Defaults to 1.
weight_scale (int, optional) – weight initialization scaling. Defaults to 1.
weight_norm (bool, optional) – flag to enable weight normalization. Defaults to False.
pre_hook_fx (optional) – a function pointer or lambda that is applied to synaptic weights before synaptic operation. None means no transformation. Defaults to None.
delay (bool, optional) – flag to enable axonal delay. Defaults to False.
delay_shift (bool, optional) – flag to simulate spike propagation delay from one layer to next. Defaults to True.
count_log (bool, optional) – flag to return event count log. If True, an additional value of average event rate is returned. Defaults to False.

class lava.lib.dl.slayer.block.rf.Dense(*args, **kwargs)

Bases: AbstractRF, AbstractDense

Resonate & Fire dense block class. The block is 8 bit quantization ready.

Parameters:

neuron_params (dict, optional) – a dictionary of RF parameter. Defaults to None.
in_neurons (int) – number of input neurons.
out_neurons (int) – number of output neurons.
weight_scale (int, optional) – weight initialization scaling. Defaults to 1.
weight_norm (bool, optional) – flag to enable weight normalization. Defaults to False.
pre_hook_fx (optional) – a function pointer or lambda that is applied to synaptic weights before synaptic operation. None means no transformation. Defaults to None.
delay (bool, optional) – flag to enable axonal delay. Defaults to False.
delay_shift (bool, optional) – flag to simulate spike propagation delay from one layer to next. Defaults to True.
mask (bool array, optional) – boolean synapse mask that only enables relevant synapses. None means no masking is applied. Defaults to None.
count_log (bool, optional) – flag to return event count log. If True, an additional value of average event rate is returned. Defaults to False.

class lava.lib.dl.slayer.block.rf.Flatten(*args, **kwargs)

Bases: AbstractFlatten

Resonate & Fire flatten block class. This should never be instantiated on its own.

Parameters:: count_log (bool, optional) – flag to return event count log. If True, an additional value of average event rate is returned. Defaults to False.

class lava.lib.dl.slayer.block.rf.Input(*args, **kwargs)

Bases: AbstractRF, AbstractInput

Resonate & Fire input block class. The block is 8 bit quantization ready.

Parameters:

neuron_params (dict, optional) – a dictionary of RF parameter. Defaults to None.
weight (float, optional) – weight for affine transform of input. None means no weight scaling. Defaults to None.
bias (float, optional) – bias for affine transform of input. None means no bias shift. Defaults to None.
delay_shift (bool, optional) – flag to simulate spike propagation delay from one layer to next. Defaults to True.
count_log (bool, optional) – flag to return event count log. If True, an additional value of average event rate is returned. Defaults to False.

class lava.lib.dl.slayer.block.rf.KWTA(*args, **kwargs)

Bases: AbstractRF, AbstractKWTA

Resonate & Fire K-Winner-Takes-All block class. This should never be instantiated on its own. The formulation is described as below:

s_\text{out}[t] = f_s\left(\mathbf{W}\,s_\text{in}[t] + \mathbf{R}\,s_{out}[t-1] + \alpha\,(N-2K)\right)\\ \mathbf{R} = \begin{bmatrix} a &-1 &\cdots &-1\\ -1 & a &\cdots &-1\\ \vdots &\vdots &\ddots &\vdots\\ -1 &-1 &\cdots & a \end{bmatrix},\qquad |a| < 1

Parameters:

neuron_params (dict, optional) – a dictionary of RF parameter. Defaults to None.
in_neurons (int) – number of input neurons.
out_neurons (int) – number of output neurons.
num_winners ([type]) – number of winners.
self_excitation (float, optional) – self excitation factor. Defaults to 0.5.
weight_scale (int, optional) – weight initialization scaling. Defaults to 1.
weight_norm (bool, optional) – flag to enable weight normalization. Defaults to False.
pre_hook_fx (optional) – a function pointer or lambda that is applied to synaptic weights before synaptic operation. None means no transformation. Defaults to None.
delay_shift (bool, optional) – flag to simulate spike propagation delay from one layer to next. Defaults to True.
requires_grad (bool, optional) – flag for learnable recurrent synapse. Defaults to True.
count_log (bool, optional) – flag to return event count log. If True, an additional value of average event rate is returned. Defaults to False.

class lava.lib.dl.slayer.block.rf.Pool(*args, **kwargs)

Bases: AbstractRF, AbstractPool

Resonate & Fire input block class. The block is 8 bit quantization ready.

Parameters:

neuron_params (dict, optional) – a dictionary of RF parameter. Defaults to None.
kernel_size (int or tuple of two ints) – size of pooling kernel.
stride (int or tuple of two ints, optional) – stride of pooling operation. Defaults to None.
padding (int or tuple of two ints, optional) – padding of pooling operation. Defaults to 0.
dilation (int or tuple of two ints, optional) – dilation of pooling kernel. Defaults to 1.
weight_scale (int, optional) – weight initialization scaling. Defaults to 1.
weight_norm (bool, optional) – flag to enable weight normalization. Defaults to False.
pre_hook_fx (optional) – a function pointer or lambda that is applied to synaptic weights before synaptic operation. None means no transformation. Defaults to None.
delay (bool, optional) – flag to enable axonal delay. Defaults to False.
delay_shift (bool, optional) – flag to simulate spike propagation delay from one layer to next. Defaults to True.
count_log (bool, optional) – flag to return event count log. If True, an additional value of average event rate is returned. Defaults to False.

class lava.lib.dl.slayer.block.rf.Recurrent(*args, **kwargs)

Bases: AbstractRF, AbstractRecurrent

Resonate & Fire recurrent block class. This should never be instantiated on its own. The recurrent formulation is described below:

s_\text{out}[t] = f_s\left(\mathbf{W}\,s_\text{in}[t] + \mathbf{R}\,s_{out}[t-1]\right)

Parameters:

neuron_params (dict, optional) – a dictionary of RF parameter.Defaults to None.
in_neurons (int) – number of input neurons.
out_neurons (int) – number of output neurons.
weight_scale (int, optional) – weight initialization scaling. Defaults to 1.
weight_norm (bool, optional) – flag to enable weight normalization. Defaults to False.
pre_hook_fx (optional) – a function pointer or lambda that is applied to synaptic weights before synaptic operation. None means no transformation. Defaults to None.
requires_grad (bool, optional) – flag for learnable recurrent synapse. Defaults to True.
delay (bool, optional) – flag to enable axonal delay. Defaults to False.
delay_shift (bool, optional) – flag to simulate spike propagation delay from one layer to next. Defaults to True.
count_log (bool, optional) – flag to return event count log. If True, an additional value of average event rate is returned. Defaults to False.

class lava.lib.dl.slayer.block.rf.TimeDecimation(*args, **kwargs)

Bases: AbstractTimeDecimation

Resonate & Fire time decimation block class. This should never be instantiated on its own.

Parameters:

factor (int, optional) – number of time units to decimate in a single bin. Must be in powers of 2.
count_log (bool, optional) – flag to return event count log. If True, an additional value of average event rate is returned. Defaults to False.

class lava.lib.dl.slayer.block.rf.Unpool(*args, **kwargs)

Bases: AbstractRF, AbstractUnpool

Resonate & Fire Unpool block class. The block is 8 bit quantization ready.

Parameters:

neuron_params (dict, optional) – a dictionary of RF parameter. Defaults to None.
kernel_size (int or tuple of two ints) – size of unpooling kernel.
stride (int or tuple of two ints, optional) – stride of unpooling operation. Defaults to None.
padding (int or tuple of two ints, optional) – padding of unpooling operation. Defaults to 0.
dilation (int or tuple of two ints, optional) – dilation of unpooling kernel. Defaults to 1.
weight_scale (int, optional) – weight initialization scaling. Defaults to 1.
weight_norm (bool, optional) – flag to enable weight normalization. Defaults to False.
pre_hook_fx (optional) – a function pointer or lambda that is applied to synaptic weights before synaptic operation. None means no transformation. Defaults to None.
delay (bool, optional) – flag to enable axonal delay. Defaults to False.
delay_shift (bool, optional) – flag to simulate spike propagation delay from one layer to next. Defaults to True.
count_log (bool, optional) – flag to return event count log. If True, an additional value of average event rate is returned. Defaults to False.

Resonate and Fire Izhikevich (R&F Iz) Block

Resonate and Fire - Izhikevich layer blocks.

class lava.lib.dl.slayer.block.rf_iz.AbstractRFIz(*args, **kwargs)

Bases: Module

Abstract Resonate and Fire - Izhikevich block class. This should never be instantiated on it’s own.

class lava.lib.dl.slayer.block.rf_iz.Affine(*args, **kwargs)

Bases: AbstractRFIz, AbstractAffine

Resonate & Fire Izhikevich affine transform class. This should never be instantiated on its own.

Parameters:

neuron_params (dict, optional) – a dictionary of RF-Izhikevich neuron parameter. Defaults to None.
in_neurons (int) – number of input neurons.
out_neurons (int) – number of output neurons.
weight_scale (int, optional) – weight initialization scaling. Defaults to 1.
weight_norm (bool, optional) – flag to enable weight normalization. Defaults to False.
pre_hook_fx (optional) – a function pointer or lambda that is applied to synaptic weights before synaptic operation. None means no transformation. Defaults to None.
dynamics (bool, optional) – flag to enable neuron dynamics. If False, only the dendrite current is returned. Defaults to True.
mask (bool array, optional) – boolean synapse mask that only enables relevant synapses. None means no masking is applied. Defaults to None.
count_log (bool, optional) – flag to return event count log. If True, an additional value of average event rate is returned. Defaults to False.

class lava.lib.dl.slayer.block.rf_iz.Average(*args, **kwargs)

Bases: AbstractAverage

Resonate & Fire Izhikevich average block class. This should never be instantiated on its own.

Parameters:

num_outputs (int) – number of output population groups.
count_log (bool, optional) – flag to return event count log. If True, an additional value of average event rate is returned. Defaults to False.

class lava.lib.dl.slayer.block.rf_iz.Conv(*args, **kwargs)

Bases: AbstractRFIz, AbstractConv

Resonate & Fire Izhikevich convolution block class. This should never be instantiated on its own.

Parameters:

neuron_params (dict, optional) – a dictionary of RF-Izhikevich neuron parameter. Defaults to None.
in_features (int) – number of input features.
out_features (int) – number of output features.
kernel_size (int) – kernel size.
stride (int or tuple of two ints, optional) – convolution stride. Defaults to 1.
padding (int or tuple of two ints, optional) – convolution padding. Defaults to 0.
dilation (int or tuple of two ints, optional) – convolution dilation. Defaults to 1.
groups (int, optional) – number of blocked connections. Defaults to 1.
weight_scale (int, optional) – weight initialization scaling. Defaults to 1.
weight_norm (bool, optional) – flag to enable weight normalization. Defaults to False.
pre_hook_fx (optional) – a function pointer or lambda that is applied to synaptic weights before synaptic operation. None means no transformation. Defaults to None.
delay (bool, optional) – flag to enable axonal delay. Defaults to False.
delay_shift (bool, optional) – flag to simulate spike propagation delay from one layer to next. Defaults to True.
count_log (bool, optional) – flag to return event count log. If True, an additional value of average event rate is returned. Defaults to False.

class lava.lib.dl.slayer.block.rf_iz.ConvT(*args, **kwargs)

Bases: AbstractRFIz, AbstractConvT

Resonate & Fire Izhikevich convolution Traspose block class. This should never be instantiated on its own.

Parameters:

neuron_params (dict, optional) – a dictionary of RF-Izhikevich neuron parameter. Defaults to None.
in_features (int) – number of input features.
out_features (int) – number of output features.
kernel_size (int) – kernel size.
stride (int or tuple of two ints, optional) – convolutionT stride. Defaults to 1.
padding (int or tuple of two ints, optional) – convolutionT padding. Defaults to 0.
dilation (int or tuple of two ints, optional) – convolutionT dilation. Defaults to 1.
groups (int, optional) – number of blocked connections. Defaults to 1.
weight_scale (int, optional) – weight initialization scaling. Defaults to 1.
weight_norm (bool, optional) – flag to enable weight normalization. Defaults to False.
pre_hook_fx (optional) – a function pointer or lambda that is applied to synaptic weights before synaptic operation. None means no transformation. Defaults to None.
delay (bool, optional) – flag to enable axonal delay. Defaults to False.
delay_shift (bool, optional) – flag to simulate spike propagation delay from one layer to next. Defaults to True.
count_log (bool, optional) – flag to return event count log. If True, an additional value of average event rate is returned. Defaults to False.

class lava.lib.dl.slayer.block.rf_iz.Dense(*args, **kwargs)

Bases: AbstractRFIz, AbstractDense

Resonate & Fire Izhikevich dense block class. The block is 8 bit quantization ready.

Parameters:

neuron_params (dict, optional) – a dictionary of RF-Izhikevich neuron parameter. Defaults to None.
in_neurons (int) – number of input neurons.
out_neurons (int) – number of output neurons.
weight_scale (int, optional) – weight initialization scaling. Defaults to 1.
weight_norm (bool, optional) – flag to enable weight normalization. Defaults to False.
pre_hook_fx (optional) – a function pointer or lambda that is applied to synaptic weights before synaptic operation. None means no transformation. Defaults to None.
delay (bool, optional) – flag to enable axonal delay. Defaults to False.
delay_shift (bool, optional) – flag to simulate spike propagation delay from one layer to next. Defaults to True.
mask (bool array, optional) – boolean synapse mask that only enables relevant synapses. None means no masking is applied. Defaults to None.
count_log (bool, optional) – flag to return event count log. If True, an additional value of average event rate is returned. Defaults to False.

class lava.lib.dl.slayer.block.rf_iz.Flatten(*args, **kwargs)

Bases: AbstractFlatten

Resonate & Fire Izhikevich flatten block class. This should never be instantiated on its own.

Parameters:: count_log (bool, optional) – flag to return event count log. If True, an additional value of average event rate is returned. Defaults to False.

class lava.lib.dl.slayer.block.rf_iz.Input(*args, **kwargs)

Bases: AbstractRFIz, AbstractInput

Resonate & Fire Izhikevich input block class. The block is 8 bit quantization ready.

Parameters:

neuron_params (dict, optional) – a dictionary of RF-Izhikevich neuron parameter. Defaults to None.
weight (float, optional) – weight for affine transform of input. None means no weight scaling. Defaults to None.
bias (float, optional) – bias for affine transform of input. None means no bias shift. Defaults to None.
delay_shift (bool, optional) – flag to simulate spike propagation delay from one layer to next. Defaults to True.
count_log (bool, optional) – flag to return event count log. If True, an additional value of average event rate is returned. Defaults to False.

class lava.lib.dl.slayer.block.rf_iz.KWTA(*args, **kwargs)

Bases: AbstractRFIz, AbstractKWTA

Resonate & Fire Izhikevich K-Winner-Takes-All block class. This should never be instantiated on its own. The formulation is described as below:

s_\text{out}[t] = f_s\left(\mathbf{W}\,s_\text{in}[t] + \mathbf{R}\,s_{out}[t-1] + \alpha\,(N-2K)\right)\\ \mathbf{R} = \begin{bmatrix} a &-1 &\cdots &-1\\ -1 & a &\cdots &-1\\ \vdots &\vdots &\ddots &\vdots\\ -1 &-1 &\cdots & a \end{bmatrix},\qquad |a| < 1

Parameters:

neuron_params (dict, optional) – a dictionary of RF-Izhikevich neuron parameter. Defaults to None.
in_neurons (int) – number of input neurons.
out_neurons (int) – number of output neurons.
num_winners ([type]) – number of winners.
self_excitation (float, optional) – self excitation factor. Defaults to 0.5.
weight_scale (int, optional) – weight initialization scaling. Defaults to 1.
weight_norm (bool, optional) – flag to enable weight normalization. Defaults to False.
pre_hook_fx (optional) – a function pointer or lambda that is applied to synaptic weights before synaptic operation. None means no transformation. Defaults to None.
delay_shift (bool, optional) – flag to simulate spike propagation delay from one layer to next. Defaults to True.
requires_grad (bool, optional) – flag for learnable recurrent synapse. Defaults to True.
count_log (bool, optional) – flag to return event count log. If True, an additional value of average event rate is returned. Defaults to False.

class lava.lib.dl.slayer.block.rf_iz.Pool(*args, **kwargs)

Bases: AbstractRFIz, AbstractPool

Resonate & Fire Izhikevich input block class. The block is 8 bit quantization ready.

Parameters:

neuron_params (dict, optional) – a dictionary of RF-Izhikevich neuron parameter. Defaults to None.
kernel_size (int or tuple of two ints) – size of pooling kernel.
stride (int or tuple of two ints, optional) – stride of pooling operation. Defaults to None.
padding (int or tuple of two ints, optional) – padding of pooling operation. Defaults to 0.
dilation (int or tuple of two ints, optional) – dilation of pooling kernel. Defaults to 1.
weight_scale (int, optional) – weight initialization scaling. Defaults to 1.
weight_norm (bool, optional) – flag to enable weight normalization. Defaults to False.
pre_hook_fx (optional) – a function pointer or lambda that is applied to synaptic weights before synaptic operation. None means no transformation. Defaults to None.
delay (bool, optional) – flag to enable axonal delay. Defaults to False.
delay_shift (bool, optional) – flag to simulate spike propagation delay from one layer to next. Defaults to True.
count_log (bool, optional) – flag to return event count log. If True, an additional value of average event rate is returned. Defaults to False.

class lava.lib.dl.slayer.block.rf_iz.Recurrent(*args, **kwargs)

Bases: AbstractRFIz, AbstractRecurrent

Resonate & Fire Izhikevich recurrent block class. This should never be instantiated on its own. The recurrent formulation is described below:

s_\text{out}[t] = f_s\left(\mathbf{W}\,s_\text{in}[t] + \mathbf{R}\,s_{out}[t-1]\right)

Parameters:

neuron_params (dict, optional) – a dictionary of RF-Izhikevich neuron parameter.Defaults to None.
in_neurons (int) – number of input neurons.
out_neurons (int) – number of output neurons.
weight_scale (int, optional) – weight initialization scaling. Defaults to 1.
weight_norm (bool, optional) – flag to enable weight normalization. Defaults to False.
pre_hook_fx (optional) – a function pointer or lambda that is applied to synaptic weights before synaptic operation. None means no transformation. Defaults to None.
requires_grad (bool, optional) – flag for learnable recurrent synapse. Defaults to True.
delay (bool, optional) – flag to enable axonal delay. Defaults to False.
delay_shift (bool, optional) – flag to simulate spike propagation delay from one layer to next. Defaults to True.
count_log (bool, optional) – flag to return event count log. If True, an additional value of average event rate is returned. Defaults to False.

class lava.lib.dl.slayer.block.rf_iz.TimeDecimation(*args, **kwargs)

Bases: AbstractTimeDecimation

Resonate & Fire Izhikevich time decimation block class. This should never be instantiated on its own.

Parameters:

factor (int, optional) – number of time units to decimate in a single bin. Must be in powers of 2.
count_log (bool, optional) – flag to return event count log. If True, an additional value of average event rate is returned. Defaults to False.

class lava.lib.dl.slayer.block.rf_iz.Unpool(*args, **kwargs)

Bases: AbstractRFIz, AbstractUnpool

Resonate & Fire Izhikevich Unpool block class. The block is 8 bit quantization ready.

Parameters:

neuron_params (dict, optional) – a dictionary of RF-Izhikevich neuron parameter. Defaults to None.
kernel_size (int or tuple of two ints) – size of unpooling kernel.
stride (int or tuple of two ints, optional) – stride of unpooling operation. Defaults to None.
padding (int or tuple of two ints, optional) – padding of unpooling operation. Defaults to 0.
dilation (int or tuple of two ints, optional) – dilation of unpooling kernel. Defaults to 1.
weight_scale (int, optional) – weight initialization scaling. Defaults to 1.
weight_norm (bool, optional) – flag to enable weight normalization. Defaults to False.
pre_hook_fx (optional) – a function pointer or lambda that is applied to synaptic weights before synaptic operation. None means no transformation. Defaults to None.
delay (bool, optional) – flag to enable axonal delay. Defaults to False.
delay_shift (bool, optional) – flag to simulate spike propagation delay from one layer to next. Defaults to True.
count_log (bool, optional) – flag to return event count log. If True, an additional value of average event rate is returned. Defaults to False.

Adaptive Leaky Integrate and Fire (ALIF) Block

Adaptive Leaky Integrate and Fire block layers.

class lava.lib.dl.slayer.block.alif.AbstractALIF(*args, **kwargs)

Bases: Module

Abstract Leaky Integrae and Fire block. This should never be instantiated on it’s own.

class lava.lib.dl.slayer.block.alif.Average(*args, **kwargs)

Bases: AbstractAverage

Adaptive LIF average block class. This should never be instantiated on its own.

Parameters:

num_outputs (int) – number of output population groups.
count_log (bool, optional) – flag to return event count log. If True, an additional value of average event rate is returned. Defaults to False.

class lava.lib.dl.slayer.block.alif.Conv(*args, **kwargs)

Bases: AbstractALIF, AbstractConv

Adaptive LIF convolution block class. This should never be instantiated on its own.

Parameters:

neuron_params (dict, optional) – a dictionary of Adaptive LIF neuron parameter. Defaults to None.
in_features (int) – number of input features.
out_features (int) – number of output features.
kernel_size (int) – kernel size.
stride (int or tuple of two ints, optional) – convolution stride. Defaults to 1.
padding (int or tuple of two ints, optional) – convolution padding. Defaults to 0.
dilation (int or tuple of two ints, optional) – convolution dilation. Defaults to 1.
groups (int, optional) – number of blocked connections. Defaults to 1.
weight_scale (int, optional) – weight initialization scaling. Defaults to 1.
weight_norm (bool, optional) – flag to enable weight normalization. Defaults to False.
pre_hook_fx (optional) – a function pointer or lambda that is applied to synaptic weights before synaptic operation. None means no transformation. Defaults to None.
delay (bool, optional) – flag to enable axonal delay. Defaults to False.
delay_shift (bool, optional) – flag to simulate spike propagation delay from one layer to next. Defaults to True.
count_log (bool, optional) – flag to return event count log. If True, an additional value of average event rate is returned. Defaults to False.

class lava.lib.dl.slayer.block.alif.ConvT(*args, **kwargs)

Bases: AbstractALIF, AbstractConvT

Adaptive LIF convolution Traspose block class. This should never be instantiated on its own.

Parameters:

neuron_params (dict, optional) – a dictionary of Adaptive LIF neuron parameter. Defaults to None.
in_features (int) – number of input features.
out_features (int) – number of output features.
kernel_size (int) – kernel size.
stride (int or tuple of two ints, optional) – convolutionT stride. Defaults to 1.
padding (int or tuple of two ints, optional) – convolutionT padding. Defaults to 0.
dilation (int or tuple of two ints, optional) – convolutionT dilation. Defaults to 1.
groups (int, optional) – number of blocked connections. Defaults to 1.
weight_scale (int, optional) – weight initialization scaling. Defaults to 1.
weight_norm (bool, optional) – flag to enable weight normalization. Defaults to False.
pre_hook_fx (optional) – a function pointer or lambda that is applied to synaptic weights before synaptic operation. None means no transformation. Defaults to None.
delay (bool, optional) – flag to enable axonal delay. Defaults to False.
delay_shift (bool, optional) – flag to simulate spike propagation delay from one layer to next. Defaults to True.
count_log (bool, optional) – flag to return event count log. If True, an additional value of average event rate is returned. Defaults to False.

class lava.lib.dl.slayer.block.alif.Dense(*args, **kwargs)

Bases: AbstractALIF, AbstractDense

Adaptive LIF dense block class. The block is 8 bit quantization ready.

Parameters:

neuron_params (dict, optional) – a dictionary of Adaptive LIF neuron parameter. Defaults to None.
in_neurons (int) – number of input neurons.
out_neurons (int) – number of output neurons.
weight_scale (int, optional) – weight initialization scaling. Defaults to 1.
weight_norm (bool, optional) – flag to enable weight normalization. Defaults to False.
pre_hook_fx (optional) – a function pointer or lambda that is applied to synaptic weights before synaptic operation. None means no transformation. Defaults to None.
delay (bool, optional) – flag to enable axonal delay. Defaults to False.
delay_shift (bool, optional) – flag to simulate spike propagation delay from one layer to next. Defaults to True.
mask (bool array, optional) – boolean synapse mask that only enables relevant synapses. None means no masking is applied. Defaults to None.
count_log (bool, optional) – flag to return event count log. If True, an additional value of average event rate is returned. Defaults to False.

class lava.lib.dl.slayer.block.alif.Flatten(*args, **kwargs)

Bases: AbstractFlatten

Adaptive LIF flatten block class. This should never be instantiated on its own.

Parameters:: count_log (bool, optional) – flag to return event count log. If True, an additional value of average event rate is returned. Defaults to False.

class lava.lib.dl.slayer.block.alif.Input(*args, **kwargs)

Bases: AbstractALIF, AbstractInput

Adaptive LIF input block class. The block is 8 bit quantization ready.

Parameters:

neuron_params (dict, optional) – a dictionary of Adaptive LIF neuron parameter. Defaults to None.
weight (float, optional) – weight for affine transform of input. None means no weight scaling. Defaults to None.
bias (float, optional) – bias for affine transform of input. None means no bias shift. Defaults to None.
delay_shift (bool, optional) – flag to simulate spike propagation delay from one layer to next. Defaults to True.
count_log (bool, optional) – flag to return event count log. If True, an additional value of average event rate is returned. Defaults to False.

class lava.lib.dl.slayer.block.alif.KWTA(*args, **kwargs)

Bases: AbstractALIF, AbstractKWTA

Adaptive LIF K-Winner-Takes-All block class. This should never be instantiated on its own. The formulation is described as below:

s_\text{out}[t] = f_s\left(\mathbf{W}\,s_\text{in}[t] + \mathbf{R}\,s_{out}[t-1] + \alpha\,(N-2K)\right)\\ \mathbf{R} = \begin{bmatrix} a &-1 &\cdots &-1\\ -1 & a &\cdots &-1\\ \vdots &\vdots &\ddots &\vdots\\ -1 &-1 &\cdots & a \end{bmatrix},\qquad |a| < 1

Parameters:

neuron_params (dict, optional) – a dictionary of Adaptive LIF neuron parameter. Defaults to None.
in_neurons (int) – number of input neurons.
out_neurons (int) – number of output neurons.
num_winners ([type]) – number of winners.
self_excitation (float, optional) – self excitation factor. Defaults to 0.5.
weight_scale (int, optional) – weight initialization scaling. Defaults to 1.
weight_norm (bool, optional) – flag to enable weight normalization. Defaults to False.
pre_hook_fx (optional) – a function pointer or lambda that is applied to synaptic weights before synaptic operation. None means no transformation. Defaults to None.
delay_shift (bool, optional) – flag to simulate spike propagation delay from one layer to next. Defaults to True.
requires_grad (bool, optional) – flag for learnable recurrent synapse. Defaults to True.
count_log (bool, optional) – flag to return event count log. If True, an additional value of average event rate is returned. Defaults to False.

class lava.lib.dl.slayer.block.alif.Pool(*args, **kwargs)

Bases: AbstractALIF, AbstractPool

Adaptive LIF input block class. The block is 8 bit quantization ready.

Parameters:

neuron_params (dict, optional) – a dictionary of Adaptive LIF neuron parameter. Defaults to None.
kernel_size (int or tuple of two ints) – size of pooling kernel.
stride (int or tuple of two ints, optional) – stride of pooling operation. Defaults to None.
padding (int or tuple of two ints, optional) – padding of pooling operation. Defaults to 0.
dilation (int or tuple of two ints, optional) – dilation of pooling kernel. Defaults to 1.
weight_scale (int, optional) – weight initialization scaling. Defaults to 1.
weight_norm (bool, optional) – flag to enable weight normalization. Defaults to False.
pre_hook_fx (optional) – a function pointer or lambda that is applied to synaptic weights before synaptic operation. None means no transformation. Defaults to None.
delay (bool, optional) – flag to enable axonal delay. Defaults to False.
delay_shift (bool, optional) – flag to simulate spike propagation delay from one layer to next. Defaults to True.
count_log (bool, optional) – flag to return event count log. If True, an additional value of average event rate is returned. Defaults to False.

class lava.lib.dl.slayer.block.alif.Recurrent(*args, **kwargs)

Bases: AbstractALIF, AbstractRecurrent

Adaptive LIF recurrent block class. This should never be instantiated on its own. The recurrent formulation is described below:

s_\text{out}[t] = f_s\left(\mathbf{W}\,s_\text{in}[t] + \mathbf{R}\,s_{out}[t-1]\right)

Parameters:

neuron_params (dict, optional) – a dictionary of Adaptive LIF neuron parameter.Defaults to None.
in_neurons (int) – number of input neurons.
out_neurons (int) – number of output neurons.
weight_scale (int, optional) – weight initialization scaling. Defaults to 1.
weight_norm (bool, optional) – flag to enable weight normalization. Defaults to False.
pre_hook_fx (optional) – a function pointer or lambda that is applied to synaptic weights before synaptic operation. None means no transformation. Defaults to None.
requires_grad (bool, optional) – flag for learnable recurrent synapse. Defaults to True.
delay (bool, optional) – flag to enable axonal delay. Defaults to False.
delay_shift (bool, optional) – flag to simulate spike propagation delay from one layer to next. Defaults to True.
count_log (bool, optional) – flag to return event count log. If True, an additional value of average event rate is returned. Defaults to False.

class lava.lib.dl.slayer.block.alif.TimeDecimation(*args, **kwargs)

Bases: AbstractTimeDecimation

Adaptive LIF time decimation block class. This should never be instantiated on its own.

Parameters:

factor (int, optional) – number of time units to decimate in a single bin. Must be in powers of 2.
count_log (bool, optional) – flag to return event count log. If True, an additional value of average event rate is returned. Defaults to False.

class lava.lib.dl.slayer.block.alif.Unpool(*args, **kwargs)

Bases: AbstractALIF, AbstractUnpool

Adaptive LIF Unpool block class. The block is 8 bit quantization ready.

Parameters:

neuron_params (dict, optional) – a dictionary of Adaptive LIF neuron parameter. Defaults to None.
kernel_size (int or tuple of two ints) – size of unpooling kernel.
stride (int or tuple of two ints, optional) – stride of unpooling operation. Defaults to None.
padding (int or tuple of two ints, optional) – padding of unpooling operation. Defaults to 0.
dilation (int or tuple of two ints, optional) – dilation of unpooling kernel. Defaults to 1.
weight_scale (int, optional) – weight initialization scaling. Defaults to 1.
weight_norm (bool, optional) – flag to enable weight normalization. Defaults to False.
pre_hook_fx (optional) – a function pointer or lambda that is applied to synaptic weights before synaptic operation. None means no transformation. Defaults to None.
delay (bool, optional) – flag to enable axonal delay. Defaults to False.
delay_shift (bool, optional) – flag to simulate spike propagation delay from one layer to next. Defaults to True.
count_log (bool, optional) – flag to return event count log. If True, an additional value of average event rate is returned. Defaults to False.

Adaptive Resonate and Fire (Ad R&F) Block

Adaptive Resonate and Fire - Phase Threshold layer

class lava.lib.dl.slayer.block.adrf.AbstractADRF(*args, **kwargs)

Bases: Module

Abstract Adaptive Resonate and Fire block. This should never be instantiated on its own.

class lava.lib.dl.slayer.block.adrf.Average(*args, **kwargs)

Bases: AbstractAverage

Adaptive Resonate & Fire average block class. This should never be instantiated on its own.

Parameters:

num_outputs (int) – number of output population groups.
count_log (bool, optional) – flag to return event count log. If True, an additional value of average event rate is returned. Defaults to False.

class lava.lib.dl.slayer.block.adrf.Conv(*args, **kwargs)

Bases: AbstractADRF, AbstractConv

Adaptive Resonate & Fire convolution block class. This should never be instantiated on its own.

Parameters:

neuron_params (dict, optional) – a dictionary of Adaptive RF neuron parameter. Defaults to None.
in_features (int) – number of input features.
out_features (int) – number of output features.
kernel_size (int) – kernel size.
stride (int or tuple of two ints, optional) – convolution stride. Defaults to 1.
padding (int or tuple of two ints, optional) – convolution padding. Defaults to 0.
dilation (int or tuple of two ints, optional) – convolution dilation. Defaults to 1.
groups (int, optional) – number of blocked connections. Defaults to 1.
weight_scale (int, optional) – weight initialization scaling. Defaults to 1.
weight_norm (bool, optional) – flag to enable weight normalization. Defaults to False.
pre_hook_fx (optional) – a function pointer or lambda that is applied to synaptic weights before synaptic operation. None means no transformation. Defaults to None.
delay (bool, optional) – flag to enable axonal delay. Defaults to False.
delay_shift (bool, optional) – flag to simulate spike propagation delay from one layer to next. Defaults to True.
count_log (bool, optional) – flag to return event count log. If True, an additional value of average event rate is returned. Defaults to False.

class lava.lib.dl.slayer.block.adrf.ConvT(*args, **kwargs)

Bases: AbstractADRF, AbstractConvT

Adaptive Resonate & Fire convolution Traspose block class. This should never be instantiated on its own.

Parameters:

neuron_params (dict, optional) – a dictionary of Adaptive RF neuron parameter. Defaults to None.
in_features (int) – number of input features.
out_features (int) – number of output features.
kernel_size (int) – kernel size.
stride (int or tuple of two ints, optional) – convolutionT stride. Defaults to 1.
padding (int or tuple of two ints, optional) – convolutionT padding. Defaults to 0.
dilation (int or tuple of two ints, optional) – convolutionT dilation. Defaults to 1.
groups (int, optional) – number of blocked connections. Defaults to 1.
weight_scale (int, optional) – weight initialization scaling. Defaults to 1.
weight_norm (bool, optional) – flag to enable weight normalization. Defaults to False.
pre_hook_fx (optional) – a function pointer or lambda that is applied to synaptic weights before synaptic operation. None means no transformation. Defaults to None.
delay (bool, optional) – flag to enable axonal delay. Defaults to False.
delay_shift (bool, optional) – flag to simulate spike propagation delay from one layer to next. Defaults to True.
count_log (bool, optional) – flag to return event count log. If True, an additional value of average event rate is returned. Defaults to False.

class lava.lib.dl.slayer.block.adrf.Dense(*args, **kwargs)

Bases: AbstractADRF, AbstractDense

Adaptive Resonate & Fire dense block class. The block is 8 bit quantization ready.

Parameters:

neuron_params (dict, optional) – a dictionary of Adaptive RF neuron parameter. Defaults to None.
in_neurons (int) – number of input neurons.
out_neurons (int) – number of output neurons.
weight_scale (int, optional) – weight initialization scaling. Defaults to 1.
weight_norm (bool, optional) – flag to enable weight normalization. Defaults to False.
pre_hook_fx (optional) – a function pointer or lambda that is applied to synaptic weights before synaptic operation. None means no transformation. Defaults to None.
delay (bool, optional) – flag to enable axonal delay. Defaults to False.
delay_shift (bool, optional) – flag to simulate spike propagation delay from one layer to next. Defaults to True.
mask (bool array, optional) – boolean synapse mask that only enables relevant synapses. None means no masking is applied. Defaults to None.
count_log (bool, optional) – flag to return event count log. If True, an additional value of average event rate is returned. Defaults to False.

class lava.lib.dl.slayer.block.adrf.Flatten(*args, **kwargs)

Bases: AbstractFlatten

Adaptive Resonate & Fire flatten block class. This should never be instantiated on its own.

Parameters:: count_log (bool, optional) – flag to return event count log. If True, an additional value of average event rate is returned. Defaults to False.

class lava.lib.dl.slayer.block.adrf.Input(*args, **kwargs)

Bases: AbstractADRF, AbstractInput

Adaptive Resonate & Fire input block class. The block is 8 bit quantization ready.

Parameters:

neuron_params (dict, optional) – a dictionary of Adaptive RF neuron parameter. Defaults to None.
weight (float, optional) – weight for affine transform of input. None means no weight scaling. Defaults to None.
bias (float, optional) – bias for affine transform of input. None means no bias shift. Defaults to None.
delay_shift (bool, optional) – flag to simulate spike propagation delay from one layer to next. Defaults to True.
count_log (bool, optional) – flag to return event count log. If True, an additional value of average event rate is returned. Defaults to False.

class lava.lib.dl.slayer.block.adrf.KWTA(*args, **kwargs)

Bases: AbstractADRF, AbstractKWTA

Adaptive Resonate & Fire K-Winner-Takes-All block class. This should never be instantiated on its own. The formulation is described as below:

s_\text{out}[t] = f_s\left(\mathbf{W}\,s_\text{in}[t] + \mathbf{R}\,s_{out}[t-1] + \alpha\,(N-2K)\right)\\ \mathbf{R} = \begin{bmatrix} a &-1 &\cdots &-1\\ -1 & a &\cdots &-1\\ \vdots &\vdots &\ddots &\vdots\\ -1 &-1 &\cdots & a \end{bmatrix},\qquad |a| < 1

Parameters:

neuron_params (dict, optional) – a dictionary of Adaptive RF neuron parameter. Defaults to None.
in_neurons (int) – number of input neurons.
out_neurons (int) – number of output neurons.
num_winners ([type]) – number of winners.
self_excitation (float, optional) – self excitation factor. Defaults to 0.5.
weight_scale (int, optional) – weight initialization scaling. Defaults to 1.
weight_norm (bool, optional) – flag to enable weight normalization. Defaults to False.
pre_hook_fx (optional) – a function pointer or lambda that is applied to synaptic weights before synaptic operation. None means no transformation. Defaults to None.
delay_shift (bool, optional) – flag to simulate spike propagation delay from one layer to next. Defaults to True.
requires_grad (bool, optional) – flag for learnable recurrent synapse. Defaults to True.
count_log (bool, optional) – flag to return event count log. If True, an additional value of average event rate is returned. Defaults to False.

class lava.lib.dl.slayer.block.adrf.Pool(*args, **kwargs)

Bases: AbstractADRF, AbstractPool

Adaptive Resonate & Fire input block class. The block is 8 bit quantization ready.

Parameters:

neuron_params (dict, optional) – a dictionary of Adaptive RF neuron parameter. Defaults to None.
kernel_size (int or tuple of two ints) – size of pooling kernel.
stride (int or tuple of two ints, optional) – stride of pooling operation. Defaults to None.
padding (int or tuple of two ints, optional) – padding of pooling operation. Defaults to 0.
dilation (int or tuple of two ints, optional) – dilation of pooling kernel. Defaults to 1.
weight_scale (int, optional) – weight initialization scaling. Defaults to 1.
weight_norm (bool, optional) – flag to enable weight normalization. Defaults to False.
pre_hook_fx (optional) – a function pointer or lambda that is applied to synaptic weights before synaptic operation. None means no transformation. Defaults to None.
delay (bool, optional) – flag to enable axonal delay. Defaults to False.
delay_shift (bool, optional) – flag to simulate spike propagation delay from one layer to next. Defaults to True.
count_log (bool, optional) – flag to return event count log. If True, an additional value of average event rate is returned. Defaults to False.

class lava.lib.dl.slayer.block.adrf.Recurrent(*args, **kwargs)

Bases: AbstractADRF, AbstractRecurrent

Adaptive Resonate & Fire recurrent block class. This should never be instantiated on its own. The recurrent formulation is described below:

s_\text{out}[t] = f_s\left(\mathbf{W}\,s_\text{in}[t] + \mathbf{R}\,s_{out}[t-1]\right)

Parameters:

neuron_params (dict, optional) – a dictionary of Adaptive RF neuron parameter.Defaults to None.
in_neurons (int) – number of input neurons.
out_neurons (int) – number of output neurons.
weight_scale (int, optional) – weight initialization scaling. Defaults to 1.
weight_norm (bool, optional) – flag to enable weight normalization. Defaults to False.
pre_hook_fx (optional) – a function pointer or lambda that is applied to synaptic weights before synaptic operation. None means no transformation. Defaults to None.
requires_grad (bool, optional) – flag for learnable recurrent synapse. Defaults to True.
delay (bool, optional) – flag to enable axonal delay. Defaults to False.
delay_shift (bool, optional) – flag to simulate spike propagation delay from one layer to next. Defaults to True.
count_log (bool, optional) – flag to return event count log. If True, an additional value of average event rate is returned. Defaults to False.

class lava.lib.dl.slayer.block.adrf.TimeDecimation(*args, **kwargs)

Bases: AbstractTimeDecimation

Adaptive Resonate & Fire time decimation block class. This should never be instantiated on its own.

Parameters:

factor (int, optional) – number of time units to decimate in a single bin. Must be in powers of 2.
count_log (bool, optional) – flag to return event count log. If True, an additional value of average event rate is returned. Defaults to False.

class lava.lib.dl.slayer.block.adrf.Unpool(*args, **kwargs)

Bases: AbstractADRF, AbstractUnpool

Adaptive Resonate & Fire Unpool block class. The block is 8 bit quantization ready.

Parameters:

neuron_params (dict, optional) – a dictionary of Adaptive RF neuron parameter. Defaults to None.
kernel_size (int or tuple of two ints) – size of unpooling kernel.
stride (int or tuple of two ints, optional) – stride of unpooling operation. Defaults to None.
padding (int or tuple of two ints, optional) – padding of unpooling operation. Defaults to 0.
dilation (int or tuple of two ints, optional) – dilation of unpooling kernel. Defaults to 1.
weight_scale (int, optional) – weight initialization scaling. Defaults to 1.
weight_norm (bool, optional) – flag to enable weight normalization. Defaults to False.
pre_hook_fx (optional) – a function pointer or lambda that is applied to synaptic weights before synaptic operation. None means no transformation. Defaults to None.
delay (bool, optional) – flag to enable axonal delay. Defaults to False.
delay_shift (bool, optional) – flag to simulate spike propagation delay from one layer to next. Defaults to True.
count_log (bool, optional) – flag to return event count log. If True, an additional value of average event rate is returned. Defaults to False.

Adaptive Resonate and Fire Izhikevich (Ad R&F Iz) Block

Adaptive Resonate and Fire - Izhikevich layer blocks

class lava.lib.dl.slayer.block.adrf_iz.AbstractADRFIZ(*args, **kwargs)

Bases: Module

Abstract Adaptive Resonate and Fire - Izhikevich block. This should never be instantiated on its own.

class lava.lib.dl.slayer.block.adrf_iz.Average(*args, **kwargs)

Bases: AbstractAverage

Adaptive Resonate & Fire Izhikevich average block class. This should never be instantiated on its own.

Parameters:

num_outputs (int) – number of output population groups.
count_log (bool, optional) – flag to return event count log. If True, an additional value of average event rate is returned. Defaults to False.

class lava.lib.dl.slayer.block.adrf_iz.Conv(*args, **kwargs)

Bases: AbstractADRFIZ, AbstractConv

Adaptive Resonate & Fire Izhikevich convolution block class. This should never be instantiated on its own.

Parameters:

neuron_params (dict, optional) – a dictionary of Adaptive RF-Izhikevich neuron parameter. Defaults to None.
in_features (int) – number of input features.
out_features (int) – number of output features.
kernel_size (int) – kernel size.
stride (int or tuple of two ints, optional) – convolution stride. Defaults to 1.
padding (int or tuple of two ints, optional) – convolution padding. Defaults to 0.
dilation (int or tuple of two ints, optional) – convolution dilation. Defaults to 1.
groups (int, optional) – number of blocked connections. Defaults to 1.
weight_scale (int, optional) – weight initialization scaling. Defaults to 1.
weight_norm (bool, optional) – flag to enable weight normalization. Defaults to False.
pre_hook_fx (optional) – a function pointer or lambda that is applied to synaptic weights before synaptic operation. None means no transformation. Defaults to None.
delay (bool, optional) – flag to enable axonal delay. Defaults to False.
delay_shift (bool, optional) – flag to simulate spike propagation delay from one layer to next. Defaults to True.
count_log (bool, optional) – flag to return event count log. If True, an additional value of average event rate is returned. Defaults to False.

class lava.lib.dl.slayer.block.adrf_iz.ConvT(*args, **kwargs)

Bases: AbstractADRFIZ, AbstractConvT

Adaptive Resonate & Fire Izhikevich convolution Traspose block class. This should never be instantiated on its own.

Parameters:

neuron_params (dict, optional) – a dictionary of Adaptive RF-Izhikevich neuron parameter. Defaults to None.
in_features (int) – number of input features.
out_features (int) – number of output features.
kernel_size (int) – kernel size.
stride (int or tuple of two ints, optional) – convolutionT stride. Defaults to 1.
padding (int or tuple of two ints, optional) – convolutionT padding. Defaults to 0.
dilation (int or tuple of two ints, optional) – convolutionT dilation. Defaults to 1.
groups (int, optional) – number of blocked connections. Defaults to 1.
weight_scale (int, optional) – weight initialization scaling. Defaults to 1.
weight_norm (bool, optional) – flag to enable weight normalization. Defaults to False.
pre_hook_fx (optional) – a function pointer or lambda that is applied to synaptic weights before synaptic operation. None means no transformation. Defaults to None.
delay (bool, optional) – flag to enable axonal delay. Defaults to False.
delay_shift (bool, optional) – flag to simulate spike propagation delay from one layer to next. Defaults to True.
count_log (bool, optional) – flag to return event count log. If True, an additional value of average event rate is returned. Defaults to False.

class lava.lib.dl.slayer.block.adrf_iz.Dense(*args, **kwargs)

Bases: AbstractADRFIZ, AbstractDense

Adaptive Resonate & Fire Izhikevich dense block class. The block is 8 bit quantization ready.

Parameters:

neuron_params (dict, optional) – a dictionary of Adaptive RF-Izhikevich neuron parameter. Defaults to None.
in_neurons (int) – number of input neurons.
out_neurons (int) – number of output neurons.
weight_scale (int, optional) – weight initialization scaling. Defaults to 1.
weight_norm (bool, optional) – flag to enable weight normalization. Defaults to False.
pre_hook_fx (optional) – a function pointer or lambda that is applied to synaptic weights before synaptic operation. None means no transformation. Defaults to None.
delay (bool, optional) – flag to enable axonal delay. Defaults to False.
delay_shift (bool, optional) – flag to simulate spike propagation delay from one layer to next. Defaults to True.
mask (bool array, optional) – boolean synapse mask that only enables relevant synapses. None means no masking is applied. Defaults to None.
count_log (bool, optional) – flag to return event count log. If True, an additional value of average event rate is returned. Defaults to False.

class lava.lib.dl.slayer.block.adrf_iz.Flatten(*args, **kwargs)

Bases: AbstractFlatten

Adaptive Resonate & Fire Izhikevich flatten block class. This should never be instantiated on its own.

Parameters:: count_log (bool, optional) – flag to return event count log. If True, an additional value of average event rate is returned. Defaults to False.

class lava.lib.dl.slayer.block.adrf_iz.Input(*args, **kwargs)

Bases: AbstractADRFIZ, AbstractInput

Adaptive Resonate & Fire Izhikevich input block class. The block is 8 bit quantization ready.

Parameters:

neuron_params (dict, optional) – a dictionary of Adaptive RF-Izhikevich neuron parameter. Defaults to None.
weight (float, optional) – weight for affine transform of input. None means no weight scaling. Defaults to None.
bias (float, optional) – bias for affine transform of input. None means no bias shift. Defaults to None.
delay_shift (bool, optional) – flag to simulate spike propagation delay from one layer to next. Defaults to True.
count_log (bool, optional) – flag to return event count log. If True, an additional value of average event rate is returned. Defaults to False.

class lava.lib.dl.slayer.block.adrf_iz.KWTA(*args, **kwargs)

Bases: AbstractADRFIZ, AbstractKWTA

Adaptive Resonate & Fire Izhikevich K-Winner-Takes-All block class. This should never be instantiated on its own. The formulation is described as below:

s_\text{out}[t] = f_s\left(\mathbf{W}\,s_\text{in}[t] + \mathbf{R}\,s_{out}[t-1] + \alpha\,(N-2K)\right)\\ \mathbf{R} = \begin{bmatrix} a &-1 &\cdots &-1\\ -1 & a &\cdots &-1\\ \vdots &\vdots &\ddots &\vdots\\ -1 &-1 &\cdots & a \end{bmatrix},\qquad |a| < 1

Parameters:

neuron_params (dict, optional) – a dictionary of Adaptive RF-Izhikevich neuron parameter. Defaults to None.
in_neurons (int) – number of input neurons.
out_neurons (int) – number of output neurons.
num_winners ([type]) – number of winners.
self_excitation (float, optional) – self excitation factor. Defaults to 0.5.
weight_scale (int, optional) – weight initialization scaling. Defaults to 1.
weight_norm (bool, optional) – flag to enable weight normalization. Defaults to False.
pre_hook_fx (optional) – a function pointer or lambda that is applied to synaptic weights before synaptic operation. None means no transformation. Defaults to None.
delay_shift (bool, optional) – flag to simulate spike propagation delay from one layer to next. Defaults to True.
requires_grad (bool, optional) – flag for learnable recurrent synapse. Defaults to True.
count_log (bool, optional) – flag to return event count log. If True, an additional value of average event rate is returned. Defaults to False.

class lava.lib.dl.slayer.block.adrf_iz.Pool(*args, **kwargs)

Bases: AbstractADRFIZ, AbstractPool

Adaptive Resonate & Fire Izhikevich input block class. The block is 8 bit quantization ready.

Parameters:

neuron_params (dict, optional) – a dictionary of Adaptive RF-Izhikevich neuron parameter. Defaults to None.
kernel_size (int or tuple of two ints) – size of pooling kernel.
stride (int or tuple of two ints, optional) – stride of pooling operation. Defaults to None.
padding (int or tuple of two ints, optional) – padding of pooling operation. Defaults to 0.
dilation (int or tuple of two ints, optional) – dilation of pooling kernel. Defaults to 1.
weight_scale (int, optional) – weight initialization scaling. Defaults to 1.
weight_norm (bool, optional) – flag to enable weight normalization. Defaults to False.
pre_hook_fx (optional) – a function pointer or lambda that is applied to synaptic weights before synaptic operation. None means no transformation. Defaults to None.
delay (bool, optional) – flag to enable axonal delay. Defaults to False.
delay_shift (bool, optional) – flag to simulate spike propagation delay from one layer to next. Defaults to True.
count_log (bool, optional) – flag to return event count log. If True, an additional value of average event rate is returned. Defaults to False.

class lava.lib.dl.slayer.block.adrf_iz.Recurrent(*args, **kwargs)

Bases: AbstractADRFIZ, AbstractRecurrent

Adaptive Resonate & Fire Izhikevich recurrent block class. This should never be instantiated on its own. The recurrent formulation is described below:

s_\text{out}[t] = f_s\left(\mathbf{W}\,s_\text{in}[t] + \mathbf{R}\,s_{out}[t-1]\right)

Parameters:

neuron_params (dict, optional) – a dictionary of Adaptive RF-Izhikevich neuron parameter.Defaults to None.
in_neurons (int) – number of input neurons.
out_neurons (int) – number of output neurons.
weight_scale (int, optional) – weight initialization scaling. Defaults to 1.
weight_norm (bool, optional) – flag to enable weight normalization. Defaults to False.
pre_hook_fx (optional) – a function pointer or lambda that is applied to synaptic weights before synaptic operation. None means no transformation. Defaults to None.
requires_grad (bool, optional) – flag for learnable recurrent synapse. Defaults to True.
delay (bool, optional) – flag to enable axonal delay. Defaults to False.
delay_shift (bool, optional) – flag to simulate spike propagation delay from one layer to next. Defaults to True.
count_log (bool, optional) – flag to return event count log. If True, an additional value of average event rate is returned. Defaults to False.

class lava.lib.dl.slayer.block.adrf_iz.TimeDecimation(*args, **kwargs)

Bases: AbstractTimeDecimation

Adaptive Resonate & Fire Izhikevich time decimation block class. This should never be instantiated on its own.

Parameters:

factor (int, optional) – number of time units to decimate in a single bin. Must be in powers of 2.
count_log (bool, optional) – flag to return event count log. If True, an additional value of average event rate is returned. Defaults to False.

class lava.lib.dl.slayer.block.adrf_iz.Unpool(*args, **kwargs)

Bases: AbstractADRFIZ, AbstractUnpool

Adaptive Resonate & Fire Izhikevich Unpool block class. The block is 8 bit quantization ready.

Parameters:

neuron_params (dict, optional) – a dictionary of Adaptive RF-Izhikevich neuron parameter. Defaults to None.
kernel_size (int or tuple of two ints) – size of unpooling kernel.
stride (int or tuple of two ints, optional) – stride of unpooling operation. Defaults to None.
padding (int or tuple of two ints, optional) – padding of unpooling operation. Defaults to 0.
dilation (int or tuple of two ints, optional) – dilation of unpooling kernel. Defaults to 1.
weight_scale (int, optional) – weight initialization scaling. Defaults to 1.
weight_norm (bool, optional) – flag to enable weight normalization. Defaults to False.
pre_hook_fx (optional) – a function pointer or lambda that is applied to synaptic weights before synaptic operation. None means no transformation. Defaults to None.
delay (bool, optional) – flag to enable axonal delay. Defaults to False.
delay_shift (bool, optional) – flag to simulate spike propagation delay from one layer to next. Defaults to True.
count_log (bool, optional) – flag to return event count log. If True, an additional value of average event rate is returned. Defaults to False.

Sigma Delta Block (SDN)

Sigma Delta layer blocks.

class lava.lib.dl.slayer.block.sigma_delta.AbstractSDRelu(*args, **kwargs)

Bases: Module

Abstract Sigma Delta block class. This should never be instantiated on it’s own.

class lava.lib.dl.slayer.block.sigma_delta.Average(*args, **kwargs)

Bases: AbstractAverage

Sigma Delta average block class. This should never be instantiated on its own.

Parameters:

num_outputs (int) – number of output population groups.
count_log (bool, optional) – flag to return event count log. If True, an additional value of average event rate is returned. Defaults to False.

class lava.lib.dl.slayer.block.sigma_delta.Conv(*args, **kwargs)

Bases: AbstractSDRelu, AbstractConv

Sigma Delta convolution block class. This should never be instantiated on its own.

Parameters:

neuron_params (dict, optional) – a dictionary of Sigma Delta neuron parameter. Defaults to None.
in_features (int) – number of input features.
out_features (int) – number of output features.
kernel_size (int) – kernel size.
stride (int or tuple of two ints, optional) – convolution stride. Defaults to 1.
padding (int or tuple of two ints, optional) – convolution padding. Defaults to 0.
dilation (int or tuple of two ints, optional) – convolution dilation. Defaults to 1.
groups (int, optional) – number of blocked connections. Defaults to 1.
weight_scale (int, optional) – weight initialization scaling. Defaults to 1.
weight_norm (bool, optional) – flag to enable weight normalization. Defaults to False.
pre_hook_fx (optional) – a function pointer or lambda that is applied to synaptic weights before synaptic operation. None means no transformation. Defaults to None.
delay (bool, optional) – flag to enable axonal delay. Defaults to False.
delay_shift (bool, optional) – flag to simulate spike propagation delay from one layer to next. Defaults to True.
count_log (bool, optional) – flag to return event count log. If True, an additional value of average event rate is returned. Defaults to False.

class lava.lib.dl.slayer.block.sigma_delta.ConvT(*args, **kwargs)

Bases: AbstractSDRelu, AbstractConvT

Sigma Delta convolution Traspose block class. This should never be instantiated on its own.

Parameters:

neuron_params (dict, optional) – a dictionary of Sigma Delta neuron parameter. Defaults to None.
in_features (int) – number of input features.
out_features (int) – number of output features.
kernel_size (int) – kernel size.
stride (int or tuple of two ints, optional) – convolutionT stride. Defaults to 1.
padding (int or tuple of two ints, optional) – convolutionT padding. Defaults to 0.
dilation (int or tuple of two ints, optional) – convolutionT dilation. Defaults to 1.
groups (int, optional) – number of blocked connections. Defaults to 1.
weight_scale (int, optional) – weight initialization scaling. Defaults to 1.
weight_norm (bool, optional) – flag to enable weight normalization. Defaults to False.
pre_hook_fx (optional) – a function pointer or lambda that is applied to synaptic weights before synaptic operation. None means no transformation. Defaults to None.
delay (bool, optional) – flag to enable axonal delay. Defaults to False.
delay_shift (bool, optional) – flag to simulate spike propagation delay from one layer to next. Defaults to True.
count_log (bool, optional) – flag to return event count log. If True, an additional value of average event rate is returned. Defaults to False.

class lava.lib.dl.slayer.block.sigma_delta.Dense(*args, **kwargs)

Bases: AbstractSDRelu, AbstractDense

Sigma Delta dense block class. The block is 8 bit quantization ready.

Parameters:

neuron_params (dict, optional) – a dictionary of Sigma Delta neuron parameter. Defaults to None.
in_neurons (int) – number of input neurons.
out_neurons (int) – number of output neurons.
weight_scale (int, optional) – weight initialization scaling. Defaults to 1.
weight_norm (bool, optional) – flag to enable weight normalization. Defaults to False.
pre_hook_fx (optional) – a function pointer or lambda that is applied to synaptic weights before synaptic operation. None means no transformation. Defaults to None.
delay (bool, optional) – flag to enable axonal delay. Defaults to False.
delay_shift (bool, optional) – flag to simulate spike propagation delay from one layer to next. Defaults to True.
mask (bool array, optional) – boolean synapse mask that only enables relevant synapses. None means no masking is applied. Defaults to None.
count_log (bool, optional) – flag to return event count log. If True, an additional value of average event rate is returned. Defaults to False.

class lava.lib.dl.slayer.block.sigma_delta.Flatten(*args, **kwargs)

Bases: AbstractFlatten

Sigma Delta flatten block class. This should never be instantiated on its own.

Parameters:: count_log (bool, optional) – flag to return event count log. If True, an additional value of average event rate is returned. Defaults to False.

class lava.lib.dl.slayer.block.sigma_delta.Input(*args, **kwargs)

Bases: AbstractSDRelu, AbstractInput

Sigma Delta input block class. The block is 8 bit quantization ready.

Parameters:

neuron_params (dict, optional) – a dictionary of Sigma Delta neuron parameter. Defaults to None.
weight (float, optional) – weight for affine transform of input. None means no weight scaling. Defaults to None.
bias (float, optional) – bias for affine transform of input. None means no bias shift. Defaults to None.
delay_shift (bool, optional) – flag to simulate spike propagation delay from one layer to next. Defaults to True.
count_log (bool, optional) – flag to return event count log. If True, an additional value of average event rate is returned. Defaults to False.

forward(x)

class lava.lib.dl.slayer.block.sigma_delta.Output(neuron_params, in_neurons, out_neurons, weight_scale=1, weight_norm=False, count_log=False)

Bases: AbstractSDRelu

Sigma Delta output block class. The block is 8 bit quantization ready.

Parameters:

neuron_params (dict) – a dictionary of sigma delta neuron parameters.
in_neurons (int) – number of input neurons.
out_neurons (int) – number of output neurons.
weight_scale (int, optional) – weight initialization scaling. Defaults to 1.
weight_norm (bool, optional) – flag to enable weight normalization. Defaults to False.
count_log (bool, optional) – flag to return event count log. If True, an additional value of average event rate is returned. Defaults to False.

export_hdf5(handle)

Hdf5 export method for the block.

Parameters:: handle (file handle) – hdf5 handle to export block description.

forward(x)

property shape: Shape of the block.

class lava.lib.dl.slayer.block.sigma_delta.Pool(*args, **kwargs)

Bases: AbstractSDRelu, AbstractPool

Sigma Delta input block class. The block is 8 bit quantization ready.

Parameters:

neuron_params (dict, optional) – a dictionary of Sigma Delta neuron parameter. Defaults to None.
kernel_size (int or tuple of two ints) – size of pooling kernel.
stride (int or tuple of two ints, optional) – stride of pooling operation. Defaults to None.
padding (int or tuple of two ints, optional) – padding of pooling operation. Defaults to 0.
dilation (int or tuple of two ints, optional) – dilation of pooling kernel. Defaults to 1.
weight_scale (int, optional) – weight initialization scaling. Defaults to 1.
weight_norm (bool, optional) – flag to enable weight normalization. Defaults to False.
pre_hook_fx (optional) – a function pointer or lambda that is applied to synaptic weights before synaptic operation. None means no transformation. Defaults to None.
delay (bool, optional) – flag to enable axonal delay. Defaults to False.
delay_shift (bool, optional) – flag to simulate spike propagation delay from one layer to next. Defaults to True.
count_log (bool, optional) – flag to return event count log. If True, an additional value of average event rate is returned. Defaults to False.

class lava.lib.dl.slayer.block.sigma_delta.Unpool(*args, **kwargs)

Bases: AbstractSDRelu, AbstractUnpool

Sigma Delta Unpool block class. The block is 8 bit quantization ready.

Parameters:

neuron_params (dict, optional) – a dictionary of Sigma Delta neuron parameter. Defaults to None.
kernel_size (int or tuple of two ints) – size of unpooling kernel.
stride (int or tuple of two ints, optional) – stride of unpooling operation. Defaults to None.
padding (int or tuple of two ints, optional) – padding of unpooling operation. Defaults to 0.
dilation (int or tuple of two ints, optional) – dilation of unpooling kernel. Defaults to 1.
weight_scale (int, optional) – weight initialization scaling. Defaults to 1.
weight_norm (bool, optional) – flag to enable weight normalization. Defaults to False.
pre_hook_fx (optional) – a function pointer or lambda that is applied to synaptic weights before synaptic operation. None means no transformation. Defaults to None.
delay (bool, optional) – flag to enable axonal delay. Defaults to False.
delay_shift (bool, optional) – flag to simulate spike propagation delay from one layer to next. Defaults to True.
count_log (bool, optional) – flag to return event count log. If True, an additional value of average event rate is returned. Defaults to False.

Block Module

Abstract Block

CUrrent BAsed Leaky Integrate and Fire (CUBA) Block

Resonate and Fire (R&F) Block

Resonate and Fire Izhikevich (R&F Iz) Block

Adaptive Leaky Integrate and Fire (ALIF) Block

Adaptive Resonate and Fire (Ad R&F) Block

Adaptive Resonate and Fire Izhikevich (Ad R&F Iz) Block

Sigma Delta Block (SDN)

Module contents