hxtorch.spiking.functional
Modules
Adaptive exponential leaky-integrate and fire neurons |
|
Custom BatchDropout function |
|
Leaky-integrate neurons |
|
Leaky-integrate and fire neurons |
|
|
Wrap linear to allow signature inspection |
Refractory update for neurons with refractory behaviour |
|
Define different input spike sources |
|
Surrograte gradient for SuperSpike. |
|
|
Selection of the used threshold function. :param input: Input tensor to threshold function. :param method: The string indicator of the the threshold function. Currently supported: ‘super_spike’. :param alpha: Parameter controlling the slope of the surrogate derivative in case of ‘superspike’. :return: Returns the tensor of the threshold function. |
Autograd function to ‘unterjubel’ (german for ‘inject’) hardware observables and allow correct gradient back-propagation. |
Classes
|
|
|
Define gradient using adjoint code (EventProp) from norse |
|
Synapse function for proper gradient transport when using EventPropLIF. |
|
Define Surrogate Gradient ‘SuperSpike’ (negative side of Fast Sigmoid) See: https://arxiv.org/abs/1705.11146 |
Functions
-
hxtorch.spiking.functional.batch_dropout(input: torch.Tensor, mask: torch.Tensor) → torch.Tensor Applies a dropout mask to a batch of inputs.
- Parameters
input – The input tensor to apply dropout to.
mask – The dropout mask. Entires in the mask which are False will disable their corresponding entry in input.
- Returns
The input tensor with dropout mask applied.
-
hxtorch.spiking.functional.cuba_aelif_integration(input: Union[Tuple[torch.Tensor], torch.Tensor], *, leak: Union[torch.Tensor, float, int], reset: Union[torch.Tensor, float, int], threshold: Union[torch.Tensor, float, int], tau_syn: Union[torch.Tensor, float, int], c_mem: Union[torch.Tensor, float, int], g_l: Union[torch.Tensor, float, int], refractory_time: Union[torch.Tensor, float, int], method: str, alpha: float, exp_slope: Union[torch.Tensor, float, int], exp_threshold: Union[torch.Tensor, float, int], subthreshold_adaptation_strength: Union[torch.Tensor, float, int], spike_triggered_adaptation_increment: Union[torch.Tensor, float, int], tau_adap: Union[torch.Tensor, float, int], hw_data: Optional[Tuple[Optional[torch.Tensor, None]], None] = None, dt: float = 1e-06, leaky: bool = True, fire: bool = True, refractory: bool = False, exponential: bool = False, subthreshold_adaptation: bool = False, spike_triggered_adaptation: bool = False, integration_step_code: str) Adaptive exponential leaky-integrate and fire neuron integration for realization of AdEx neurons with exponential synapses. Certain terms of the differential equations of the membrane voltage v and the adaptation current w can be disabled or enabled via flags.
- If all flags are set, it integrates according to:
i^{t+1} = i^t * (1 - dt / au_{syn}) + x^t v^{t+1} = dt / c_{mem} * (g_l * (v_l - v^t + T * exp((v^t - v_T) / T))
i^t - w^t) + v^t
z^{t+1} = 1 if v^{t+1} > params.threshold w^{t+1} = w^t + dt / au_{adap} * (a * (v^{t+1} - v_l) - w^t)
b * z^{t+1}
v^{t+1} = params.reset if z^{t+1} == 1
Assumes i^0, v^0 = v_leak, if leak term is enabled, else v^0 = 0 and w^0 = 0. :note: One dt synaptic delay between input and output
- Parameters
input – torch.Tensor holding ‘graded_spikes’ in shape (batch, time, neurons) or tuple which holds one of such tensors for each input synapse.
leak – The leak voltage.
reset – The reset voltage.
threshold – The threshold voltage.
tau_syn – The synaptic time constant.
c_mem – The membrane capacitance.
g_l – The leak conductance.
refractory_time – The refractory time constant.
method – The method used for the surrogate gradient, e.g., ‘superspike’.
alpha – The slope of the surrogate gradient in case of ‘superspike’.
exp_slope – The exponential slope.
exp_threshold – The exponential threshold.
subthreshold_adaptation_strength – The subthreshold adaptation strength.
spike_triggered_adaptation_increment – The spike-triggered adaptation increment.
tau_adap – The adaptive time constant.
hw_data – An optional tuple holding optional hardware observables in the order (spikes, membrane_cadc, membrane_madc).
dt – Integration step width.
leaky – Flag that enables / disables the leak term when set to true / false
fire – Flag that enables / disables firing behaviour when set to true / false.
refractory – Flag used to omit the execution of the refractory update in case the refractory time is set to zero.
exponential – Flag that enables / disables the exponential term in the differential equation for the membrane potential when set to true / false.
subthreshold_adaptation – Flag that enables / disables the subthreshold adaptation term in the differential equation of the adaptation when set to true / false.
spike_triggered_adaptation – Flag that enables / disables spike-triggered adaptation when set to true / false.
- Returns
Returns tuple holding tensors with spikes, membrane traces, adaptation current and synaptic current. Tensors are of shape (time, batch, neurons).
-
hxtorch.spiking.functional.exp_cuba_li_integration(input: torch.Tensor, *, leak: torch.Tensor, tau_syn_exp: torch.Tensor, tau_mem_exp: torch.Tensor, hw_data: Optional[torch.Tensor, None] = None) → torch.Tensor
-
hxtorch.spiking.functional.exp_cuba_lif_integration(input: torch.Tensor, *, leak: torch.Tensor, reset: torch.Tensor, threshold: torch.Tensor, tau_syn_exp: torch.Tensor, tau_mem_exp: torch.Tensor, method: torch.Tensor, alpha: torch.Tensor, hw_data: Optional[torch.Tensor, None] = None) → Tuple[torch.Tensor, …]
-
hxtorch.spiking.functional.input_neuron(input: torch.Tensor, hw_data: Optional[torch.Tensor, None] = None) → types.Handle_current_membrane_cadc_membrane_madc_spikes Input neuron, forwards spikes without modification in non-hardware runs but injects loop-back recorded spikes if available.
- Parameters
input – Input spike tensor.
hw_data – Loop-back spikes, if available.
- Returns
Returns the input spike tensor.
-
hxtorch.spiking.functional.linear(input: torch.Tensor, weight: torch.nn.parameter.Parameter, bias: torch.nn.parameter.Parameter = None) → torch.Tensor Wrap linear to allow signature inspection
-
hxtorch.spiking.functional.linear_exponential_clamp(inputs: torch.Tensor, weight: torch.nn.parameter.Parameter, bias: torch.nn.parameter.Parameter = None, cap: float = 1.5, start_weight: float = 61.0, quantize: bool = False) → torch.Tensor Clamps the weights with an exponential roll-off towards saturation.
- Parameters
input – The input neuron tensor holding spikes to be multiplied with the params tensor weight.
weight – Weight Tensor to be clamped.
bias – The bias of the linear operation.
cap – Upper resp. -1 * lower boundary of the weights. Choose this value to be 1 / weight_scaling (see hxtorch.spiking.Synapse) to saturate the software weights where theirs scaled values saturate on hardware.
start_weight – Indicating at which hardware-weight the roll off begins. Has to be in range (0, 63).
quantize – If true, the weights are rounded to multiples of cap / 63 to match the discrete hardware representation.
- Returns
Clamped weights and possibly rounded weights
-
hxtorch.spiking.functional.linear_sparse(input: torch.Tensor, weight: torch.nn.parameter.Parameter, connections: torch.Tensor = None, bias: torch.nn.parameter.Parameter = None) → torch.Tensor Wrap linear to allow signature inspection. Disable inactive connections in weight tensor.
- Parameters
input – The input neuron tensor holding spikes to be multiplied with the params tensor weight.
weight – The weight parameter tensor. This tensor is expected to be dense since pytorch, see issue: 4039.
bias – The bias of the linear operation.
connections – A dense boolean connection mask indicating active connections. If None, the weight tensor remains untouched.
-
hxtorch.spiking.functional.threshold(input: torch.Tensor, method: str, alpha: float) → torch.Tensor Selection of the used threshold function. :param input: Input tensor to threshold function. :param method: The string indicator of the the threshold function.
Currently supported: ‘super_spike’.
- Parameters
alpha – Parameter controlling the slope of the surrogate derivative in case of ‘superspike’.
- Returns
Returns the tensor of the threshold function.