qualia_plugin_snn.learningmodel.pytorch.layers.CustomNode module

ATIF-u: spiking neuron with learnable quantization steps.

Author: Andrea Castagetti <Andrea.CASTAGNETTI@univ-cotedazur.fr>

qualia_plugin_snn.learningmodel.pytorch.layers.CustomNode.heaviside(x: Tensor) → Tensor[source]

Heaviside function.

Parameters:: x (Tensor) – Input tensor
Returns:: Boolean tensor of the same dimension as x where each element is True if the element in x is greater than or equal to 0, False otherwise
Return type:: Tensor

class qualia_plugin_snn.learningmodel.pytorch.layers.CustomNode.SpikeFunctionSigmoid[source]

Bases: Function

Spike functions with surrogate bkw gradient.

static forward(ctx: FunctionCtx, *args: Tensor, **_: Any) → Tensor[source]

Forward of heaviside() function.

Parameters:

ctx (FunctionCtx) – A context object used to save tensors for backward()
args (Tensor) – Tuple of 2 tensors for x and alpha, respectively, saved in ctx for backward pass
_ (Any) – Unused

Returns:

Tensor of heaviside() applied over x.

Return type:

Tensor

static backward(ctx: Function, *grad_outputs: Tensor) → tuple[Tensor | None, None][source]

Backward pass of surrogate gradient using torch.Tensor.sigmoid_() function.

Parameters:

ctx (Function) – Context to restore the x and alpha tensors from
grad_outputs (Tensor) – Output tensor from the computation of the forward() pass

Returns:

A tuple of Tensor and None with the first element being the computed gradient for x or None if there is no gradient to compute and the second element a placeholder for the gradient of alpha.

Return type:

tuple[Tensor | None, None]

classmethod apply(*args: Tensor, **kwargs: Any) → Tensor[source]

Apply heaviside activation with sigmoid surrogate gradient.

Parameters:

args (Tensor) – Input tensor
kwargs (Any) – Unused

Returns:

Output tensor

Return type:

Tensor

class qualia_plugin_snn.learningmodel.pytorch.layers.CustomNode.IFSRL[source]

Bases: Module

IFSRL: Integrate and Fire soft-reset with learnable Vth and activation scaling.

__init__(v_threshold: float = 1.0, vth_init_l: float = 0.8, vth_init_h: float = 1.0, alpha: float = 1.0, device: str = 'cpu') → None[source]

Construct IFSRL.

Parameters:

v_threshold (float) – Factor to apply to the uniform initialization bounds
vth_init_l (float) – Lower bound for uniform initialization of threshold Tensor
vth_init_h (float) – Higher bound for uniform initialization of threshold Tensor
alpha (float) – Sigmoig surrogate scale factor
device (str) – Device to run the computation on

Return type:

None

v: torch.Tensor

get_coeffs() → Tensor[source]

Return the Tensor of threshold vp_th.

Returns:: Tensor of threshold vp_th
Return type:: Tensor

set_coeffs(vp_th: Tensor) → None[source]

Replace the Tensor of threshold vp_th.

Parameters:: vp_th (Tensor) – New Tensor of threshold to replace vp_th
Return type:: None

reset() → None[source]

Reset potential to 0.

Return type:: None

ifsrl_fn(x: Tensor) → Tensor[source]

Integrate-and-Fire soft-reset neuron with learnable threshold.

Parameters:: x (Tensor) – Input tensor
Returns:: Output tensor
Return type:: Tensor

forward(input: Tensor) → Tensor[source]

Forward of ifsrl_fn().

Parameters:: input (Tensor) – Input tensor
Returns:: Output tensor
Return type:: Tensor

training: bool

class qualia_plugin_snn.learningmodel.pytorch.layers.CustomNode.ATIF[source]

Bases: BaseNode

IFSRLSJ: Integrate and Fire soft-reset with learnable Vth and activation scaling, based on spikingjelly.

__init__(v_threshold: float = 1.0, vth_init_l: float = 0.8, vth_init_h: float = 1.0, alpha: float = 1.0, device: str = 'cpu') → None[source]

Construct ATIF.

Parameters:

v_threshold (float) – Factor to apply to the uniform initialization bounds
vth_init_l (float) – Lower bound for uniform initialization of threshold Tensor
vth_init_h (float) – Higher bound for uniform initialization of threshold Tensor
alpha (float) – Sigmoig surrogate scale factor
device (str) – Device to run the computation on

Return type: