Expand description
Wire-accurate NIR computational node types.
The closed NirNode enum mirrors upstream HDF5 type strings exactly
(CubaLIF, Conv2d, SumPool2d, I, …). Do not invent marketing
aliases (CurrLIF, Convolution, Integrator) — those break
interoperability with Python NIR.
Field names use snake_case matching the neuromorphs/NIR Python dataclasses.
Numeric array parameters are Tensor values (Python: numpy.ndarray).
Structs§
- Affine
- Affine map
y = W x + b. - AvgPool2d
- 2-D average pooling.
- Conv1d
- 1-D convolution.
- Conv2d
- 2-D convolution.
- CubaLi
- Current-based leaky integrator (
CubaLI). - CubaLif
- Current-based leaky integrate-and-fire (
CubaLIF). - Delay
- Pure delay
y(t) = x(t − τ). - Flatten
- Flatten a contiguous range of dimensions.
- I
- Integrator neuron (
I):dv/dt = R I. - If
- Integrate-and-fire neuron (
IF). - Input
- Input port: virtual node feeding data into the graph.
- Li
- Leaky integrator (
LI). - Lif
- Leaky integrate-and-fire (
LIF). - Linear
- Linear map without bias
y = W x. - Output
- Output port: virtual node collecting graph results.
- Scale
- Elementwise scale
y = x ⊙ s. - SumPool2d
- 2-D sum pooling.
- Threshold
- Heaviside threshold / surrogate step.