qualia_plugin_snn.learningmodel.pytorch.SCNN module

Contains the template for a convolutional spiking neural network.

class qualia_plugin_snn.learningmodel.pytorch.SCNN.SCNN

Bases: SNN

Convolutional spiking neural network template.

Similar to qualia_core.learningmodel.pytorch.CNN.CNN but with spiking neuron activation layers (e.g., IF) instead of torch.nn.ReLU.

Example TOML configuration for a 2D SVGG16 over 4 timesteps with soft-reset multi-step IF based on the SCNN template:

[[model]]
name = "VGG16"
params.filters      = [ 64, 64, 128, 128, 256, 256, 256, 512, 512, 512, 512, 512, 512]
params.kernel_sizes = [  3,  3,   3,   3,   3,   3,   3,   3,   3,   3,   3,   3,   3]
params.paddings     = [  1,  1,   1,   1,   1,   1,   1,   1,   1,   1,   1,   1,   1]
params.strides      = [  1,  1,   1,   1,   1,   1,   1,   1,   1,   1,   1,   1,   1]
params.pool_sizes   = [  0,  2,   0,   2,   0,   0,   2,   0,   0,   2,   0,   0,   2]
params.dropouts     = [  0,  0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0,   0, 0, 0]
params.fc_units     = [4096, 4096]
params.batch_norm   = true
params.timesteps    = 4
params.dims         = 2
params.neuron.kind                = 'IFNode'
params.neuron.params.v_reset      = false # Soft reset
params.neuron.params.v_threshold  = 1.0
params.neuron.params.detach_reset = true
params.neuron.params.step_mode    = 'm' # Multi-step mode, make sure to use SpikingJellyMultiStep learningframework
params.neuron.params.backend      = 'cupy'

__init__(input_shape: tuple[int, ...], output_shape: tuple[int, ...], filters: list[int], kernel_sizes: list[int], paddings: list[int], strides: list[int], dropouts: float | list[float], pool_sizes: list[int], fc_units: list[int], batch_norm: bool = False, prepool: int | list[int] = 1, postpool: int | list[int] = 1, neuron: RecursiveConfigDict | None = None, timesteps: int = 4, gsp: bool = False, dims: int = 1) → None

Construct SCNN.

Parameters:

• input_shape (tuple[int, ...]) – Input shape

• output_shape (tuple[int, ...]) – Output shape

• filters (list[int]) – List of out_channels for each Conv layer, also defines the number of Conv layers

• kernel_sizes (list[int]) – List of kernel_size for each Conv layer, must of the same size as filters

• paddings (list[int]) – List of padding for each Conv layer, must of the same size as filters

• strides (list[int]) – List of stride for each Conv layer, must of the same size as filters

• dropouts (float | list[float]) – List of Dropout layer p to apply after each Conv or Linear layer, must be of the same size as filters + fc_units, no layer added if element is 0

• pool_sizes (list[int]) – List of MaxPool layer kernel_size to apply after each Conv layer, must be of the same size as filters, no layer added if element is 0

• fc_units (list[int]) – List of torch.nn.Linear layer out_features to add at the end of the network, no layer added if empty

• batch_norm (bool) – If True, add a BatchNorm layer after each Conv layer, otherwise no layer added

• prepool (int | list[int]) – AvgPool layer kernel_size to add at the beginning of the network, no layer added if 0

• postpool (int | list[int]) – AvgPool layer kernel_size to add after all Conv layers, no layer added if 0

• neuron (RecursiveConfigDict | None) – Spiking neuron configuration, see qualia_plugin_snn.learningmodel.pytorch.SNN.SNN.__init__()

• timesteps (int) – Number of timesteps

• gsp (bool) – If True, a single GlobalSumPool layer is added instead of Linear layers

• dims (int) – Either 1 or 2 for 1D or 2D convolutional network.

Return type:

None

layers: nn.ModuleDict

List of sequential layers of the SCNN model

timesteps: int

Number of timesteps

input_shape: tuple[int, ...]

output_shape: tuple[int, ...]

training: bool

forward(input: Tensor) → Tensor

Forward calls each of the SCNN layers sequentially.

Parameters:

input (Tensor) – Input tensor

Returns:

Output tensor

Return type:

Tensor