qualia_plugin_snn.learningframework package

Submodules

• qualia_plugin_snn.learningframework.SpikingJelly module
  • SpikingJelly
    • SpikingJelly.experimenttrackings
    • SpikingJelly.learningmodels
    • SpikingJelly.TrainerModule
      • SpikingJelly.TrainerModule.model
      • SpikingJelly.TrainerModule.forward()
    • SpikingJelly.trace_model()
• qualia_plugin_snn.learningframework.SpikingJellyMultiStep module
  • SpikingJellyMultiStep
    • SpikingJellyMultiStep.TrainerModule
      • SpikingJellyMultiStep.TrainerModule.forward()
• qualia_plugin_snn.learningframework.SpikingJellyMultiStepTimeStepsInData module
  • SpikingJellyMultiStepTimeStepsInData
    • SpikingJellyMultiStepTimeStepsInData.TrainerModule
      • SpikingJellyMultiStepTimeStepsInData.TrainerModule.forward()
• qualia_plugin_snn.learningframework.SpikingJellyTimeStepsInData module
  • SpikingJellyTimeStepsInData
    • SpikingJellyTimeStepsInData.TrainerModule
      • SpikingJellyTimeStepsInData.TrainerModule.apply_dataaugmentation()
      • SpikingJellyTimeStepsInData.TrainerModule.forward()
    • SpikingJellyTimeStepsInData.channels_last_to_channels_first()
    • SpikingJellyTimeStepsInData.channels_first_to_channels_last()
    • SpikingJellyTimeStepsInData.DatasetFromArray
      • SpikingJellyTimeStepsInData.DatasetFromArray.__init__()

Module contents

Qualia-Plugin-SNN learningframework package contains the SpikingJelly interface to use with the bundled learningmodels.

class qualia_plugin_snn.learningframework.SpikingJelly

Bases: PyTorch

SpikingJelly single-step LearningFramework implementation extending PyTorch.

qualia_core.learningframework.PyTorch.PyTorch.learningmodels are replaced by the Spiking Neural Networks from qualia_plugin_snn.learningmodel.pytorch

experimenttrackings: ModuleType

qualia_plugin_snn.experimenttracking.pytorch additional experimenttrackings for Spiking Neural Networks.

Merged to qualia_core.learningframework.PyTorch.PyTorch.experimenttrackings.

learningmodels: ModuleType

qualia_plugin_snn.learningmodel.pytorch additional learningmodels for Spiking Neural Networks.

class TrainerModule

Bases: TrainerModule

SpikingJelly single-step TrainerModule implementation extending PyTorch TrainerModule.

model: SNN

Spiking Neural Network model used by this TrainerModule

forward(x: Tensor) → Tensor

Forward pass for a Spiking Neural Network model with duplicated timesteps.

First calls SpikingJelly’s reset on the model to reset neurons potentials. Then duplicate the input to generate the number of timesteps given by qualia_plugin_snn.learningmodel.pytorch.SNN.SNN.timesteps Call qualia_plugin_snn.learningmodel.pytorch.SNN.SNN.forward() for each timestep. Finally, average the output of the model over the timesteps.

Parameters:

x (Tensor) – Input data

Returns:

Output predictions

Return type:

Tensor

trace_model(model: Module, extra_custom_layers: tuple[type[Module], ...] = ()) → tuple[Graph, GraphModule]

Parameters:

• model (Module)

• extra_custom_layers (tuple[type[Module], ...])

Return type:

tuple[Graph, GraphModule]

class qualia_plugin_snn.learningframework.SpikingJellyMultiStep

Bases: SpikingJelly

SpikingJelly multi-step LearningFramework implementation extending SpikingJelly single-step.

class TrainerModule

Bases: TrainerModule

SpikingJelly multi-step TrainerModule implementation extending SpikingJelly single-step TrainerModule.

forward(x: Tensor) → Tensor

Forward pass for a Spiking Neural Network model with duplicated timesteps in multi-step mode.

First calls SpikingJelly’s reset on the model to reset neurons potentials. Then duplicate the input to generate the number of timesteps given by qualia_plugin_snn.learningmodel.pytorch.SNN.SNN.timesteps. Timesteps are generated as a new dimension of the input tensor: [N, C, …] → [T, N, C, …] Call qualia_plugin_snn.learningmodel.pytorch.SNN.SNN.forward() for the whole tensor. Finally, average the output of the model over the timestep dimension.

Parameters:

x (Tensor) – Input data

Returns:

Output predictions

Return type:

Tensor

class qualia_plugin_snn.learningframework.SpikingJellyMultiStepTimeStepsInData

Bases: SpikingJellyTimeStepsInData

SpikingJelly multi-step with timesteps in data LearningFramework implementation extending SpikingJelly single-step.

class TrainerModule

Bases: TrainerModule

SpikingJelly multi-step with timesteps in data TrainerModule extending SpikingJelly single-step TrainerModule.

forward(x: Tensor) → Tensor

Forward pass for a Spiking Neural Network model with timesteps in input data in multi-step mode.

First calls SpikingJelly’s reset on the model to reset neurons potentials. Call qualia_plugin_snn.learningmodel.pytorch.SNN.SNN.forward() for each timestep of the input data. Finally, average the output of the model over the timesteps.

Parameters:

x (Tensor) – Input data with timestep dimension in [N, T, C, S] or [N, T, C, H, W] order

Returns:

Output predictions

Raises:

ValueError – when the input data does not have the correct number of dimenions or the timestep dimension does not match qualia_plugin_snn.learningmodel.pytorch.SNN.SNN.timesteps

Return type:

Tensor

class qualia_plugin_snn.learningframework.SpikingJellyTimeStepsInData

Bases: SpikingJelly

SpikingJelly single-step with timesteps in data LearningFramework implementation extending SpikingJelly single-step.

class TrainerModule

Bases: TrainerModule

SpikingJelly single-step with timesteps in data TrainerModule extending SpikingJelly single-step TrainerModule.

apply_dataaugmentation(batch: tuple[Tensor, Tensor], dataaugmentation: DataAugmentationPyTorch) → tuple[Tensor, Tensor]

Call a dataaugmentation module on the current batch.

If the dataaugmentation module is not a qualia_plugin_snn.dataaugmentation.pytorch.DataAugmentationPyTorchTimeStepsInData, or if qualia_plugin_snn.dataaugmentation.pytorch.DataAugmentationPyTorchTimeStepsInData.collapse_timesteps is True, then the timestep dimension is automatically merged with the batch dimension in order to support existing non-timestep-aware Qualia-Core dataaugmentation modules transparently.

Parameters:

• batch (tuple[Tensor, Tensor]) – Current batch of data and targets

• dataaugmentation (DataAugmentationPyTorch) – dataaugmentation module to apply

Returns:

Augmented batch

Return type:

tuple[Tensor, Tensor]

forward(x: Tensor) → Tensor

Forward pass for a Spiking Neural Network model with timesteps in input data in single-step mode.

First calls SpikingJelly’s reset on the model to reset neurons potentials. Call qualia_plugin_snn.learningmodel.pytorch.SNN.SNN.forward() for each timestep of the input data. Finally, average the output of the model over the timesteps.

Parameters:

x (Tensor) – Input data with timestep dimension in [N, T, C, S] or [N, T, C, H, W] order

Returns:

Output predictions

Raises:

ValueError – when the input data does not have the correct number of dimenions or the timestep dimension does not match qualia_plugin_snn.learningmodel.pytorch.SNN.SNN.timesteps

Return type:

Tensor

static channels_last_to_channels_first(x: ndarray[Any, Any]) → ndarray[Any, Any]

Channels last to channels first conversion with consideration of timestep dimension.

For 2D data with timestep: [N, T, H, W, C] → [N, T, C, H, W]

For 1D data with timestep: [N, T, S, C] → [N, T, C, S]

Parameters:

x (ndarray[Any, Any]) – NumPy array in channels last format

Returns:

NumPy array reordered to channels first format

Return type:

ndarray[Any, Any]

static channels_first_to_channels_last(x: ndarray[Any, Any]) → ndarray[Any, Any]

Channels first to channels last conversion with consideration of timestep dimension.

For 2D data with timestep: [N, T, C, H, W] → [N, T, H, W, C]

For 1D data with timestep: [N, T, C, S] → [N, T, S, C]

Parameters:

x (ndarray[Any, Any]) – NumPy array in channels first format

Returns:

NumPy array reordered to channels last format

Return type:

ndarray[Any, Any]

class DatasetFromArray

Bases: DatasetFromArray

Override qualia_core.learningframework.PyTorch.PyTorch.DatasetFromArray to reorder data with timestep dim.

__init__(dataset: RawData) → None

Load a Qualia dataset partition as a PyTorch dataset.

Reorders the input data from channels_last to channels_first, with consideration of timestep dimension.

Parameters:

dataset (RawData) – Dataset partition to load

Return type:

None