aihwkit_lightning.simulator.parameters.enums module

Utility enumerators for resistive processing units.

class aihwkit_lightning.simulator.parameters.enums.WeightClipType(value)

Bases: Enum

Weight clipper type.

LAYER_GAUSSIAN = 'LayerGaussian'

Calculates the second moment of the whole weight matrix and clips at sigma times the result symmetrically around zero.

LAYER_GAUSSIAN_PER_CHANNEL = 'LayerGaussianPerChannel'

Calculates the second moment of the whole weight matrix per output column and clips at sigma times the result symmetrically around zero.

LEARNABLE_PER_CHANNEL = 'LearnablePerChannel'

Implements ParetoQ-like learnable weight clipping (originally LSQ) see eq. 3 from https://arxiv.org/pdf/1902.08153. Implements this on a per-channel basis.

NONE = 'None'

None.

class aihwkit_lightning.simulator.parameters.enums.WeightModifierType(value)

Bases: Enum

Weight modifier type.

ADD_NORMAL = 'AddNormal'

Additive Gaussian noise.

ADD_NORMAL_PER_CHANNEL = 'AddNormalPerChannel'

Additive Gaussian noise per channel.

DISCRETIZE = 'Discretize'

Quantize the weights.

DISCRETIZE_ADD_NORMAL = 'DiscretizeAddNormal'

First discretize and then additive Gaussian noise.

DISCRETIZE_ADD_NORMAL_PER_CHANNEL = 'DiscretizeAddNormalPerChannel'

First discretize per channel and then additive Gaussian noise per channel.

DISCRETIZE_PER_CHANNEL = 'DiscretizePerChannel'

Quantize the weights per channel.

NONE = 'None'

No weight modifier. Nothing happens to the weight.

class aihwkit_lightning.simulator.parameters.enums.WeightNoiseInjectionType(value)

Bases: Enum

Weight modifier type.

ADD_NORMAL = 'AddNormal'

Additive Gaussian noise.

ADD_NORMAL_PER_CHANNEL = 'AddNormalPerChannel'

Additive Gaussian noise per channel.

NONE = 'None'

No weight modifier. Nothing happens to the weight.

class aihwkit_lightning.simulator.parameters.enums.WeightQuantizationType(value)

Bases: Enum

Weight quantization type.

DISCRETIZE = 'Discretize'

Quantize the weights.

DISCRETIZE_PER_CHANNEL = 'DiscretizePerChannel'

Quantize the weights per channel.

NONE = 'None'

No weight quantizer. Nothing happens to the weight.