qualia_plugin_snn.learningmodel.pytorch.layers.quantized_SNN_layers module

Contains quantized spiking neuron implementations.

class qualia_plugin_snn.learningmodel.pytorch.layers.quantized_SNN_layers.QuantizedLIFNode

Bases: LIFNode, QuantizerInputProtocol, QuantizerActProtocol, QuantizedLayer

Quantized variant of SpikingJelly’s spikingjelly.activation_based.neuron.LIFNode.

Hyperparameters v_threshold, v_reset and tau are quantized as well as membrane potential v.

v: Tensor

v_threshold: float

v_reset: float | None

tau: float

__init__(quant_params: QuantizationConfigDict, tau: float = 2.0, decay_input: bool = True, v_threshold: float = 1.0, v_reset: float = 0.0, detach_reset: bool = False, step_mode: str = 's', backend: str = 'torch') → None

Construct QuantizedLIFNode.

For more information about spiking neuron parameters, see: spikingjelly.activation_based.neuron.LIFNode.__init__()

Parameters:

• tau (float) – Membrane time constant

• decay_input (bool) – Whether the input will decay

• v_threshold (float) – Threshold of this neurons layer

• v_reset (float) – Reset voltage of this neurons layer. If not None, the neuron’s voltage will be set to v_reset after firing a spike. If None, the neuron’s voltage will subtract v_threshold after firing a spike

• detach_reset (bool) – Whether detach the computation graph of reset in backward

• step_mode (str) – The step mode, which can be s (single-step) or m (multi-step)

• backend (str) – backend fot this neurons layer, only ‘torch’ is supported

• quant_params (QuantizationConfigDict) – Quantization configuration dict, see qualia_core.learningmodel.pytorch.layers.Quantizer.Quantizer

Return type:

None

property supported_backends: tuple[Literal['torch']]

Supported step_mode and backend.

Only torch backend is supported.

Returns:

Tuple of 'torch'

Raises:

ValueError – When step_mode is not 's' or 'm'

neuronal_charge(x: Tensor) → None

Quantized spikingjelly.activation_based.neuron.LIFNode.neuronal_charge().

Membrane potential and hyperparameters are quantized before and after computation using quantize_v_and_hyperparams().

Parameters:

x (Tensor) – Input tensor

Return type:

None

single_step_forward(x: Tensor) → Tensor

Quantized spikingjelly.activation_based.neuron.LIFNode.single_step_forward().

Input is (optionally) quantized. Membrane potential and hyperparameters are quantized before and after computation using quantize_v_and_hyperparams().

Parameters:

x (Tensor) – Input tensor

Return type:

Tensor

multi_step_forward(x_seq: Tensor) → Tensor

Implement multi-step as loop over single-step for quantized neurons, inefficient but at least it works.

Parameters:

x_seq (Tensor) – Input tensor with timesteps

Returns:

Output tensor with timesteps

Return type:

Tensor

property reciprocal_tau: float

Return 1 / tau.

Returns:

1 / tau

quantize_v_and_hyperparams() → None

Quantize potential and hyperparameters (v_threshold, tau, v_reset) in-place with the same quantizer at the same time.

tau is not quantized directly, instead quantize reciprocal_tau (1 / tau) because this is what is used during inference to avoid division.

Return type:

None

get_hyperparams_tensor(device: device, dtype: dtype) → Tensor

Pack v_threshold, reciprocal_tau and optionally v_reset into the same Tensor.

Parameters:

• device (device) – Device to create the tensor on

• dtype (dtype) – Data type for the created tensor

Returns:

New tensor with hyperparemeters concatenated

Return type:

Tensor

property weights_bits: int | None

Total number of bits used to encode the membrane potential and hyperparameters in case of fixed-point quantization.

See qualia_core.learningmodel.pytorch.layers.Quantizer.Quantizer.bits().

Returns:

Total bits for the membrane potential and hyperparameters or None if not applicable.

property weights_q: int | None

Number of fractional part bits for the membrane potential and hyperparameters in case of fixed-point quantization.

See qualia_core.learningmodel.pytorch.layers.Quantizer.Quantizer.fractional_bits().

Returns:

Fractional part bits for the membrane potential and hyperparameters or None if not applicable.

property weights_round_mode: str | None

class qualia_plugin_snn.learningmodel.pytorch.layers.quantized_SNN_layers.QuantizedIFNode

Bases: IFNode, QuantizerInputProtocol, QuantizerActProtocol, QuantizedLayer

Quantized variant of SpikingJelly’s spikingjelly.activation_based.neuron.IFNode.

Hyperparameters v_threshold and v_reset are quantized as well as membrane potential v.

v: Tensor

v_threshold: float

v_reset: float | None

__init__(quant_params: QuantizationConfigDict, v_threshold: float = 1.0, v_reset: float = 0.0, detach_reset: bool = False, step_mode: str = 's', backend: str = 'torch') → None

Construct QuantizedIFNode.

For more information about spiking neuron parameters, see: spikingjelly.activation_based.neuron.IFNode.__init__()

Parameters:

• v_threshold (float) – Threshold of this neurons layer

• v_reset (float) – Reset voltage of this neurons layer. If not None, the neuron’s voltage will be set to v_reset after firing a spike. If None, the neuron’s voltage will subtract v_threshold after firing a spike

• detach_reset (bool) – Whether detach the computation graph of reset in backward

• step_mode (str) – The step mode, which can be s (single-step) or m (multi-step)

• backend (str) – backend fot this neurons layer, only ‘torch’ is supported

• quant_params (QuantizationConfigDict) – Quantization configuration dict, see qualia_core.learningmodel.pytorch.layers.Quantizer.Quantizer

Return type:

None

property supported_backends: tuple[Literal['torch']]

Supported step_mode and backend.

Only torch backend is supported.

Returns:

Tuple of 'torch'

Raises:

ValueError – When step_mode is not 's' or 'm'

neuronal_charge(x: Tensor) → None

Quantized spikingjelly.activation_based.neuron.IFNode.neuronal_charge().

Membrane potential and hyperparameters are quantized before and after computation using quantize_v_and_hyperparams().

Parameters:

x (Tensor) – Input tensor

Return type:

None

single_step_forward(x: Tensor) → Tensor

Quantized spikingjelly.activation_based.neuron.IFNode.single_step_forward().

Input is (optionally) quantized. Membrane potential and hyperparameters are quantized before and after computation using quantize_v_and_hyperparams().

Parameters:

x (Tensor) – Input tensor

Return type:

Tensor

multi_step_forward(x_seq: Tensor) → Tensor

Implement multi-step as loop over single-step for quantized neurons, inefficient but at least it works.

Parameters:

x_seq (Tensor) – Input tensor with timesteps

Returns:

Output tensor with timesteps

Return type:

Tensor

quantize_v_and_hyperparams() → None

Quantize potential and hyperparameters (v_threshold, v_reset) in-place with the same quantizer at the same time.

Return type:

None

get_hyperparams_tensor(device: device, dtype: dtype) → Tensor

Pack v_threshold and optionally v_reset into the same Tensor.

Parameters:

• device (device) – Device to create the tensor on

• dtype (dtype) – Data type for the created tensor

Returns:

New tensor with hyperparemeters concatenated

Return type:

Tensor

property weights_bits: int | None

Total number of bits used to encode the membrane potential and hyperparameters in case of fixed-point quantization.

See qualia_core.learningmodel.pytorch.layers.Quantizer.Quantizer.bits().

Returns:

Total bits for the membrane potential and hyperparameters or None if not applicable.

property weights_q: int | None

Number of fractional part bits for the membrane potential and hyperparameters in case of fixed-point quantization.

See qualia_core.learningmodel.pytorch.layers.Quantizer.Quantizer.fractional_bits().

Returns:

Fractional part bits for the membrane potential and hyperparameters or None if not applicable.

property weights_round_mode: str | None

class qualia_plugin_snn.learningmodel.pytorch.layers.quantized_SNN_layers.QuantizedATIF

Bases: ATIF, QuantizerInputProtocol, QuantizerActProtocol, QuantizedLayer

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

__init__(quant_params: QuantizationConfigDict, 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

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

• quant_params (QuantizationConfigDict) – Quantization configuration dict, see qualia_core.learningmodel.pytorch.layers.Quantizer.Quantizer

Return type:

None

ifsrl_fn(x: Tensor) → Tensor

Quantized qualia_plugin_snn.learningmodel.pytorch.layers.CustomNode.ATIF.ifsrl_fn().

Input is quantized. Membrane potential is quantized before and after computation. Threshold is quantized by get_coeffs().

Parameters:

x (Tensor) – Input tensor

Returns:

Output tensor

Return type:

Tensor

get_coeffs() → Tensor

Return the quantized Tensor of threshold v_threshold.

Returns:

Quantized Tensor of threshold v_threshold

Return type:

Tensor

set_coeffs(v_threshold: Tensor) → None

Quantized and replace the Tensor of threshold v_threshold.

Parameters:

v_threshold (Tensor) – New Tensor of quantized threshold to replace v_threshold

Return type:

None

property weights_bits: int | None

Total number of bits used to encode the membrane potential and hyperparameters in case of fixed-point quantization.

See qualia_core.learningmodel.pytorch.layers.Quantizer.Quantizer.bits().

Returns:

Total bits for the membrane potential and hyperparameters or None if not applicable.

property weights_q: int | None

Number of fractional part bits for the membrane potential and hyperparameters in case of fixed-point quantization.

See qualia_core.learningmodel.pytorch.layers.Quantizer.Quantizer.fractional_bits().

Returns:

Fractional part bits for the membrane potential and hyperparameters or None if not applicable.

property weights_round_mode: str | None