aihwkit.simulator.noise_models module

Phenomenological noise models for inference.

class aihwkit.simulator.noise_models.BaseConductanceConverter

Bases: object

Base class for converting DNN weights into conductances.

convert_back_to_weights(conductances, params)

Converting a matrix of conductances into weights.

Caution

The conversion is assumed deterministic and repeatable.

Parameters

• conductances – list of conductance tensors representing a weight matrix

• params – param dictionary that was returned from the conversion

Returns

weight matrix

Return type

torch.Tensor

convert_to_conductances(weights)

Converting a weight matrix into conductances.

Caution

The conversion is assumed deterministic and repeatable.

Parameters

weights – weight matrix tensor.

Returns

Tuple of the list of conductance tensors and a params dictionary that is used for the reverse conversion.

Return type

Tuple[List[torch.Tensor], Dict]

class aihwkit.simulator.noise_models.BaseDriftCompensation

Bases: object

Base class for drift compensations.

apply(forward_output, ref_value)

Read out the current value from the output of the forward pass and returns the drift compensation alpha scale.

Return type

torch.Tensor

get_readout_tensor(in_size)

Return the read-out tensor.

Called once during init_baseline().

Return type

torch.Tensor

init_baseline(forward_output)

Initialize the base line for applying the compensation.

Uses a all one tensor for read_out.

Parameters

forward_output – forward output of the read out vector to compensate

Returns

reference tensor readout

Return type

Tuple[torch.Tensor, torch.Tensor]

readout(out_tensor)

Implements the read out math.

Return type

torch.Tensor

class aihwkit.simulator.noise_models.BaseNoiseModel(g_converter=None)

Bases: object

Base class for phenomenological noise models for inference.

apply_drift_noise(weights, nu_drift_list, t_inference)

Applies the expected drift noise to weights.

Uses the apply_drift_noise_to_conductances() on each of the conductance slices.

Parameters

• weights – weights tensor (usually with programming noise already applied)

• nu_drift_list – list of drift nu for each conductance slice

• t_inference – assumed time of inference (in sec)

Returns

weight tensor with drift noise applied

Return type

torch.Tensor

apply_drift_noise_to_conductance(g_prog, nu_drift, t_inference)

Apply the noise and drift up to the assumed inference time point.

Parameters

• g_prog – Tensor of conductance values after programming (in \(\muS\))

• nu_drift – drift nu

• t_inference – assumed time of inference (in sec)

Returns

conductance Tensor with applied noise and drift

Return type

torch.Tensor

apply_noise(weights, t_inference)

Applies the expected noise.

Applies the noise to a non-perturbed conductance matrix weights at time of inference t_inference (in seconds) where 0 sec refers to the time when weight programming has finished.

Note

The drift coefficients and intermediate noises etc. are sampled for each application of this function anew from the distributions, thus it samples the expected noise and drift behavior at time t_inference but not a continual trajectory of a given device instance over time (having e.g. constant drift coefficients).

Return type

torch.Tensor

apply_programming_noise(weights)

Applies the expected programming noise to weights.

Uses the apply_programming_noise_to_conductances() on each of the conductance slices.

Parameters

weights – weights tensor

Returns

weight tensor with programming noise applied, and tuple of all drift coefficients (per conductances slice) that are determined during programming.

Return type

Tuple[torch.Tensor, List[torch.Tensor]]

apply_programming_noise_to_conductance(g_target)

Apply programming noise to a target conductance Tensor.

Parameters

g_target – Target conductances

Returns

Tensor of sampled drift coefficients \(\nu\), one for each target conductance value.

Return type

torch.Tensor

generate_drift_coefficients(g_target)

Generates drift coefficients nu based on the target conductances.

Return type

torch.Tensor

class aihwkit.simulator.noise_models.GlobalDriftCompensation

Bases: aihwkit.simulator.noise_models.BaseDriftCompensation

Global drift compensation.

Uses a constant factor for compensating the drift.

get_readout_tensor(in_size)

Returns the read-out tensor.

Uses a single all one vector.

Return type

torch.Tensor

readout(out_tensor)

Read outs the abs max.

Return type

torch.Tensor

class aihwkit.simulator.noise_models.PCMLikeNoiseModel(prog_coeff=None, g_converter=None, g_max=None, t_read=2.5e-07, t_0=20.0)

Bases: aihwkit.simulator.noise_models.BaseNoiseModel

Noise model that was fitted and characterized on real PCM devices.

Expected weight noise at assumed time of inference with expected programming noise at 0.

The statistical noise model is based on measured PCM devices.

Parameters

• prog_coeff – programming polynomial coeffs in \(\mu S\), c(0) + c(1)*gt + c(2)*gt^2)

• g_converter – instantiated class of the conductance converter (defaults to single pair)

• t_read – parameter of the 1/f fit (in seconds)

• t_0 – parameter of the drift fit (first reading time)

• Note – The t_inference is relative to this time t0 e.g. t_inference counts from the completion of the programming of a device.

apply_drift_noise_to_conductance(g_prog, nu_drift, t_inference)

Apply the noise and drift up to the assumed inference time point based on PCM measurements.

Return type

torch.Tensor

apply_programming_noise_to_conductance(g_target)

Apply programming noise to a target conductance Tensor.

Programming noise with additive Gaussian noise with conductance dependency of the variance given by a 2-degree polynomial.

Return type

torch.Tensor

generate_drift_coefficients(g_target)

Return drift coefficients nu based on PCM measurements.

Return type

torch.Tensor

class aihwkit.simulator.noise_models.SinglePairConductanceConverter(g_max=None, g_min=None)

Bases: aihwkit.simulator.noise_models.BaseConductanceConverter

Single pair of devices.

Assuming a single pair of devices per cross-point, taking positive and negative weights, respectively, where one device is always at 0.

Parameters

• g_max – In \(\mu S\), the maximal conductance, ie the value the absolute max of the weights will be mapped to.

• g_min – In \(\mu S\), the minimal conductance, ie the value the logical zero of the weights will be mapped to.

convert_back_to_weights(conductances, params)

Converting a matrix of conductances into weights.

Caution

The conversion is assumed deterministic and repeatable.

Parameters

• conductances – list of conductance tensors representing a weight matrix

• params – param dictionary that was returned from the conversion

Returns

weight matrix

Return type

torch.Tensor

convert_to_conductances(weights)

Converting a weight matrix into conductances.

Caution

The conversion is assumed deterministic and repeatable.

Parameters

weights – weight matrix tensor.

Returns

Tuple of the list of conductance tensors and a params dictionary that is used for the reverse conversion.

Return type

Tuple[List[torch.Tensor], Dict]