model

Train and test neuromorphic models.

Models are networks endowed with the methods fit(), predict(), similarity(), load(), and save(). This class provides an ML-focused API for the training and usage of sapicore Network output for practical purposes.

class model.Model(network: Network = None, **kwargs)

Model base class.

Loosely follows the design of scikit-learn’s sklearn.base.BaseEstimator interface.

Note

In a machine learning context, spiking networks can be used in diverse ways, e.g. as classifiers or as generative models. While engine is mostly about form (network architecture and information flow), model is all about function (how to fit the model to data and how to utilize it once trained).

fit(data: Tensor | Sequence[Tensor], duration: int | Sequence[int], rinse: int | Sequence[int] = 0, **kwargs)

Serves a batch of data, then turns off learning for all synapses.

Warning

fit() does not return intermediate output. Users should register forward hooks to efficiently stream data to memory or disk throughout the simulation (see add_data_hook()).

load(path: str) → Network

Loads a engine.network.Network from path and assigns it to this object’s network attribute.

Parameters:

path (str) – Path to the file containing the model.

Returns:

A reference to the loaded engine.network.Network object in case it is required by the calling function.

Return type:

Network

Note

The default implementation uses torch.load(), for the common case where files are used. Users may override this method when other formats are called for.

predict(data: Tensor, labels: Sequence, **kwargs) → Sequence

Predicts the labels of data.

Parameters:

  • data (Tensor) – Standalone 2D tensor of data buffer, sample X feature.

  • labels (Sequence) – Label values corresponding to classification layer cell indices.

Returns:

Vector of predicted labels.

Return type:

Sequence

save(path: str)

Saves the engine.network.Network object owned by this model to path.

Parameters:

path (str) – Destination path, inclusive of the file to which the network should be saved.

Note

The default implementation uses torch.save(), for the common case where files are used. Users may override this method when other formats are called for.

serve(data: Tensor | Sequence[Tensor], duration: int | Sequence[int], rinse: int | Sequence[int] = 0, **kwargs)

Applies engine.network.Network.forward() sequentially on a batch of buffer data.

Parameters:

  • data (Tensor or Sequence of Tensor) – 2D tensor(s) of data buffer to be fed to the root ensemble(s) of this object’s network, formatted sample X feature.

  • duration (int or Sequence of int) – Duration of sample presentation. Simulates duration of exposure to a particular input. If a list or a tensor is provided, the i-th sample in the batch is maintained for duration[i] steps.

  • rinse (int or Sequence of int) – Null stimulation steps (0s in-between samples). If a list or a tensor is provided, the i-th sample is followed by rinse[i] rinse steps.

  • duration[i] (Each sample i is presented for) –

  • rinse[i]. (followed by all 0s stimulation for) –

similarity(data: Tensor, metric: str | Callable, **kwargs) → Tensor

Performs a similarity analysis on network responses to data, yielding a pairwise distance matrix.

Parameters:

  • data (Tensor) – 2D tensor of data buffer, formatted sample X feature.

  • metric (str or Callable) – Distance metric to be used. If a string value is provided, it should correspond to one of the available scipy.spatial.distance metrics. If a custom function is provided, it should accept the data tensor and return a scalar corresponding to their distance.