Dataset

Dataset wrapper for use with SLAYER.

class slayerSNN.auto.dataset.SlayerDataset(dataset, network, randomShift=False, binningMode='OR', fullDataset=False)[source]

This class wraps a basic dataset class to be used in SLAYER training. This allows the use of the same basic dataset definition on some other platform other than SLAYER, for e.g. for implementation in a neuromorphic hardware with its SDK.

The basic dataset must return a numpy array of events where each row consists of an AER event represented by x, y, polarity and time (in ms).

Arguments:

  • dataset: basic dataset to be wrapped.

  • network: an auto module network with which the dataset is intended to be used with.

    The shape of the tensor is determined from the netowrk definition.

  • randomShift: a flag to indicate if the sample must be randomly shifted in time over the

    entire sample length. Default: False

  • binningMode: the way the overlapping events are binned. Supports SUM and OR binning.

    Default: OR

  • fullDataset: a flag that indicates weather the full dataset is to be processed or not.

    If True, full length of the events is loaded into tensor. This will cause problems with default batching, as the number of time bins will not match for all the samples in a minibatch. In this case, the dataloader’s collate_fn must be custom defined or a batch size of 1 should be used. Default: False

Usage:

dataset = SlayerDataset(dataset, net)