spyx.loaders

Module Contents

Classes

MNIST_loader	Dataloader for the MNIST dataset. The data is returned in a packed format after using the pixel intensities as the p-value for sampling from
NMNIST_loader	Dataloading wrapper for the Neuromorphic MNIST dataset.
SHD_loader	Dataloading wrapper for the Spiking Heidelberg Dataset. The entire dataset is loaded to vRAM in a temporally compressed format. The user must

Attributes

optional_dependencies_installed
State

spyx.loaders.optional_dependencies_installed = True

spyx.loaders.State

class spyx.loaders.MNIST_loader(batch_size=32, sample_T=64, max_rate=0.75, val_size=0.3, data_subsample=1, key=0, download_dir='./MNIST')

Dataloader for the MNIST dataset. The data is returned in a packed format after using the pixel intensities as the p-value for sampling from a Bernoulli distribution.

Batch_size:

Number of samples per batch.

Sample_T:

Length of the time axis for each sample.

Max_rate:

Maximum number of spikes possible.

Val_size:

Fraction of the training set to set aside for validation.

Data_subsample:

use a subsample of the training/validation data to reduce computational demand.

Key:

An integer for setting the dataset loading random state.

Download_dir:

The directory to download the dataset to.

class spyx.loaders.NMNIST_loader(batch_size=32, sample_T=40, data_subsample=1, val_size=0.3, key=0, download_dir='./NMNIST')

Dataloading wrapper for the Neuromorphic MNIST dataset.

Batch_size:

Samples per batch.

Sample_T:

Timesteps per sample/length of time axis.

Data_subsample:

Use a fraction of the training/validation sets to reduce computational demand.

Val_size:

Proportion of dataset to set aside for validation.

Key:

Integer specifying the random seed for the train/val split.

Download_dir:

The local directory to save the data to.

class spyx.loaders.SHD_loader(batch_size=256, sample_T=128, channels=128, val_size=0.2)

Dataloading wrapper for the Spiking Heidelberg Dataset. The entire dataset is loaded to vRAM in a temporally compressed format. The user must apply jnp.unpackbits(events, axis=<time axis>) prior to feeding to an SNN.

https://zenkelab.org/resources/spiking-heidelberg-datasets-shd/

Batch_size:

Number of samples per batch.

Sample_T:

Number of time steps per sample.

Channels:

Number of frequency channels used.

Val_size:

Fraction of the training dataset to set aside for validation.