Available Operations

Available Operations#

Available torch.nn.modules for Devkits#

Current we support:

  • nn.Conv2d

    • convolution kernel size < 16

    • convolution channel number < 1024

  • nn.AvgPool2d(will be converted to SumPool2d when deployed to devkit)

    • pooling kernel size = [1, 2, 4]

  • nn.Linear(will be converted to equivalent convolution layer when deployed to devkit)

  • nn.Flatten

Available sinabs.layers for Devkits#

  • IAFSqueeze

  • SumPool2d

    • pooling kernel size = [1, 2, 4]

Neuron Leakage#

Our devkit support a constant leakage on the membrane-potential of the IAF neuron. To achieve that, all you need to do is:

  • Enable the bias term of the nn.Conv2d layer.

  • Make sure the bias is lower than 0 after the training.

  • Turn on the slow-clock of the devkit which driven the leakage functionality when deploy the SNN to the devkit.

For more details see how to leak neuron