latency() got an unexpected keyword argument 'num_outputs' and latency() got multiple values for argument 'num_steps'

• snntorch version: 0.2.11
• Python version: 3.7.10

Description

In the latest updates, I see that the num_output has been removed from the spikegen. No clue why!

Moreover, also when I just decide about removing the num_output parameter. I get another error from the num_steps when I set it to any int value?

BTW, the same exists if I tried to do rate encoding instead of latency encoding.

What I Did

Here: you will find a quick ipynb file that shows the errors.

[jeshraghian]

Gotcha!

This should work.

Note that spikegen.latency is now intended for input features. So including the number of outputs/classes would be meaningless.

You can instead separately enter that into spikegen.targets_latency.

# Spiking Data
spike_data = spikegen.latency(data_it, num_steps=num_steps, tau=5, threshold=0.01, linear=True, normalize=True, clip=True)

spike_targets = spikegen.targets_latency(targets_it, num_classes=num_outputs, num_steps=num_steps, tau=5, threshold=0.01,
                                                     linear=True, normalize=True, clip=True)

[jeshraghian]

Hi @Dola47,

Just a reminder that issues are intended for reporting bugs. When in doubt, open a new thread in discussions so that users with the same question can refer to that easily. If it turns out it’s a bug, then we can convert the discussion into an issue!

We will have 2d targets and we will not be able to use it with the NNLLoss for example, or even the MSELoss. So I wonder how you expect the spiking targets to be used within your code?

A 2d target of [batch_size x num_outputs] can be used to apply a separate MSELoss to each output neuron. Such a target represents the ‘goal’ of the membrane potential of each neuron. The 10 losses are then summed together both over time, and over neurons. You can can check this notebook out as an example:

https://colab.research.google.com/drive/1j9-ddyadCMyn6N12zjPmu98VAmYF3YIl?usp=sharing

Another note; I’ve since added a couple other neuron types.

In my own experiments, Leaky consistently outperforms Synaptic (used to be Stein) for the right set of parameters. The only difference is that Leaky takes an input, integrates it into membrane potential (decay at rate beta), and then emits a spike. Stein/Synaptic take an input, integrate it into a synaptic current (decay at rate alpha), which is integrated again into membrane potential (decay at rate beta), and then emits a spike. The notebook above shows a code sample with Leaky.

Next question -

Then we do not need the spiking targets anymore when we call print_batch_accracy but we also need the ground truth targets of the batch without any encoding?

Yeah - you are absolutely right. No need for the converted targets for this particular task. You only compare the raw ground truth targets to the spike count.

There are obviously other tasks that might require the converted targets to measure accuracy, e.g., if you are trying to encourage a neuron to fire at very specific times, or the goal is to achieve a given pattern/evolution of membrane potential.