• This is a comment, not an issue *

Hi Friedemann, First of thanks a lot for these great tutorials, I've enjoyed a lot playing with them, and I've learned a lot :-) One question: in the run_snn function, why do you bother constructing the "rst" tensor? Why don't you subtract the "out" tensor, which also contains the output spikes? I've tried, and it seems to work. Just curious. Best,

Tim

Hello,

It was a very nice and interesting tutorial, thank you for preparing it...

tutorial1 haven't any problem, but in tutorial 2, some dtype problems occurred... after their fixation, training process was very slow on GTX 980 (I've run on this config some very deep model)... could you please explain your config, and also training time and response time?

I have the impression that the spike recordings are shifted one time step in all tutorials. Could you maybe check if this is indeed the case?

From my understanding, time step 0 is recorded twice for the spikes, once during initialisation

  mem = torch.zeros((batch_size, nb_hidden), device=device, dtype=dtype)
  spk_rec = [mem]

and once within the simulation of time step 0:

  for t in range(nb_steps):
      mthr = mem-1.0
      out = spike_fn(mthr)
      ...
      spk_rec.append(out)

As a result the indeces appear shifted when comparing

print(torch.nonzero((mem_rec-1.0) > 0.0))
print(torch.nonzero(spk_rec))

Thanks, Simon

Hey Friedemann,

thanks for making the examples available, they look very helpful. However, to make them fully reproducible I think that some additional information regarding the "technical dependencies" is needed.

In particular, the list of used software packages (incl. version and build variant information) plus some specification about the machine hardware (CPU arch, GPUs).

Preferably, the former could be expressed as a recipe for constructing a container (Dockerfile, or for better HPC-compatibility, a Singularity recipe), maybe even using an explicitly versioning package manager like spack.

Cheers, Eric

Hello,

I belive I ran into a possible issue here. Due to line 37 the evaluation in line 38 will always be false if one hasnt already got the uncompressed dataset.

https://github.com/fzenke/spytorch/blob/9e91eceaf53f17be9e95a3743164224bdbb086bb/notebooks/utils.py#L35-L42

If I change line 37 to: hdf5_file_path = gz_file_path[:-3] This works for me.

Best, Aaron

Hi zenke, I have a question about the snn model. If I feed a spike image to a snn with L layers at time step n, the output of the last layer will be affected by the input at time step n + L - 1. In deep networks, the delay should be considered, because it will increase the whole time steps.

Hey Friedemann,

thank you for the very comprehensive tutorial! I have a question on the way the recurrence is computed in tutorial 4. If I understand the equation for the dynamics of the current correctly, the recurrence should be computed with the spiking neuron state:

mthr = mem-1.0
out = spike_fn(mthr)
h1 = h1_from_input[:,t] + torch.einsum("ab,bc->ac", (out, v1))

Instead in tutorial 4, a separate hidden state is kept, that ignores the spike function:

h1 = h1_from_input[:,t] + torch.einsum("ab,bc->ac", (h1, v1))

Is this done deliberately? Judging from simulating a few epochs, the two versions seem to perform similarly.

Thank you,

Simon

I don't understand why the 'rst' variable exists. It seems to always be == 'out'. Changing to rst = out yields same results...

def spike_fn(x):
    out = torch.zeros_like(x)
    out[x > 0] = 1.0
    return out
...
# Here we loop over time
for t in range(nb_steps):
    mthr = mem-1.0
    out = spike_fn(mthr) 
    rst = torch.zeros_like(mem)
    c = (mthr > 0)
    rst[c] = torch.ones_like(mem)[c] 

When I am running the following piece of code in Tutorial-4:

loss_hist = train(x_train, y_train, lr=2e-4, nb_epochs=nb_epochs)

I am getting the following error:

Can you please suggest me how to resolve this issue?