Questions about converting networks from lava-dl to lava

[fangwei123456]

lava.lib.dl.slayer.block provides a simple method to define deep SNN and conver SNN to process in Lava by lava.lib.dl.next. I build a toy network and test it:

from lava.magma.core.run_configs import Loihi1SimCfg
from lava.magma.core.run_conditions import RunSteps
from lava.proc.io.sink import RingBuffer as ReceiveProcess
from lava.proc.io.source import RingBuffer as SendProcess
import h5py
import torch
from lava.lib.dl import netx, slayer
filename ='./lava_net.net'
tau = 2.
scale = 1 << 6
in_features = 8
out_features = 1
N = 1
T = 8
neuron_param = {
            'threshold': 1.,
            'current_decay': 1.,
            'voltage_decay': 1. / tau,
            'tau_grad': 1, 'scale_grad': 1, 'scale': scale,
            'norm': None, 'dropout': None,
            'shared_param': True, 'persistent_state': False, 'requires_grad': False,
            'graded_spike': False
        }
with torch.no_grad():
    net = slayer.block.cuba.Dense(neuron_param, in_features, out_features)
    torch.nn.init.constant_(net.synapse.weight.data, 0.25)
    x = torch.rand([N, in_features, T])
    x = (x > 0.5).float()
    y = net(x)
    print('lava-dl output:\n', y.numpy()[0])
    h5_h = h5py.File(filename, 'w')
    h5_layer = h5_h.create_group('layer')
    net.export_hdf5(h5_layer.create_group(f'{0}'))

    net_x = netx.hdf5.Network(net_config=filename)
    source = SendProcess(x[0])
    sink = ReceiveProcess(shape=net_x.out.shape, buffer=T * 2)
    source.s_out.connect(net_x.inp)
    net_x.out.connect(sink.a_in)
    run_condition = RunSteps(num_steps=T)
    run_config = Loihi1SimCfg(select_tag='fixed_pt')
    net_x.run(condition=run_condition, run_cfg=run_config)
    output = sink.data.get()
    net_x.stop()
    print('lava output:\n', output)

The outputs are:

lava-dl output:
 [[0. 1. 0. 1. 1. 1. 1. 1.]]
lava output:
 [[0. 0. 1. 1. 0. 1. 1. 1. 0. 0. 0. 0. 0. 0. 0. 0.]]

I find that the two networks have different outputs (0. 1. 0. 1. 1. 1. 1. 1. is not in 0. 0. 1. 1. 0. 1. 1. 1. 0. 0. 0. 0. 0. 0. 0. 0.). Is the difference expected?

[bamsumit]

Hi @fangwei123456, thanks for reaching out to us. The neurons in slayer and lava should match except for some esoteric cases. What you have encountered is a specific discrepancy regarding the spike condition. A fix to this issue has been pushed in lava-dl PR #78. Once it is merged, the difference should be resolved and the outputs should match 1:1

The PR also includes a test inspired by your code snippet. Thanks for uncovering such issues. We highly appreciate it.

[fangwei123456]

Hi @bamsumit , I have another question about it. If a comment x = (x > 0.5).float(), then the input will be not binary, and I will get different outputs from lava/lava-dl. Does snn in lava only supports for binary input? And if I set graded_spike=True , I find in the source codes:

spikes = ((voltage >= threshold) * voltage / scale).to(dtype)

Does it means the spikes are also 8-bits, which is same with voltage ?

[bamsumit]

If graded_spike = True, we can send spikes with magnitude. This is a feature supported by Loihi 2 and hence lava-dl also supports it.

In the code above, if the voltage exceeds the threshold, a spike with a value equal to the voltage is sent.