def test_sequential_forward_state_hook():
data = torch.ones(10, 2, 8, 1)
model = norse.SequentialState(
torch.nn.Linear(1, 10, bias=False),
norse.LIFRecurrent(10, 10),
norse.LIF(),
norse.LIF(),
norse.LIF(),
)
spikes = []
states = []
def forward_state_hook(mod, inp, output):
spikes.append(inp[0])
states.append(inp[1])
model.register_forward_state_hooks(forward_state_hook)
_, s = model(data)
spikes = torch.stack(spikes).detach().cpu()
assert spikes.shape == (4, 10, 2, 8, 10)
assert spikes.max() > 0
assert len(states) == 4
assert states[0] is None
spikes = []
states = []
model(data, s)
assert hasattr(states[0], "v") # isinstance is broken for namedtuple
model.remove_forward_state_hooks()
model(data, s)
assert len(spikes) == 4
def test_state_sequence_norse():
d = torch.ones(10, 2, 10)
l1 = norse.LIFRecurrent(10, 5)
l2 = norse.LSNNRecurrent(5, 1)
z, (s1, s2) = norse.SequentialState(l1, l2)(d)
assert z.shape == (10, 2, 1)
assert s1.v.shape == (2, 5)
assert s2.v.shape == (2, 1)
def test_state_sequence_list():
d = torch.ones(10, 1, 20)
l1 = norse.LIFRecurrent(20, 6)
l2 = norse.LIFRecurrent(6, 2)
s = [None, None]
z, s = norse.SequentialState(l1, l2)(d, s)
assert z.shape == (10, 1, 2)
assert s[1].v.shape == (1, 2)
def test_state_sequence_conv():
data = torch.ones(1, 8, 16, 4, 4) # (timestep, minibatch, channels, x, y)
model = norse.SequentialState(
norse.Lift(torch.nn.Conv2d(16, 8, 3)), # (1, 8, 8, 2, 2)
torch.nn.Flatten(2), # (1, 8, 32)
norse.LSNNRecurrent(32, 6), # (1, 8, 6)
torch.nn.RNN(6, 4, 2), # (1, 6, 4) with 2 recurrent layers
norse.LIFRecurrent(4, 1), # (1, 4, 1)
)
model(data)
def test_state_sequence_mix():
d = torch.ones(10, 3, 20)
l1 = norse.LIFRecurrent(20, 10)
l2 = torch.nn.RNN(10, 4, 2) # 2 layers
l3 = norse.LSNNRecurrent(4, 1)
state = [None, torch.randn(2, 3, 4), None]
z, (s1, s2, s3) = norse.SequentialState(l1, l2, l3)(d, state)
assert z.shape == (10, 3, 1)
assert s1.v.shape == (3, 10)
assert s2.shape == (2, 3, 4)
assert s3.v.shape == (3, 1)
def test_sequential_debug_hook_twice():
def dub(mod, inp, output):
pass
model = norse.SequentialState(
torch.nn.Linear(1, 10, bias=False),
norse.LIFRecurrent(10, 10),
)
model.register_forward_state_hooks(dub)
with pytest.raises(ValueError):
model.register_forward_state_hooks(dub)
def test_backprop_works():
model = norse.SequentialState(
norse.LSNNRecurrent(1, 2),
norse.LSNNRecurrent(2, 3),
norse.LSNNRecurrent(3, 3),
norse.LSNNRecurrent(3, 2),
nn.Flatten(),
norse.Lift(torch.nn.Linear(4, 2)),
norse.Lift(torch.nn.Linear(2, 1)),
)
optimizer = torch.optim.Adam(model.parameters(), lr=0.0001)
loss_func = nn.MSELoss()
data = torch.ones(3, 1, 2, 1)
target = torch.ones(1, 1)
state = None
for x in data:
out, state = model(x, state)
loss = loss_func(out, target)
optimizer.zero_grad() # clear gradients for this training step
loss.backward() # backpropagation, compute gradients
optimizer.step()
state = [s.detach() if s is not None else None for s in state]
def test_state_sequence():
d = torch.ones(10, 1, 20)
l = norse.LIFRecurrent(20, 6)
z, s = norse.SequentialState(l)(d)
assert z.shape == (10, 1, 6)
assert s[0].v.shape == (1, 6)
