def test_kwta_suppresses_all_but_k_units() -> None:
for p in np.linspace(0.0, 1.0, 10):
n = lb.Net()
n.new_layer(name="lr1", size=1)
lr2_spec = lb.LayerSpec(
inhibition_type="kwta",
kwta_pct=p,
unit_spec=lb.UnitSpec(adapt_dt=0, spike_gain=0))
n.new_layer(name="lr2", size=10, spec=lr2_spec)
pr1_spec = lb.ProjnSpec(dist=lb.Uniform(low=0.25, high=0.75))
n.new_projn("proj1", "lr1", "lr2", pr1_spec)
pr2_spec = lb.ProjnSpec(dist=lb.Uniform(low=0.25, high=0.75))
n.new_projn("proj2", "lr1", "lr2", pr2_spec)
n.clamp_layer("lr1", [1])
for i in range(50):
n.cycle()
v_m_eq = n.observe("lr2", "unit_v_m_eq")["v_m_eq"]
assert (v_m_eq > 0.5).sum() == max(1, int(round(10 * p)))
def test_kwta_avg_suppresses_all_but_k_units() -> None:
for p in np.linspace(0.4, 1.0, 10):
n = lb.Net()
n.new_layer(name="lr1", size=1)
lr2_spec = lb.LayerSpec(
inhibition_type="kwta_avg",
kwta_pct=p,
log_on_cycle=("unit_act", ),
unit_spec=lb.UnitSpec(adapt_dt=0, spike_gain=0))
n.new_layer(name="lr2", size=10, spec=lr2_spec)
pr1_spec = lb.ProjnSpec(dist=lb.Uniform(low=0.25, high=0.75))
n.new_projn("proj1", "lr1", "lr2", pr1_spec)
pr2_spec = lb.ProjnSpec(dist=lb.Uniform(low=0.25, high=0.75))
n.new_projn("proj2", "lr1", "lr2", pr2_spec)
n.clamp_layer("lr1", [1])
for i in range(100):
n.cycle()
logs = n.logs("cycle", "lr2").parts
acts = logs[logs.time == 99]["act"]
def test_a_simple_network_can_learn_simple_things() -> None:
network = lb.Net()
network.new_layer("input", size=2)
network.new_layer("output", size=2)
projn_spec = lb.ProjnSpec(lrate=0.2)
network.new_projn("input_to_output",
pre="input",
post="output",
spec=projn_spec)
input_patterns = [[1, 0], [0, 1]]
output_patterns = [[0, 1], [1, 0]]
batch(network, input_patterns, output_patterns)
actual = [output(network, i) for i in input_patterns]
for act, exp in zip(actual, output_patterns):
assert act == exp
def test_you_can_log_projection_weights() -> None:
network = lb.Net()
network.new_layer("input", size=2)
network.new_layer("output", size=2)
projn_spec = lb.ProjnSpec(log_on_trial=["conn_wt"])
network.new_projn("input_to_output",
pre="input",
post="output",
spec=projn_spec)
trial(network, (1, 0), (0, 1))
_, part_logs = network.logs(freq="trial", name="input_to_output")
expected = pd.DataFrame({
"pre_unit": (0, 1, 0, 1),
"conn_wt": (0.5, 0.5, 0.5, 0.5),
"post_unit": (0, 0, 1, 1),
"time": (0, 0, 0, 0)
})
pd.util.testing.assert_frame_equal(part_logs, expected, check_like=True)
for i in range(200):
net.cycle()
# Begin Markdown
# ### Graphing Logs
# Begin Code
wholeLog, partLog = net.logs(freq="cycle", name="output")
partLog.plot(x='time', figsize=(10, 6))
# Begin Markdown
# # One to Many Neuron Network
# Begin Code
# Define projection spec to have random weights
projn_spec = lb.ProjnSpec(dist=lb.Gaussian(mean=0.5, var=0.3))
# Begin Markdown
# ### Network Creation
# Begin Code
# Create the network
net = lb.Net()
# Create our layers and projections
net.new_layer("input", 1, layer_spec)
net.new_layer("output", 10, layer_spec)
net.new_projn("proj1", "input", "output", spec=projn_spec)
# Begin Markdown
# ### Run Network
# CA1 Layer Spec
CA1_layer_spec = lb.LayerSpec(
inhibition_type="kwta_avg",
kwta_pct=0.25,
kwta_pt=0.7,
)
# In[4]:
## Define Projections Spec
# Input -> EC_in
Input_projn_spec = lb.ProjnSpec(
dist=lb.Uniform(low=0.25, high=0.75),
lrate=0,
projn_type="one_to_one",
minus_phase=lb.NonePhase,
plus_phase=ThetaPlus,
)
# EC_in -> DG
# EC_in -> CA3
EC_in_projn_spec = lb.ProjnSpec(
dist=lb.Uniform(low=0.25, high=0.75),
lrate=0.2,
sparsity=0.25,
minus_phase=lb.NonePhase,
plus_phase=ThetaPlus,
)
# DG -> CA3 (Mossy Fiber)
