def test_current(rng):
neuron_type = NeuronType()
n_neurons = 20
gain = rng.rand(n_neurons)
bias = rng.rand(n_neurons)
# 3 samples
x = rng.rand(3)
current = neuron_type.current(x, gain, bias)
assert np.allclose(current, gain * x.reshape(-1, 1) + bias)
assert current.shape == (3, n_neurons)
# 10 samples, different values for each neuron
x = rng.rand(10, n_neurons)
current = neuron_type.current(x, gain, bias)
assert np.allclose(current, gain * x + bias)
assert current.shape == (10, n_neurons)
with pytest.raises(ValidationError):
# Incorrect second dimension
x = rng.rand(10, 2)
current = neuron_type.current(x, gain, bias)
with pytest.raises(ValidationError):
# Too many dimensions
x = rng.rand(10, n_neurons, 1)
current = neuron_type.current(x, gain, bias)
def test_gain_bias(rng, NonDirectNeuronType, generic, allclose):
if NonDirectNeuronType == Sigmoid and generic:
# the generic method doesn't work with sigmoid neurons (because they're
# always positive). that's not a failure, because the sigmoid neurons
# never need to use the generic method normally, so we'll just skip
# it for this test.
return
n = 100
max_rates = rng.uniform(300, 400, size=n)
intercepts = rng.uniform(-0.5, 0.5, size=n)
neuron_type = NonDirectNeuronType()
tolerance = 0.1 if generic else 1e-8
if generic:
gain, bias = NeuronType.gain_bias(neuron_type, max_rates, intercepts)
else:
gain, bias = neuron_type.gain_bias(max_rates, intercepts)
assert allclose(neuron_type.rates(1, gain, bias),
max_rates,
atol=tolerance)
if NonDirectNeuronType == Sigmoid:
threshold = 0.5 / neuron_type.tau_ref
else:
threshold = 0
x = (intercepts - tolerance)[np.newaxis, :]
assert np.all(neuron_type.rates(x, gain, bias) <= threshold)
x = (intercepts + tolerance)[np.newaxis, :]
assert np.all(neuron_type.rates(x, gain, bias) > threshold)
if generic:
max_rates0, intercepts0 = NeuronType.max_rates_intercepts(
neuron_type, gain, bias)
else:
max_rates0, intercepts0 = neuron_type.max_rates_intercepts(gain, bias)
assert allclose(max_rates, max_rates0, atol=tolerance)
assert allclose(intercepts, intercepts0, atol=tolerance)
def test_current(rng, allclose):
neuron_type = NeuronType()
n_neurons = 20
gain = rng.rand(n_neurons)
bias = rng.rand(n_neurons)
# 3 samples
x = rng.rand(3)
current = neuron_type.current(x, gain, bias)
assert allclose(current, gain * x.reshape(-1, 1) + bias)
assert current.shape == (3, n_neurons)
# 10 samples, different values for each neuron
x = rng.rand(10, n_neurons)
current = neuron_type.current(x, gain, bias)
assert allclose(current, gain * x + bias)
assert current.shape == (10, n_neurons)
with pytest.raises(ValidationError):
# Incorrect second dimension
x = rng.rand(10, 2)
current = neuron_type.current(x, gain, bias)
with pytest.raises(ValidationError):
# Too many dimensions
x = rng.rand(10, n_neurons, 1)
current = neuron_type.current(x, gain, bias)
def test_gain_bias(rng, nl_nodirect, generic):
if nl_nodirect == Sigmoid and generic:
# the generic method doesn't work with sigmoid neurons (because they're
# always positive). that's not a failure, because the sigmoid neurons
# never need to use the generic method normally, so we'll just skip
# it for this test.
return
n = 100
max_rates = rng.uniform(300, 400, size=n)
intercepts = rng.uniform(-0.5, 0.5, size=n)
nl = nl_nodirect()
tolerance = 0.1 if generic else 1e-8
if generic:
gain, bias = NeuronType.gain_bias(nl, max_rates, intercepts)
else:
gain, bias = nl.gain_bias(max_rates, intercepts)
assert np.allclose(nl.rates(1, gain, bias), max_rates, atol=tolerance)
if nl_nodirect == Sigmoid:
threshold = 0.5 / nl.tau_ref
else:
threshold = 0
x = (intercepts - tolerance)[np.newaxis, :]
assert np.all(nl.rates(x, gain, bias) <= threshold)
x = (intercepts + tolerance)[np.newaxis, :]
assert np.all(nl.rates(x, gain, bias) > threshold)
if generic:
max_rates0, intercepts0 = (
NeuronType.max_rates_intercepts(nl, gain, bias))
else:
max_rates0, intercepts0 = nl.max_rates_intercepts(gain, bias)
assert np.allclose(max_rates, max_rates0, atol=tolerance)
assert np.allclose(intercepts, intercepts0, atol=tolerance)
def test_gain_bias(rng, nl_nodirect, generic):
if nl_nodirect == Sigmoid and generic:
# the generic method doesn't work with sigmoid neurons (because they're
# always positive). that's not a failure, because the sigmoid neurons
# never need to use the generic method normally, so we'll just skip
# it for this test.
return
n = 100
max_rates = rng.uniform(300, 400, size=n)
intercepts = rng.uniform(-0.5, 0.5, size=n)
nl = nl_nodirect()
tolerance = 0.1 if generic else 1e-8
if generic:
gain, bias = NeuronType.gain_bias(nl, max_rates, intercepts)
else:
gain, bias = nl.gain_bias(max_rates, intercepts)
assert np.allclose(nl.rates(np.ones(n), gain, bias),
max_rates,
atol=tolerance)
if nl_nodirect == Sigmoid:
threshold = 0.5 / nl.tau_ref
else:
threshold = 0
assert np.all(nl.rates(intercepts - tolerance, gain, bias) <= threshold)
assert np.all(nl.rates(intercepts + tolerance, gain, bias) > threshold)
if generic:
max_rates0, intercepts0 = (NeuronType.max_rates_intercepts(
nl, gain, bias))
else:
max_rates0, intercepts0 = nl.max_rates_intercepts(gain, bias)
assert np.allclose(max_rates, max_rates0, atol=tolerance)
assert np.allclose(intercepts, intercepts0, atol=tolerance)