def test_optimizer_does_not_change_result(seed):
dtype_resolution = np.finfo(nengo.rc.float_dtype).resolution
dtype_decimal = int(np.floor(-np.log10(dtype_resolution) * 0.5))
model = learning_net()
model.seed = seed
with model:
# Add the default probe for every non-Probe object
probes = [
nengo.Probe(obj) for obj in model.all_objects
if not isinstance(obj, (nengo.Probe, nengo.Network))
]
# Also probe learning rules and neurons
probes.extend(nengo.Probe(ens.neurons) for ens in model.all_ensembles)
probes.extend(
nengo.Probe(conn.learning_rule) for conn in model.all_connections
if conn.learning_rule is not None)
with nengo.Simulator(model, optimize=False) as sim:
sim.run(0.1)
with nengo.Simulator(model, optimize=True) as sim_opt:
sim_opt.run(0.1)
for probe in probes:
assert_almost_equal(sim.data[probe],
sim_opt.data[probe],
decimal=dtype_decimal)
def test_unsupervised_exact_match(Simulator, learning_rule, seed, rng):
sim_time = 1.0
m, activity_p, trans_p = learning_net(
learning_rule, nengo.Network(seed=seed), rng)
refimpl_sim = nengo.Simulator(m)
refimpl_sim.run(sim_time)
mpi_sim = Simulator(m)
mpi_sim.run(sim_time)
assert np.allclose(
refimpl_sim.data[activity_p], mpi_sim.data[activity_p],
atol=0.00001, rtol=0.00)
assert np.allclose(
refimpl_sim.data[trans_p], mpi_sim.data[trans_p],
atol=0.00001, rtol=0.00)
refimpl_sim.reset()
mpi_sim.reset()
refimpl_sim.run(sim_time)
mpi_sim.run(sim_time)
assert np.allclose(
refimpl_sim.data[activity_p], mpi_sim.data[activity_p],
atol=0.00001, rtol=0.00)
assert np.allclose(
refimpl_sim.data[trans_p], mpi_sim.data[trans_p],
atol=0.00001, rtol=0.00)