def test_lmu(Simulator, native_nengo, pytestconfig):
n_steps = 1000
net = benchmarks.lmu(
n_steps, 1, native_nengo=native_nengo, dtype=pytestconfig.getoption("--dtype")
)
# TODO: It would be good to optimize LMU performance as the NengoDL implementation
# is a bit slower than the original TensorFlow implementation.
# benchmarks.run_profile(
# net,
# train=True,
# n_steps=n_steps if native_nengo else 1,
# do_profile=False,
# minibatch_size=100,
# unroll_simulation=25 if native_nengo else 1,
# reps=5,
# )
with Simulator(net) as sim:
n_trainable = sum(
np.prod(w.shape.as_list()) for w in sim.keras_model.trainable_weights
)
assert n_trainable == 102017
assert net.inp.size_out == 1 if native_nengo else n_steps
assert net.p.size_in == 10
(benchmarks.integrator(128, 32, nengo.LIF()), True, 64, 1.25, 1.45),
(
benchmarks.random_network(
64,
32,
nengo.RectifiedLinear(),
n_ensembles=20,
connections_per_ensemble=5,
seed=0,
),
False,
None,
0.5,
0.7,
),
(benchmarks.lmu(1000, 1, native_nengo=True), True, 100, 1.05, 1.25),
],
)
def test_performance(net, train, minibatch_size, min, max):
# performance is based on Azure NC6 VM
# CPU: Intel Xeon E5-2690 v3 @ 2.60Ghz
# GPU: Nvidia Tesla K80
# Python version: 3.6.10
# TensorFlow GPU version: 2.1.0
# Nengo version: 3.1.0
# NengoDL version: 3.1.0
time = benchmarks.run_profile(
net,
minibatch_size=minibatch_size,
train=train,
(
benchmarks.random_network(
64,
32,
nengo.RectifiedLinear(),
n_ensembles=20,
connections_per_ensemble=5,
seed=0,
),
False,
None,
True,
0.5,
0.7,
),
(benchmarks.lmu(1000, 1, native_nengo=True), True, 100, True, 1.3, 1.5),
(benchmarks.lmu(1000, 1, native_nengo=True), True, 100, False, 1.05, 1.25),
],
)
def test_performance(net, train, minibatch_size, eager, min, max):
# performance is based on Azure NC6 VM
# CPU: Intel Xeon E5-2690 v3 @ 2.60Ghz
# GPU: Nvidia Tesla K80
# Python version: 3.6.10
# TensorFlow GPU version: 2.3.0
# Nengo version: 3.1.0
# NengoDL version: 3.3.0
if not eager:
tf.compat.v1.disable_eager_execution()
tf.compat.v1.disable_control_flow_v2()
(benchmarks.integrator(128, 32, nengo.LIF()), True, 64, 0.95, 1.25),
(
benchmarks.random_network(
64,
32,
nengo.RectifiedLinear(),
n_ensembles=20,
connections_per_ensemble=5,
seed=0,
),
False,
None,
0.35,
0.55,
),
(benchmarks.lmu(1000, 1, native_nengo=True), True, 100, 0.85, 1.15),
],
)
def test_performance(net, train, minibatch_size, min, max):
# performance is based on ABR GPU server
# CPU: Intel Xeon E5-1650 v3 @ 3.50GHz
# GPU: GeForce GTX Titan X
# Python version: 3.6.8
# TensorFlow GPU version: 2.0.0
# Nengo version: 3.1.0
# NengoDL version: 3.1.0
time = benchmarks.run_profile(
net,
minibatch_size=minibatch_size,
train=train,