def test_tile(plt, rng, color):
n = 200
angles = rng.uniform(0, np.pi, size=n)
colors = rng.uniform(size=(n, 3))
genimg = lambda k: wave_image(
(30, 31), angles[k], color=colors[k] if color else None)
images = np.array([genimg(k) for k in range(n)])
ne_plt.tile(images, grid=True, gridcolor=(1, 1, 0))
def test_tile(plt, rng, color):
n = 200
angles = rng.uniform(0, np.pi, size=n)
colors = rng.uniform(size=(n, 3))
genimg = lambda k: wave_image(
(30, 31), angles[k], color=colors[k] if color else None)
images = np.array([genimg(k) for k in range(n)])
ne_plt.tile(images, grid=True, gridcolor=(1, 1, 0))
def show(images):
"""
Render a given numpy.uint8 2D array of pixel data.
"""
from matplotlib import pyplot
import matplotlib as mpl
from nengo_extras.matplotlib import tile
image_array = np.stack(images)
show_params = {'interpolation': 'nearest'}
tile(image_array, rows=4, cols=5, **show_params)
#ax.xaxis.set_ticks_position('top')
#ax.yaxis.set_ticks_position('left')
pyplot.show()
def make_gabors():
from nengo.dists import Uniform
from nengo_extras.matplotlib import tile
from nengo_extras.vision import Gabor
from hunse_thesis.dists import LogUniform
rng = np.random.RandomState(3)
# r, c = 10, 20
r, c = 9, 12
# gabor = Gabor()
# gabor = Gabor(freq=Uniform(0.5, 1.5))
gabor = Gabor(freq=LogUniform(np.log(0.5), np.log(1.5)))
gabors = gabor.generate(r * c, (32, 32), rng=rng)
tile(gabors, rows=r, cols=c)
plt.savefig('gabors.pdf')
#sigma_x=Tile(np.sort(np.repeat(np.linspace(0.2, 1.0, n_sigma), n_exc//n_sigma))),
#sigma_y=Tile(np.sort(np.repeat(np.linspace(0.2, 1.0, n_sigma), n_exc//n_sigma))),
input_encoders = Gabor(sigma_x=Choice([0.5]),
sigma_y=Choice([0.5]),
theta=Tile(np.sort(np.repeat(np.linspace(-np.pi,np.pi,n_gabor), n_exc//n_gabor))),
freq=Tile(np.sort(np.repeat(np.linspace(0.2,2.0,n_gabor), n_exc//n_gabor))),
phase=Tile(np.sort(np.repeat(np.linspace(-np.pi,np.pi,n_gabor), n_exc//n_gabor))),
).generate(n_exc, gabor_size, rng=rng)
input_encoders = SequentialMask((n_x, n_y)).populate(input_encoders, rng=rng, flatten=True)
#input_encoders = rng.normal(size=(n_exc, n_x * n_y))
#input_encoders = rng.normal(size=(n_exc, 5, 5))
#input_encoders = SequentialMask((n_x, n_y)).populate(input_encoders, rng=rng, random_positions=True, flatten=True)
print(f'input_encoders.shape = {input_encoders.shape}')
tile(input_encoders.reshape((-1, n_x, n_y)), rows=10, cols=n_gabor, grid=True)
plt.show()
model_dict = build_network(params, dimensions=input_dimensions, inputs=input_data,
input_encoders=input_encoders, direct_input=False,
presentation_time=presentation_time, pause_time=pause_time,
coords=coords_dict,
t_learn_exc=t_train, t_learn_inh=t_train,
sample_weights_every=t_end // 5.0)
network = model_dict['network']
plt.imshow(network.srf_network.weights_initial_E.T, aspect='auto', interpolation='nearest')
plt.colorbar(label='Synaptic weight')
plt.show()
def test_load_mnist(plt):
(trainX, _), (_, _) = load_mnist()
tile(trainX.reshape(-1, 28, 28))
def test_conv_split(Simulator, rng, plt, allclose):
channels_last = False
# load data
with open(os.path.join(test_dir, 'mnist10.pkl'), 'rb') as f:
test10 = pickle.load(f)
input_shape = nengo_transforms.ChannelShape((1, 28, 28),
channels_last=channels_last)
n_filters = 8
kernel_size = (7, 7)
kernel = Gabor(freq=Uniform(0.5, 1)).generate(n_filters,
kernel_size,
rng=rng)
kernel = kernel[None, :, :, :] # single channel
kernel = np.transpose(kernel, (2, 3, 0, 1))
strides = (2, 2)
seed = 3 # fix seed to do the same computation for both channel positions
with nengo.Network(seed=seed) as net:
nengo_loihi.add_params(net)
a = nengo.Node(test10[0][0].ravel())
# --- make population to turn image into spikes
nc = 1
in_kernel = np.array([1.]).reshape((1, 1, 1, nc))
transform = nengo_transforms.Convolution(1,
input_shape,
kernel_size=(1, 1),
init=in_kernel,
channels_last=channels_last)
b = nengo.Ensemble(transform.output_shape.size,
1,
neuron_type=nengo.SpikingRectifiedLinear(),
max_rates=nengo.dists.Choice([50]),
intercepts=nengo.dists.Choice([0]))
net.config[b].on_chip = False
nengo.Connection(a, b.neurons, transform=transform)
in_shape = transform.output_shape
transform = nengo_transforms.Convolution(n_filters,
in_shape,
kernel_size=kernel_size,
strides=strides,
init=kernel,
channels_last=channels_last)
out_shape = transform.output_shape
split_slices = conv.split_channels(out_shape,
max_size=1024,
max_channels=4)
# --- make convolution population, split across ensembles
cc = []
cp = []
out_shapes = []
xslice = conv.ImageSlice(in_shape)
for yslice in split_slices:
transform_xy = conv.split_transform(transform, xslice, yslice)
out_shapes.append(transform_xy.output_shape)
c = nengo.Ensemble(transform_xy.output_shape.size,
1,
neuron_type=nengo.LIF(),
max_rates=nengo.dists.Choice([15]),
intercepts=nengo.dists.Choice([0]))
nengo.Connection(b.neurons, c.neurons, transform=transform_xy)
cc.append(c)
cp.append(nengo.Probe(c.neurons))
simtime = 0.3
with nengo.Simulator(net, optimize=False) as sim_nengo:
sim_nengo.run(simtime)
hw_opts = dict(snip_max_spikes_per_step=100)
with Simulator(net, seed=seed, hardware_options=hw_opts) as sim_loihi:
sim_loihi.run(simtime)
nengo_out = []
loihi_out = []
for p, out_shape_i in zip(cp, out_shapes):
nengo_out.append(
(sim_nengo.data[p] > 0).sum(axis=0).reshape(out_shape_i.shape))
loihi_out.append(
(sim_loihi.data[p] > 0).sum(axis=0).reshape(out_shape_i.shape))
if channels_last:
nengo_out = np.concatenate(nengo_out, axis=2)
loihi_out = np.concatenate(loihi_out, axis=2)
# put channels first to display them separately
nengo_out = np.transpose(nengo_out, (2, 0, 1))
loihi_out = np.transpose(loihi_out, (2, 0, 1))
else:
nengo_out = np.concatenate(nengo_out, axis=0)
loihi_out = np.concatenate(loihi_out, axis=0)
out_max = np.maximum(nengo_out.max(), loihi_out.max())
# --- plot results
rows = 2
cols = 3
ax = plt.subplot(rows, cols, 1)
imshow(test10[0][0].reshape((28, 28)), vmin=0, vmax=1, ax=ax)
ax = plt.subplot(rows, cols, 2)
tile(np.transpose(kernel[0], (2, 0, 1)), cols=8, ax=ax)
ax = plt.subplot(rows, cols, 3)
plt.hist(nengo_out.ravel(), bins=31)
plt.hist(loihi_out.ravel(), bins=31)
ax = plt.subplot(rows, cols, 4)
tile(nengo_out, vmin=0, vmax=out_max, cols=8, ax=ax)
ax = plt.subplot(rows, cols, 6)
tile(loihi_out, vmin=0, vmax=out_max, cols=8, ax=ax)
assert allclose(loihi_out, nengo_out, atol=0.05 * out_max, rtol=0.15)
def test_conv_connection(channels, channels_last, Simulator, seed, rng, plt,
allclose):
if channels_last:
plt.saveas = None
pytest.xfail("Blocked by CxBase cannot be > 256 bug")
# load data
with open(os.path.join(test_dir, 'mnist10.pkl'), 'rb') as f:
test10 = pickle.load(f)
test_x = test10[0][0].reshape(28, 28)
test_x = 1.999 * test_x - 0.999 # range (-1, 1)
input_shape = nengo_transforms.ChannelShape(
(test_x.shape + (channels, )) if channels_last else
((channels, ) + test_x.shape),
channels_last=channels_last)
filters = Gabor(freq=Uniform(0.5, 1)).generate(8, (7, 7), rng=rng)
filters = filters[None, :, :, :] # single channel
filters = np.transpose(filters, (2, 3, 0, 1))
strides = (2, 2)
tau_rc = 0.02
tau_ref = 0.002
tau_s = 0.005
dt = 0.001
neuron_type = LoihiLIF(tau_rc=tau_rc, tau_ref=tau_ref)
pres_time = 0.1
with nengo.Network(seed=seed) as model:
nengo_loihi.add_params(model)
u = nengo.Node(test_x.ravel(), label="u")
a = nengo.Ensemble(input_shape.size,
1,
neuron_type=LoihiSpikingRectifiedLinear(),
max_rates=nengo.dists.Choice([40 / channels]),
intercepts=nengo.dists.Choice([0]),
label='a')
model.config[a].on_chip = False
if channels == 1:
nengo.Connection(u, a.neurons, transform=1, synapse=None)
elif channels == 2:
# encode image into spikes using two channels (on/off)
if input_shape.channels_last:
nengo.Connection(u, a.neurons[0::2], transform=1, synapse=None)
nengo.Connection(u,
a.neurons[1::2],
transform=-1,
synapse=None)
else:
k = input_shape.spatial_shape[0] * input_shape.spatial_shape[1]
nengo.Connection(u, a.neurons[:k], transform=1, synapse=None)
nengo.Connection(u, a.neurons[k:], transform=-1, synapse=None)
filters = np.concatenate([filters, -filters], axis=2)
else:
raise ValueError("Test not configured for more than two channels")
conv2d_transform = nengo_transforms.Convolution(
8,
input_shape,
strides=strides,
kernel_size=(7, 7),
channels_last=channels_last,
init=filters)
output_shape = conv2d_transform.output_shape
gain, bias = neuron_type.gain_bias(max_rates=100, intercepts=0)
gain = gain * 0.01 # account for `a` max_rates
b = nengo.Ensemble(output_shape.size,
1,
neuron_type=neuron_type,
gain=nengo.dists.Choice([gain[0]]),
bias=nengo.dists.Choice([bias[0]]),
label='b')
nengo.Connection(a.neurons,
b.neurons,
synapse=tau_s,
transform=conv2d_transform)
bp = nengo.Probe(b.neurons)
with nengo.Simulator(model, dt=dt, optimize=False) as sim:
sim.run(pres_time)
ref_out = sim.data[bp].mean(axis=0).reshape(output_shape.shape)
# Currently, non-gpu TensorFlow does not support channels first in conv
use_nengo_dl = HAS_DL and channels_last
ndl_out = np.zeros_like(ref_out)
if use_nengo_dl:
with nengo_dl.Simulator(model, dt=dt) as sim_dl:
sim_dl.run(pres_time)
ndl_out = sim_dl.data[bp].mean(axis=0).reshape(output_shape.shape)
with nengo_loihi.Simulator(model, dt=dt, target='simreal') as sim_real:
sim_real.run(pres_time)
real_out = sim_real.data[bp].mean(axis=0).reshape(output_shape.shape)
with Simulator(model, dt=dt) as sim_loihi:
if "loihi" in sim_loihi.sims:
sim_loihi.sims["loihi"].snip_max_spikes_per_step = 800
sim_loihi.run(pres_time)
sim_out = sim_loihi.data[bp].mean(axis=0).reshape(output_shape.shape)
if not output_shape.channels_last:
ref_out = np.transpose(ref_out, (1, 2, 0))
ndl_out = np.transpose(ndl_out, (1, 2, 0))
real_out = np.transpose(real_out, (1, 2, 0))
sim_out = np.transpose(sim_out, (1, 2, 0))
out_max = max(ref_out.max(), sim_out.max())
# --- plot results
rows = 2
cols = 3
ax = plt.subplot(rows, cols, 1)
imshow(test_x, vmin=0, vmax=1, ax=ax)
ax = plt.subplot(rows, cols, 2)
tile(np.transpose(filters[0], (2, 0, 1)), cols=8, ax=ax)
ax = plt.subplot(rows, cols, 3)
plt.hist(ref_out.ravel(), bins=31)
plt.hist(sim_out.ravel(), bins=31)
ax = plt.subplot(rows, cols, 4)
tile(np.transpose(ref_out, (2, 0, 1)), vmin=0, vmax=out_max, cols=8, ax=ax)
ax = plt.subplot(rows, cols, 5)
tile(np.transpose(ndl_out, (2, 0, 1)), vmin=0, vmax=out_max, cols=8, ax=ax)
ax = plt.subplot(rows, cols, 6)
tile(np.transpose(sim_out, (2, 0, 1)), vmin=0, vmax=out_max, cols=8, ax=ax)
if use_nengo_dl:
assert allclose(ndl_out, ref_out, atol=1e-5, rtol=1e-5)
assert allclose(real_out, ref_out, atol=1, rtol=1e-3)
assert allclose(sim_out, ref_out, atol=10, rtol=1e-3)
def test_pop_tiny(pop_type, out_channels_last, request, plt, seed, rng,
allclose):
nc = 2
tau_rc = 0.02
tau_ref = 0.001
tau_s = 0.0
dt = 0.001
neuron_bias = 1.
pres_time = 0.4
sti, stj = 1, 1
if nc == 1:
filters = np.array([[-0.5, 2., -0.25], [-0.75, 2., -1.0],
[-0.5, 3., -0.5], [-1.0, 6.,
-0.25]]).reshape(1, 4, 1, 3)
filters = np.transpose(filters, (0, 2, 3, 1))
test_x = np.array([[1, 5, 1], [2, 1, 2]])
test_x = test_x[:, :, None]
elif nc == 2:
filters = np.array([[[-0.5, 2., -0.2], [-0.7, 2., -1.0],
[-0.5, 3., -0.5], [-1.0, 6., -0.2]],
[[-1.0, 2., -1.0], [-0.5, 2., -0.5],
[-0.8, 3., -0.2], [-1.0, 4.,
-0.2]]]).reshape(2, 4, 1, 3)
filters = np.transpose(filters, (0, 2, 3, 1))
test_x = np.array([[[1, 5, 1], [2, 1, 2]], [[0, 3, 1], [4, 2, 1]]])
test_x = np.transpose(test_x, (1, 2, 0))
test_x = test_x / (test_x.max() + 0.001)
# --- compute nengo_loihi outputs
inp_biases = test_x
inp_shape = ImageShape.from_shape(inp_biases.shape, channels_last=True)
ni, nj, nk = inp_shape.shape(channels_last=True)
nc, si, sj, nf = filters.shape
nij = ni * nj
nyi = 1 + (ni - si) // sti
nyj = 1 + (nj - sj) // stj
out_size = nyi * nyj * nf
assert out_size <= 1024
model = loihi_cx.CxModel()
# input group
inp = loihi_cx.CxGroup(ni * nj * nk, label='inp')
assert inp.n <= 1024
inp.configure_relu()
inp.bias[:] = inp_biases.ravel()
inp_ax = loihi_cx.CxAxons(nij, label='inp_ax')
inp_ax.set_axon_map(inp_shape.pixel_idxs(), inp_shape.channel_idxs())
inp.add_axons(inp_ax)
model.add_group(inp)
# conv group
neurons = loihi_cx.CxGroup(out_size, label='neurons')
assert neurons.n <= 1024
neurons.configure_lif(tau_rc=tau_rc, tau_ref=tau_ref, dt=dt)
neurons.configure_filter(tau_s, dt=dt)
neurons.bias[:] = neuron_bias
synapses = loihi_cx.CxSynapses(inp_shape.n_pixels, label='synapses')
conv2d_transform = Conv2D.from_kernel(
filters,
inp_shape,
strides=(sti, stj),
output_channels_last=out_channels_last)
weights, indices, axon_to_weight_map, cx_bases = conv2d_loihi_weights(
conv2d_transform)
synapses.set_population_weights(weights,
indices,
axon_to_weight_map,
cx_bases,
pop_type=pop_type)
neurons.add_synapses(synapses)
out_probe = loihi_cx.CxProbe(target=neurons, key='s')
neurons.add_probe(out_probe)
inp_ax.target = synapses
model.add_group(neurons)
# simulation
model.discretize()
n_steps = int(pres_time / dt)
target = request.config.getoption("--target")
if target == 'loihi':
with LoihiSimulator(model, use_snips=False, seed=seed) as sim:
sim.run_steps(n_steps)
sim_out = sim.get_probe_output(out_probe)
else:
with CxSimulator(model, seed=seed) as sim:
sim.run_steps(n_steps)
sim_out = sim.get_probe_output(out_probe)
sim_out = np.sum(sim_out, axis=0) * (dt / pres_time)
if out_channels_last:
sim_out.shape = (nyi, nyj, nf)
sim_out = np.transpose(sim_out, (2, 0, 1))
else:
sim_out.shape = (nf, nyi, nyj)
out_max = sim_out.max()
# --- plot results
rows = 1
cols = 2
ax = plt.subplot(rows, cols, 1)
plt.hist(sim_out.ravel(), bins=11)
ax = plt.subplot(rows, cols, 2)
tile(sim_out, vmin=0, vmax=out_max, grid=True, ax=ax)
print("sim_out:\n%r" % (sim_out[:, :, 0], ))
# ref_out determined by emulator running code known to work
if nc == 1:
ref_out = np.array([[0.06, 0.02], [0.055, 0.], [0.0825, 0.0225],
[0.125, 0.04]])
elif nc == 2:
ref_out = np.array([[0.0975, 0.02], [0.0825, 0.02], [0.125, 0.055],
[0.2475, 0.0825]])
assert allclose(sim_out[:, :, 0], ref_out, rtol=0, atol=1e-7)
def test_conv2d_weights(request, plt, seed, rng, allclose):
pop_type = 32
out_channels_last = False
# load data
with open(os.path.join(test_dir, 'mnist10.pkl'), 'rb') as f:
test10 = pickle.load(f)
test_x = test10[0][0].reshape(28, 28)
test_x = test_x[3:24, 3:24]
test_x = 1.999 * test_x - 0.999
filters = Gabor(freq=Uniform(0.5, 1)).generate(8, (7, 7), rng=rng)
sti, stj = 2, 2
tau_rc = 0.02
tau_ref = 0.002
tau_s = 0.005
dt = 0.001
encode_type = nengo.SpikingRectifiedLinear()
encode_gain = 1. / dt
encode_bias = 0.
neuron_type = nengo.LIF(tau_rc=tau_rc, tau_ref=tau_ref)
neuron_gain = 1.
neuron_bias = 1.
pres_time = 0.2
# --- compute ideal outputs
def conv_pm(x, kernel):
y0 = scipy.signal.correlate2d(x[0], kernel, mode='valid')[::sti, ::stj]
y1 = scipy.signal.correlate2d(x[1], kernel, mode='valid')[::sti, ::stj]
return [y0, -y1]
ref_out = np.array([test_x, -test_x])
ref_out = loihi_rates(encode_type, ref_out, encode_gain, encode_bias, dt)
ref_out = ref_out / encode_gain
ref_out = np.array([conv_pm(ref_out, kernel) for kernel in filters])
ref_out = ref_out.sum(axis=1) # sum positive and negative parts
ref_out = loihi_rates(neuron_type, ref_out, neuron_gain, neuron_bias, dt)
# --- compute nengo_loihi outputs
inp_biases = np.array([test_x, -test_x])
inp_shape = ImageShape.from_shape(inp_biases.shape, channels_last=False)
kernel = np.array([filters, -filters]) # two channels, pos and neg
kernel = np.transpose(kernel, (0, 2, 3, 1))
conv2d_transform = Conv2D.from_kernel(
kernel,
inp_shape,
strides=(sti, stj),
output_channels_last=out_channels_last)
ni, nj, nk = inp_shape.shape(channels_last=True)
out_size = ref_out.size
nf, nyi, nyj = ref_out.shape
assert out_size <= 1024
model = loihi_cx.CxModel()
# input group
inp = loihi_cx.CxGroup(inp_shape.size, label='inp')
assert inp.n <= 1024
inp.configure_relu()
inp.bias[:] = inp_biases.ravel()
inp_ax = loihi_cx.CxAxons(inp_shape.n_pixels, label='inp_ax')
inp_ax.set_axon_map(inp_shape.pixel_idxs(), inp_shape.channel_idxs())
inp.add_axons(inp_ax)
model.add_group(inp)
# conv group
neurons = loihi_cx.CxGroup(out_size, label='neurons')
assert neurons.n <= 1024
neurons.configure_lif(tau_rc=tau_rc, tau_ref=tau_ref, dt=dt)
neurons.configure_filter(tau_s, dt=dt)
neurons.bias[:] = neuron_bias
synapses = loihi_cx.CxSynapses(inp_shape.n_pixels, label='synapses')
weights, indices, axon_to_weight_map, cx_bases = conv2d_loihi_weights(
conv2d_transform)
synapses.set_population_weights(weights,
indices,
axon_to_weight_map,
cx_bases,
pop_type=pop_type)
neurons.add_synapses(synapses)
out_probe = loihi_cx.CxProbe(target=neurons, key='s')
neurons.add_probe(out_probe)
inp_ax.target = synapses
model.add_group(neurons)
# simulation
model.discretize()
n_steps = int(pres_time / dt)
target = request.config.getoption("--target")
if target == 'loihi':
with LoihiSimulator(model, use_snips=False, seed=seed) as sim:
sim.run_steps(n_steps)
sim_out = sim.get_probe_output(out_probe)
else:
with CxSimulator(model, seed=seed) as sim:
sim.run_steps(n_steps)
sim_out = sim.get_probe_output(out_probe)
sim_out = np.sum(sim_out, axis=0) / pres_time
if out_channels_last:
sim_out.shape = (nyi, nyj, nf)
sim_out = np.transpose(sim_out, (2, 0, 1))
else:
sim_out.shape = (nf, nyi, nyj)
out_max = max(ref_out.max(), sim_out.max())
# --- plot results
rows = 2
cols = 2
ax = plt.subplot(rows, cols, 1)
tile(filters, cols=8, ax=ax)
ax = plt.subplot(rows, cols, 2)
tile(ref_out, vmin=0, vmax=out_max, cols=8, ax=ax)
ax = plt.subplot(rows, cols, 3)
plt.hist(ref_out.ravel(), bins=31)
plt.hist(sim_out.ravel(), bins=31)
ax = plt.subplot(rows, cols, 4)
# tile(sim_out, vmin=0, vmax=1, cols=8, ax=ax)
tile(sim_out, vmin=0, vmax=out_max, cols=8, ax=ax)
assert allclose(sim_out, ref_out, atol=10, rtol=1e-3)
def test_load_svhn(plt):
(trainX, _), (_, _) = load_svhn()
trainX = trainX.reshape((-1, 3, 32, 32))
trainX = np.transpose(trainX, (0, 2, 3, 1))
tile(trainX, ax=plt.gca())
def test_load_mnist(plt):
trainX = load_mnist()[0][0]
tile(trainX.reshape((-1, 28, 28)), ax=plt.gca())
def test_load_ilsvrc2012(plt):
testX, _, _, _ = load_ilsvrc2012(n_files=1)
testX = testX.reshape((-1, 3, 256, 256))
testX = np.transpose(testX, (0, 2, 3, 1))
tile(testX, ax=plt.gca())
def test_load_cifar100(plt):
trainX = load_cifar100()[0][0]
trainX = trainX.reshape((-1, 3, 32, 32))
trainX = np.transpose(trainX, (0, 2, 3, 1))
tile(trainX, ax=plt.gca())
import matplotlib.pyplot as plt
import numpy as np
import nengo_extras.data
from nengo_extras.matplotlib import tile
plt.figure(figsize=(6.4, 6.4), dpi=200)
plt.subplot(511)
(mnist, _), _ = nengo_extras.data.load_mnist()
tile(mnist.reshape(-1, 28, 28), rows=3, cols=20)
plt.title('MNIST')
plt.subplot(512)
(svhn, _), _ = nengo_extras.data.load_svhn()
tile(np.transpose(svhn.reshape(-1, 3, 32, 32), (0, 2, 3, 1)), rows=3, cols=20)
plt.title('SVHN')
plt.subplot(513)
(cifar10, _), _ = nengo_extras.data.load_cifar10(n_train=1, n_test=0)
tile(np.transpose(cifar10.reshape(-1, 3, 32, 32), (0, 2, 3, 1)), rows=3, cols=20)
plt.title('CIFAR-10')
plt.subplot(514)
(cifar100, _), _ = nengo_extras.data.load_cifar100()
tile(np.transpose(cifar100.reshape(-1, 3, 32, 32), (0, 2, 3, 1)), rows=3, cols=20)
plt.title('CIFAR-100')
plt.subplot(515)
ilsvrc2012, _, _, _ = nengo_extras.data.load_ilsvrc2012(n_files=1)
tile(np.transpose(ilsvrc2012.reshape(-1, 3, 256, 256), (0, 2, 3, 1)), rows=2, cols=14)
import nengo_extras.data
from nengo_extras.matplotlib import tile
images, labels, data_mean, label_names = nengo_extras.data.load_ilsvrc2012()
# images, labels, data_mean, label_names = (
# nengo_extras.data.load_ilsvrc2012(n_files=1))
for s in label_names:
print(s)
plt.figure()
# show = 'butcher'
show_names = ['butcher', 'restaurant', 'planetarium', 'church', 'library']
for i, show_name in enumerate(show_names):
match = [s.startswith(show_name) for s in label_names]
index = match.index(True) if True in match else None
mask = (labels == index) if index is not None else None
count = mask.sum() if mask is not None else None
if count is not None:
print("Found %d matches for %r: showing index %d, %d examples" %
(sum(match), show_name, index, count))
plt.subplot(len(show_names), 1, i + 1)
tile(np.transpose(images[mask].reshape(-1, 3, 256, 256), (0, 2, 3, 1)),
rows=3,
cols=5)
plt.show()
def test_conv2d_weights(channels_last, hw_opts, request, plt, seed, rng,
allclose):
def loihi_rates_n(neuron_type, x, gain, bias, dt):
"""Compute Loihi rates on higher dimensional inputs"""
y = x.reshape(-1, x.shape[-1])
gain = np.asarray(gain)
bias = np.asarray(bias)
if gain.ndim == 0:
gain = gain * np.ones(x.shape[-1])
if bias.ndim == 0:
bias = bias * np.ones(x.shape[-1])
rates = loihi_rates(neuron_type, y, gain, bias, dt)
return rates.reshape(*x.shape)
if channels_last:
plt.saveas = None
pytest.xfail("Blocked by CxBase cannot be > 256 bug")
target = request.config.getoption("--target")
if target != 'loihi' and len(hw_opts) > 0:
pytest.skip("Hardware options only available on hardware")
pop_type = 32
# load data
with open(os.path.join(test_dir, 'mnist10.pkl'), 'rb') as f:
test10 = pickle.load(f)
test_x = test10[0][0].reshape(28, 28)
test_x = test_x[3:24, 3:24]
test_x = 1.999 * test_x - 0.999
filters = Gabor(freq=Uniform(0.5, 1)).generate(8, (7, 7), rng=rng)
sti, stj = 2, 2
tau_rc = 0.02
tau_ref = 0.002
tau_s = 0.005
dt = 0.001
encode_type = nengo.SpikingRectifiedLinear()
encode_gain = 1. / dt
encode_bias = 0.
neuron_type = nengo.LIF(tau_rc=tau_rc, tau_ref=tau_ref)
neuron_gain = 1.
neuron_bias = 1.
pres_time = 0.2
# --- compute ideal outputs
def conv_pm(x, kernel):
y0 = scipy.signal.correlate2d(x[0], kernel, mode='valid')[::sti, ::stj]
y1 = scipy.signal.correlate2d(x[1], kernel, mode='valid')[::sti, ::stj]
return [y0, -y1]
ref_out = np.array([test_x, -test_x])
ref_out = loihi_rates_n(encode_type, ref_out, encode_gain, encode_bias, dt)
ref_out = ref_out / encode_gain
ref_out = np.array([conv_pm(ref_out, kernel) for kernel in filters])
ref_out = ref_out.sum(axis=1) # sum positive and negative parts
ref_out = loihi_rates_n(neuron_type, ref_out, neuron_gain, neuron_bias, dt)
# --- compute nengo_loihi outputs
inp_biases = np.stack([test_x, -test_x], axis=-1 if channels_last else 0)
inp_shape = nengo_transforms.ChannelShape(inp_biases.shape,
channels_last=channels_last)
kernel = np.array([filters, -filters]) # two channels, pos and neg
kernel = np.transpose(kernel, (2, 3, 0, 1))
conv2d_transform = nengo_transforms.Convolution(
8,
inp_shape,
strides=(sti, stj),
channels_last=channels_last,
kernel_size=(7, 7),
init=kernel)
out_size = ref_out.size
nf, nyi, nyj = ref_out.shape
assert out_size <= 1024
model = Model()
# input block
inp = LoihiBlock(inp_shape.size, label='inp')
assert inp.n_neurons <= 1024
inp.compartment.configure_relu()
inp.compartment.bias[:] = inp_biases.ravel()
inp_ax = Axon(np.prod(inp_shape.spatial_shape), label='inp_ax')
inp_ax.set_compartment_axon_map(target_axons=conv.pixel_idxs(inp_shape),
atoms=conv.channel_idxs(inp_shape))
inp.add_axon(inp_ax)
model.add_block(inp)
# conv block
neurons = LoihiBlock(out_size, label='neurons')
assert neurons.n_neurons <= 1024
neurons.compartment.configure_lif(tau_rc=tau_rc, tau_ref=tau_ref, dt=dt)
neurons.compartment.configure_filter(tau_s, dt=dt)
neurons.compartment.bias[:] = neuron_bias
synapse = Synapse(np.prod(inp_shape.spatial_shape), label='synapse')
weights, indices, axon_to_weight_map, bases = conv.conv2d_loihi_weights(
conv2d_transform)
synapse.set_population_weights(weights,
indices,
axon_to_weight_map,
bases,
pop_type=pop_type)
neurons.add_synapse(synapse)
out_probe = Probe(target=neurons, key='spiked')
neurons.add_probe(out_probe)
inp_ax.target = synapse
model.add_block(neurons)
# simulation
discretize_model(model)
n_steps = int(pres_time / dt)
if target == 'loihi':
with HardwareInterface(model, use_snips=False, seed=seed,
**hw_opts) as sim:
sim.run_steps(n_steps)
sim_out = sim.get_probe_output(out_probe)
else:
with EmulatorInterface(model, seed=seed) as sim:
sim.run_steps(n_steps)
sim_out = sim.get_probe_output(out_probe)
sim_out = np.sum(sim_out, axis=0) / pres_time
if channels_last:
sim_out.shape = (nyi, nyj, nf)
sim_out = np.transpose(sim_out, (2, 0, 1))
else:
sim_out.shape = (nf, nyi, nyj)
out_max = max(ref_out.max(), sim_out.max())
# --- plot results
rows = 2
cols = 2
ax = plt.subplot(rows, cols, 1)
tile(filters, cols=8, ax=ax)
ax = plt.subplot(rows, cols, 2)
tile(ref_out, vmin=0, vmax=out_max, cols=8, ax=ax)
ax = plt.subplot(rows, cols, 3)
plt.hist(ref_out.ravel(), bins=31)
plt.hist(sim_out.ravel(), bins=31)
ax = plt.subplot(rows, cols, 4)
# tile(sim_out, vmin=0, vmax=1, cols=8, ax=ax)
tile(sim_out, vmin=0, vmax=out_max, cols=8, ax=ax)
assert allclose(sim_out, ref_out, atol=10, rtol=1e-3)
def test_conv_connection(channels, Simulator, seed, rng, plt, allclose):
# channels_last = True
channels_last = False
if channels > 1:
pytest.xfail("Cannot send population spikes to chip")
# load data
with open(os.path.join(test_dir, 'mnist10.pkl'), 'rb') as f:
test10 = pickle.load(f)
test_x = test10[0][0].reshape(28, 28)
test_x = 1.999 * test_x - 0.999 # range (-1, 1)
test_x = test_x[:, :, None] # single channel
input_shape = ImageShape(test_x.shape[0],
test_x.shape[1],
channels,
channels_last=channels_last)
filters = Gabor(freq=Uniform(0.5, 1)).generate(8, (7, 7), rng=rng)
filters = filters[None, :, :, :] # single channel
filters = np.transpose(filters, (0, 2, 3, 1)) # filters last
strides = (2, 2)
tau_rc = 0.02
tau_ref = 0.002
tau_s = 0.005
dt = 0.001
neuron_type = LoihiLIF(tau_rc=tau_rc, tau_ref=tau_ref)
pres_time = 0.1
with nengo.Network(seed=seed) as model:
nengo_loihi.add_params(model)
u = nengo.Node(nengo.processes.PresentInput([test_x.ravel()],
pres_time),
label='u')
a = nengo.Ensemble(input_shape.size,
1,
neuron_type=LoihiSpikingRectifiedLinear(),
max_rates=nengo.dists.Choice([40 / channels]),
intercepts=nengo.dists.Choice([0]),
label='a')
model.config[a].on_chip = False
if channels == 1:
nengo.Connection(u, a.neurons, transform=1, synapse=None)
elif channels == 2:
# encode image into spikes using two channels (on/off)
if input_shape.channels_last:
nengo.Connection(u, a.neurons[0::2], transform=1, synapse=None)
nengo.Connection(u,
a.neurons[1::2],
transform=-1,
synapse=None)
else:
k = input_shape.rows * input_shape.cols
nengo.Connection(u, a.neurons[:k], transform=1, synapse=None)
nengo.Connection(u, a.neurons[k:], transform=-1, synapse=None)
filters = np.vstack([filters, -filters])
else:
raise ValueError("Test not configured for more than two channels")
conv2d_transform = Conv2D.from_kernel(filters,
input_shape,
strides=strides)
output_shape = conv2d_transform.output_shape
gain, bias = neuron_type.gain_bias(max_rates=100, intercepts=0)
gain = gain * 0.01 # account for `a` max_rates
b = nengo.Ensemble(output_shape.size,
1,
neuron_type=neuron_type,
gain=nengo.dists.Choice([gain[0]]),
bias=nengo.dists.Choice([bias[0]]),
label='b')
nengo.Connection(a.neurons,
b.neurons,
synapse=tau_s,
transform=conv2d_transform)
bp = nengo.Probe(b.neurons)
with nengo.Simulator(model, dt=dt, optimize=False) as sim:
sim.run(pres_time)
ref_out = sim.data[bp].mean(axis=0).reshape(output_shape.shape())
# Currently, default TensorFlow does not support channels first in conv
use_nengo_dl = nengo_dl is not None and channels_last
ndl_out = np.zeros_like(ref_out)
if use_nengo_dl:
with nengo_dl.Simulator(model, dt=dt) as sim:
sim.run(pres_time)
ndl_out = sim.data[bp].mean(axis=0).reshape(output_shape.shape())
with nengo_loihi.Simulator(model, dt=dt, target='simreal') as sim:
sim.run(pres_time)
real_out = sim.data[bp].mean(axis=0).reshape(output_shape.shape())
with Simulator(model, dt=dt) as sim:
sim.run(pres_time)
sim_out = sim.data[bp].mean(axis=0).reshape(output_shape.shape())
if not output_shape.channels_last:
ref_out = np.transpose(ref_out, (1, 2, 0))
ndl_out = np.transpose(ndl_out, (1, 2, 0))
real_out = np.transpose(real_out, (1, 2, 0))
sim_out = np.transpose(sim_out, (1, 2, 0))
out_max = max(ref_out.max(), sim_out.max())
# --- plot results
rows = 2
cols = 3
ax = plt.subplot(rows, cols, 1)
imshow(test_x, vmin=0, vmax=1, ax=ax)
ax = plt.subplot(rows, cols, 2)
tile(np.transpose(filters[0], (2, 0, 1)), cols=8, ax=ax)
ax = plt.subplot(rows, cols, 3)
plt.hist(ref_out.ravel(), bins=31)
plt.hist(sim_out.ravel(), bins=31)
ax = plt.subplot(rows, cols, 4)
tile(np.transpose(ref_out, (2, 0, 1)), vmin=0, vmax=out_max, cols=8, ax=ax)
ax = plt.subplot(rows, cols, 5)
tile(np.transpose(ndl_out, (2, 0, 1)), vmin=0, vmax=out_max, cols=8, ax=ax)
ax = plt.subplot(rows, cols, 6)
tile(np.transpose(sim_out, (2, 0, 1)), vmin=0, vmax=out_max, cols=8, ax=ax)
if use_nengo_dl:
assert allclose(ndl_out, ref_out, atol=1e-5, rtol=1e-5)
assert allclose(real_out, ref_out, atol=1, rtol=1e-3)
assert allclose(sim_out, ref_out, atol=10, rtol=1e-3)
def test_pop_tiny(pop_type, channels_last, nc, request, plt, seed, allclose):
tau_rc = 0.02
tau_ref = 0.001
tau_s = 0.0
dt = 0.001
neuron_bias = 1.
pres_time = 0.4
sti, stj = 1, 1
if nc == 1:
filters = np.array([[-0.5, 2., -0.25], [-0.75, 2., -1.0],
[-0.5, 3., -0.5], [-1.0, 6.,
-0.25]]).reshape(1, 4, 1, 3)
inp_biases = np.array([[1, 5, 1], [2, 1, 2]])
inp_biases = inp_biases[:, :, None]
elif nc == 2:
filters = np.array([[[-0.5, 2., -0.2], [-0.7, 2., -1.0],
[-0.5, 3., -0.5], [-1.0, 6., -0.2]],
[[-1.0, 2., -1.0], [-0.5, 2., -0.5],
[-0.8, 3., -0.2], [-1.0, 4.,
-0.2]]]).reshape(2, 4, 1, 3)
inp_biases = np.array([[[1, 5, 1], [2, 1, 2]], [[0, 3, 1], [4, 2, 1]]])
inp_biases = np.transpose(inp_biases, (1, 2, 0))
# rearrange to (kernel_rows, kernel_cols, in_channels, out_channels)
filters = np.transpose(filters, (2, 3, 0, 1))
inp_biases = inp_biases / (inp_biases.max() + 0.001)
# --- compute nengo_loihi outputs
ni, nj, nk = inp_biases.shape
si, sj, nc, nf = filters.shape
nij = ni * nj
nyi = 1 + (ni - si) // sti
nyj = 1 + (nj - sj) // stj
out_size = nyi * nyj * nf
assert out_size <= 1024
model = Model()
# input block
inp = LoihiBlock(ni * nj * nk, label='inp')
assert inp.n_neurons <= 1024
inp.compartment.configure_relu()
inp.compartment.bias[:] = inp_biases.ravel()
inp_ax = Axon(nij, label='inp_ax')
# we always compute the pixel/channel idxs with channels_last=True
# (not sure why?), and then set it to the correct value afterwards
inp_shape = nengo_transforms.ChannelShape((ni, nj, nk), channels_last=True)
inp_ax.set_compartment_axon_map(target_axons=conv.pixel_idxs(inp_shape),
atoms=conv.channel_idxs(inp_shape))
inp_shape.shape = (ni, nj, nk) if channels_last else (nk, ni, nj)
inp_shape.channels_last = channels_last
inp.add_axon(inp_ax)
model.add_block(inp)
# conv block
neurons = LoihiBlock(out_size, label='neurons')
assert neurons.n_neurons <= 1024
neurons.compartment.configure_lif(tau_rc=tau_rc, tau_ref=tau_ref, dt=dt)
neurons.compartment.configure_filter(tau_s, dt=dt)
neurons.compartment.bias[:] = neuron_bias
synapse = Synapse(np.prod(inp_shape.spatial_shape), label='synapse')
conv2d_transform = nengo_transforms.Convolution(
nf,
inp_shape,
strides=(sti, stj),
channels_last=channels_last,
init=filters,
kernel_size=(1, 3))
weights, indices, axon_to_weight_map, bases = conv.conv2d_loihi_weights(
conv2d_transform)
synapse.set_population_weights(weights,
indices,
axon_to_weight_map,
bases,
pop_type=pop_type)
neurons.add_synapse(synapse)
out_probe = Probe(target=neurons, key='spiked')
neurons.add_probe(out_probe)
inp_ax.target = synapse
model.add_block(neurons)
# simulation
discretize_model(model)
n_steps = int(pres_time / dt)
target = request.config.getoption("--target")
if target == 'loihi':
with HardwareInterface(model, use_snips=False, seed=seed) as sim:
sim.run_steps(n_steps)
sim_out = sim.get_probe_output(out_probe)
else:
with EmulatorInterface(model, seed=seed) as sim:
sim.run_steps(n_steps)
sim_out = sim.get_probe_output(out_probe)
sim_out = np.sum(sim_out, axis=0) * (dt / pres_time)
if channels_last:
sim_out.shape = (nyi, nyj, nf)
sim_out = np.transpose(sim_out, (2, 0, 1))
else:
sim_out.shape = (nf, nyi, nyj)
out_max = sim_out.max()
# --- plot results
rows = 1
cols = 2
ax = plt.subplot(rows, cols, 1)
plt.hist(sim_out.ravel(), bins=11)
ax = plt.subplot(rows, cols, 2)
tile(sim_out, vmin=0, vmax=out_max, grid=True, ax=ax)
# ref_out determined by emulator running code known to work
if nc == 1:
ref_out = np.array([[0.06, 0.02], [0.055, 0.], [0.0825, 0.0225],
[0.125, 0.04]])
elif nc == 2:
ref_out = np.array([[0.0975, 0.02], [0.0825, 0.02], [0.125, 0.055],
[0.2475, 0.0825]])
assert allclose(sim_out[:, :, 0], ref_out, rtol=0, atol=1e-7)
def test_conv_connection(channels, channels_last, Simulator, seed, rng, plt,
allclose):
# load data
with open(os.path.join(test_dir, "mnist10.pkl"), "rb") as f:
test10 = pickle.load(f)
test_x = test10[0][0].reshape((28, 28))
test_x = 1.999 * test_x - 0.999 # range (-1, 1)
input_shape = make_channel_shape(test_x.shape, channels, channels_last)
filters = Gabor(freq=Uniform(0.5, 1)).generate(8, (7, 7), rng=rng)
filters = filters[None, :, :, :] # single channel
filters = np.transpose(filters, (2, 3, 0, 1))
strides = (2, 2)
tau_rc = 0.02
tau_ref = 0.002
tau_s = 0.005
neuron_type = LoihiLIF(tau_rc=tau_rc, tau_ref=tau_ref)
pres_time = 0.1
with nengo.Network(seed=seed) as model:
nengo_loihi.add_params(model)
u = nengo.Node(test_x.ravel(), label="u")
a = nengo.Ensemble(
input_shape.size,
1,
neuron_type=LoihiSpikingRectifiedLinear(),
max_rates=nengo.dists.Choice([40 / channels]),
intercepts=nengo.dists.Choice([0]),
label="a",
)
model.config[a].on_chip = False
if channels == 1:
nengo.Connection(u, a.neurons, transform=1, synapse=None)
elif channels == 2:
# encode image into spikes using two channels (on/off)
if input_shape.channels_last:
nengo.Connection(u, a.neurons[0::2], transform=1, synapse=None)
nengo.Connection(u,
a.neurons[1::2],
transform=-1,
synapse=None)
else:
k = input_shape.spatial_shape[0] * input_shape.spatial_shape[1]
nengo.Connection(u, a.neurons[:k], transform=1, synapse=None)
nengo.Connection(u, a.neurons[k:], transform=-1, synapse=None)
filters = np.concatenate([filters, -filters], axis=2)
else:
raise ValueError("Test not configured for more than two channels")
conv2d_transform = nengo_transforms.Convolution(
8,
input_shape,
strides=strides,
kernel_size=(7, 7),
channels_last=channels_last,
init=filters,
)
output_shape = conv2d_transform.output_shape
gain, bias = neuron_type.gain_bias(max_rates=100, intercepts=0)
gain = gain * 0.01 # account for `a` max_rates
b = nengo.Ensemble(
output_shape.size,
1,
neuron_type=neuron_type,
gain=nengo.dists.Choice([gain[0]]),
bias=nengo.dists.Choice([bias[0]]),
label="b",
)
nengo.Connection(a.neurons,
b.neurons,
synapse=tau_s,
transform=conv2d_transform)
bp = nengo.Probe(b.neurons)
with nengo.Simulator(model, optimize=False) as sim_nengo:
sim_nengo.run(pres_time)
ref_out = sim_nengo.data[bp].mean(axis=0).reshape(output_shape.shape)
with Simulator(model, target="simreal") as sim_emu:
sim_emu.run(pres_time)
emu_out = sim_emu.data[bp].mean(axis=0).reshape(output_shape.shape)
with Simulator(model, hardware_options={"snip_max_spikes_per_step":
800}) as sim_loihi:
sim_loihi.run(pres_time)
sim_out = sim_loihi.data[bp].mean(axis=0).reshape(output_shape.shape)
if not output_shape.channels_last:
ref_out = np.transpose(ref_out, (1, 2, 0))
emu_out = np.transpose(emu_out, (1, 2, 0))
sim_out = np.transpose(sim_out, (1, 2, 0))
out_max = max(ref_out.max(), emu_out.max(), sim_out.max())
# --- plot results
rows = 2
cols = 3
ax = plt.subplot(rows, cols, 1)
imshow(test_x, vmin=0, vmax=1, ax=ax)
ax = plt.subplot(rows, cols, 2)
tile(np.transpose(filters[0], (2, 0, 1)), cols=8, ax=ax)
ax = plt.subplot(rows, cols, 3)
plt.hist(ref_out.ravel(), bins=31)
plt.hist(sim_out.ravel(), bins=31)
ax = plt.subplot(rows, cols, 4)
tile(np.transpose(ref_out, (2, 0, 1)), vmin=0, vmax=out_max, cols=8, ax=ax)
ax = plt.subplot(rows, cols, 5)
tile(np.transpose(emu_out, (2, 0, 1)), vmin=0, vmax=out_max, cols=8, ax=ax)
ax = plt.subplot(rows, cols, 6)
tile(np.transpose(sim_out, (2, 0, 1)), vmin=0, vmax=out_max, cols=8, ax=ax)
assert allclose(emu_out, ref_out, atol=10, rtol=1e-3)
assert allclose(sim_out, ref_out, atol=10, rtol=1e-3)
for learner in learners:
plot_batches(learner.batch_costs, label=type(learner).__name__)
plt.ylabel('train cost')
plt.legend(loc='best')
plt.subplot(212)
def plot_epochs(x, label=None):
epoch_inds = len(trainX) * np.arange(1, len(x) + 1)
plt.plot(epoch_inds, x, label=label)
for learner in learners:
plot_epochs(learner.test_errors, label=type(learner).__name__)
plt.xlabel('example')
plt.ylabel('test error')
plt.legend(loc='best')
# plot encoders
plt.figure(2)
rows, cols = len(learners), 1
for i, learner in enumerate(learners):
encoders = learner.network.weights[0].T.reshape(-1, 28, 28)
plt.subplot(rows, cols, i + 1)
nplt.tile(encoders, grid=True)
plt.show()
def test_conv_deepnet(
channels_last,
pop_type,
precompute,
Simulator,
request,
rng,
seed,
plt,
allclose,
):
"""Run a convolutional network with two layers on the chip.
Checks that network with block splitting on the target matches one without
on the emulator.
"""
# TODO: This case fails in NxSDK 0.9.0 but will be fixed in the next version.
# Remove this check once the next version is released.
if pop_type == 32:
pytest.skip("Pop32 multichip test requires latest NxSDK")
def set_partition(partition):
os.environ["PARTITION"] = partition
request.addfinalizer(lambda: set_partition(""))
# multichip pop_type = 16 works only on nahuku32 board currently
if pop_type == 16:
set_partition("nahuku32")
def conv_layer(x,
input_shape,
array_init=None,
label=None,
conn_args=None,
**conv_args):
conn_args = {} if conn_args is None else conn_args
if array_init is not None:
assert all(a not in conv_args
for a in ("init", "kernel_size", "n_filters"))
assert array_init.ndim == 4
conv_args["init"] = array_init
conv_args["kernel_size"] = array_init.shape[:2]
assert array_init.shape[2] == input_shape.n_channels
conv_args["n_filters"] = array_init.shape[3]
conv = nengo.Convolution(input_shape=input_shape, **conv_args)
# add an ensemble to implement the activation function
layer = nengo.Ensemble(conv.output_shape.size, 1, label=label)
# connect up the input object to the new layer
conn = nengo.Connection(x, layer.neurons, transform=conv)
return layer, conv, conn
channels = 1
n_filters0 = 1
n_filters1 = 4
n_filters2 = 4
# load data
with open(os.path.join(test_dir, "mnist10.pkl"), "rb") as f:
test10 = pickle.load(f)
test_x = test10[0][0].reshape(28, 28) # range (0, 1)
input_shape = make_channel_shape(test_x.shape, channels, channels_last)
filters0 = np.ones((1, 1, channels, n_filters0))
# use Gabor filters for first layer
filters1 = Gabor(freq=Uniform(0.5, 1),
sigma_x=Choice([0.9]),
sigma_y=Choice([0.9])).generate(n_filters1, (7, 7),
rng=rng)
assert n_filters0 == 1
filters1 = filters1[None, :, :, :] # single channel
filters1 = np.transpose(filters1,
(2, 3, 0, 1)) # rows, cols, in_chan, out_chan
# use random combinations of first-layer channels in 1x1 convolution
filters2 = rng.uniform(-0.2, 1,
size=(n_filters1, n_filters2)).clip(0, None)
filters2 *= 2 / filters2.sum(axis=0,
keepdims=True) # each filter sums to 2
filters2 = filters2[None, None, :, :] # rows, cols, in_chan, out_chan
tau_s = 0.001
max_rate = 100
amp = 1 / max_rate
f_split = 2 if pop_type == 32 else 4
# use Loihi neuron type so Nengo sim mimics Loihi neuron effects
neuron_type = LoihiSpikingRectifiedLinear(amplitude=amp)
pres_time = 0.2
with nengo.Network(seed=seed) as net:
nengo_loihi.add_params(net)
net.config[nengo.Ensemble].neuron_type = neuron_type
net.config[nengo.Ensemble].max_rates = Choice([max_rate])
net.config[nengo.Ensemble].intercepts = Choice([0])
net.config[nengo.Connection].synapse = tau_s
u = nengo.Node(test_x.ravel(), label="u")
layer0, conv0, conn0 = conv_layer(
u,
input_shape=input_shape,
array_init=filters0,
strides=(1, 1),
channels_last=channels_last,
label="layer0",
conn_args=dict(synapse=None),
)
net.config[layer0].on_chip = False
layer1, conv1, conn1 = conv_layer(
layer0.neurons,
input_shape=conv0.output_shape,
array_init=filters1,
strides=(2, 2),
channels_last=channels_last,
label="layer1",
)
net.config[layer1].block_shape = nengo_loihi.BlockShape(
make_shape((4, 4), f_split, channels_last), conv1)
net.config[conn1].pop_type = pop_type
layer2, conv2, conn2 = conv_layer(
layer1.neurons,
input_shape=conv1.output_shape,
array_init=filters2,
strides=(1, 1),
channels_last=channels_last,
label="layer2",
)
net.config[layer2].block_shape = nengo_loihi.BlockShape(
make_shape((4, 4), f_split, channels_last), conv2)
net.config[conn2].pop_type = pop_type
output_p = nengo.Probe(layer2.neurons)
output_shape = conv2.output_shape
with nengo.Simulator(net, optimize=False) as sim_nengo:
sim_nengo.run(pres_time)
ref_out = (sim_nengo.data[output_p] > 0).sum(axis=0).reshape(
output_shape.shape)
with Simulator(net, target="sim") as sim_emu:
sim_emu.run(pres_time)
emu_out = (sim_emu.data[output_p] > 0).sum(axis=0).reshape(
output_shape.shape)
# TODO: Remove the if condition when configurable timeout parameter
# is available in nxsdk
if (pop_type == 32 or
os.popen("sinfo -h --partition=nahuku32").read().find("idle") > 0):
with Simulator(
net,
precompute=precompute,
hardware_options={
"allocator": RoundRobin(),
"snip_max_spikes_per_step": 800,
},
) as sim_loihi:
sim_loihi.run(pres_time)
sim_out = ((sim_loihi.data[output_p] > 0).sum(axis=0).reshape(
output_shape.shape))
elif nengo_loihi.version.dev is None:
pytest.fail(
"Pop16 multichip test failed since Nahuku32 is unavailable")
else:
pytest.skip(
"Pop16 multichip test skipped since Nahuku32 is unavailable")
out_max = ref_out.max()
ref_out = ref_out / out_max
emu_out = emu_out / out_max
sim_out = sim_out / out_max
if channels_last:
# channels first, to display channels in separate plots
ref_out = np.transpose(ref_out, (2, 0, 1))
emu_out = np.transpose(emu_out, (2, 0, 1))
sim_out = np.transpose(sim_out, (2, 0, 1))
# --- plot results
rows = 2
cols = 3
ax = plt.subplot(rows, cols, 1)
imshow(test_x, vmin=0, vmax=1, ax=ax)
ax = plt.subplot(rows, cols, 2)
tile(np.transpose(filters1, (2, 3, 0, 1))[0],
rows=2,
cols=2,
grid=True,
ax=ax)
ax = plt.subplot(rows, cols, 3)
plt.hist((ref_out.ravel(), emu_out.ravel(), sim_out.ravel()), bins=21)
ax = plt.subplot(rows, cols, 4)
tile(ref_out, rows=2, cols=2, grid=True, ax=ax)
ax = plt.subplot(rows, cols, 5)
tile(emu_out, rows=2, cols=2, grid=True, ax=ax)
ax = plt.subplot(rows, cols, 6)
tile(sim_out, rows=2, cols=2, grid=True, ax=ax)
assert allclose(sim_out, ref_out, atol=0.15, rtol=1e-3)
assert allclose(sim_out, emu_out, atol=1e-3, rtol=1e-3)
import matplotlib.pyplot as plt
import numpy as np
from nengo.dists import Uniform
from nengo_extras.matplotlib import tile
from nengo_extras.vision import Gabor
from hunse_thesis.dists import LogUniform
rng = np.random.RandomState(3)
# r, c = 10, 20
r, c = 9, 12
# gabor = Gabor()
# gabor = Gabor(freq=Uniform(0.5, 1.5))
gabor = Gabor(freq=LogUniform(np.log(0.5), np.log(1.5)))
gabors = gabor.generate(r * c, (32, 32), rng=rng)
tile(gabors, rows=r, cols=c)
plt.savefig('gabors.pdf')
plt.show()
def test_load_cifar100(plt):
(trainX, _), (_, _) = load_cifar100()
trainX = trainX.reshape(-1, 3, 32, 32)
trainX = np.transpose(trainX, (0, 2, 3, 1))
tile(trainX)
