def test_hopfield_capacity_n(n, k=1, iters=100):
corruption = np.empty((iters, k))
# store the same number of items multiple times
for i in xrange(iters):
# generate random inputs
vecs = util.random_input(n, k)
# create hopfield net
mem = hop.hopnet(vecs)
# read the items backout
r = mem.readM(vecs, 1000)
# find the largest fraction of corrupted bits
corruption[i] = np.mean(r ^ vecs, axis=0)
return corruption
def test_hopfield_capacity_n(n, k=1, iters=100):
corruption = np.empty((iters, k))
# store the same number of items multiple times
for i in xrange(iters):
# generate random inputs
vecs = util.random_input(n, k)
# create hopfield net
mem = hop.hopnet(vecs)
# read the items backout
r = mem.readM(vecs, 1000)
# find the largest fraction of corrupted bits
corruption[i] = np.mean(r ^ vecs, axis=0)
return corruption
def test_hopfield_prototype_n(n, kp=1, ke=1, noise=0, iters=100):
corruption = np.empty((iters, kp))
bits = int(n * noise)
# store the same number of items multiple times
for i in xrange(iters):
# generate random inputs
vecs = util.random_input(n, kp)
cvecs = vecs[..., None] * np.ones((n, kp, ke), dtype="i4")
cvecs = util.corrupt(cvecs.reshape((n, kp * ke)), bits, with_replacement=True)
ex = util.corrupt(vecs, bits)
# create hopfield net
mem = hop.hopnet(cvecs)
# read the items backout
r = mem.readM(ex, 1000)
# find the largest fraction of corrupted bits
corruption[i] = np.mean(r ^ vecs, axis=0)
return corruption
def test_hopfield_noise_tolerance_n(n, k=1, noise=0, iters=100):
if noise == 0:
return test_hopfield_capacity_n(n, k=k, iters=iters)
corruption = np.empty((iters, k))
bits = int(n * noise)
# store the same number of items multiple times
for i in xrange(iters):
# generate random inputs
vecs = util.random_input(n, k)
cvecs = util.corrupt(vecs, bits)
# create hopfield net
mem = hop.hopnet(vecs)
# read the items backout
r = mem.readM(cvecs, 1000)
# find the largest fraction of corrupted bits
corruption[i] = np.mean(r ^ vecs, axis=0)
return corruption
def test_hopfield_noise_tolerance_n(n, k=1, noise=0, iters=100):
if noise == 0:
return test_hopfield_capacity_n(n, k=k, iters=iters)
corruption = np.empty((iters, k))
bits = int(n * noise)
# store the same number of items multiple times
for i in xrange(iters):
# generate random inputs
vecs = util.random_input(n, k)
cvecs = util.corrupt(vecs, bits)
# create hopfield net
mem = hop.hopnet(vecs)
# read the items backout
r = mem.readM(cvecs, 1000)
# find the largest fraction of corrupted bits
corruption[i] = np.mean(r ^ vecs, axis=0)
return corruption
def test_hopfield_prototype_n(n, kp=1, ke=1, noise=0, iters=100):
corruption = np.empty((iters, kp))
bits = int(n * noise)
# store the same number of items multiple times
for i in xrange(iters):
# generate random inputs
vecs = util.random_input(n, kp)
cvecs = vecs[..., None] * np.ones((n, kp, ke), dtype='i4')
cvecs = util.corrupt(
cvecs.reshape((n, kp*ke)),
bits, with_replacement=True)
ex = util.corrupt(vecs, bits)
# create hopfield net
mem = hop.hopnet(cvecs)
# read the items backout
r = mem.readM(ex, 1000)
# find the largest fraction of corrupted bits
corruption[i] = np.mean(r ^ vecs, axis=0)
return corruption
return newarr
data = np.load('patterns.npz')
X = data['X']
hi = data['hi']
face = data['face']
num1 = data['num1']
num2 = data['num2']
num3 = data['num3']
num4 = data['num4']
data.close()
#inputs = np.hstack([face, X, hi, num1, num2, num3, num4])
inputs = np.hstack([face, X, hi, num1, num2])
hop = hopfield.hopnet(inputs)
addresses = inputs.copy()
numNeurons, numPatterns = addresses.shape
numIters = 1000
print "uncorrupted tests"
for i in xrange(numPatterns):
a = addresses[:, i]
d = hop.read(a, numIters).reshape((10, 10), order='F')
plt.figure(i)
plot_io(a.reshape((10, 10), order='F'), d)
print "now with corruption"
for i in xrange(numPatterns):
a = corrupt(addresses[:, i], 10)
newarr = arr[:, 0].reshape((n, n), order='F').ravel()[:, None]
return newarr
data = np.load('patterns.npz')
X = data['X']
hi = data['hi']
face = data['face']
num1 = data['num1']
num2 = data['num2']
num3 = data['num3']
num4 = data['num4']
data.close()
#inputs = np.hstack([face, X, hi, num1, num2, num3, num4])
inputs = np.hstack([face, X, hi, num1, num2])
hop = hopfield.hopnet(inputs)
addresses = inputs.copy()
numNeurons, numPatterns = addresses.shape
numIters = 1000
print "uncorrupted tests"
for i in xrange(numPatterns):
a = addresses[:, i]
d = hop.read(a, numIters).reshape((10, 10), order='F')
plt.figure(i)
plot_io(a.reshape((10, 10), order='F'), d)
print "now with corruption"
for i in xrange(numPatterns):
a = corrupt(addresses[:, i], 10)
three = data["three"]
four = data["four"]
five = data["five"]
six = data["six"]
inputs = np.hstack([zero, one, two, three, four, five, six])
vec = six.copy()
cvecs = util.corrupt(vec * np.ones((n, nex), dtype="i4"), b)
ex = util.corrupt(vec.copy(), b)
mem1 = sdm.SDM(n, m, D)
mem1.writeM(cvecs, cvecs)
r1 = mem1.readM(ex)
mem2 = hop.hopnet(cvecs)
r2 = mem2.readM(ex, 1000)
for i in xrange(nex):
plt.clf()
plt.imshow(cvecs[:, i].reshape((16, 16), order="F"), cmap="gray", vmin=0, vmax=1, interpolation="nearest")
plt.xticks([], [])
plt.yticks([], [])
save("vi_ex%d" % i)
plt.clf()
util.plot_io(ex.reshape((16, 16), order="F"), r1.reshape((16, 16), order="F"))
save("vi_sdm")
plt.clf()
util.plot_io(ex.reshape((16, 16), order="F"), 1 - r2.reshape((16, 16), order="F"))
three = data['three']
four = data['four']
five = data['five']
six = data['six']
inputs = np.hstack([zero, one, two, three, four, five, six])
vec = six.copy()
cvecs = util.corrupt(vec * np.ones((n, nex), dtype='i4'), b)
ex = util.corrupt(vec.copy(), b)
mem1 = sdm.SDM(n, m, D)
mem1.writeM(cvecs, cvecs)
r1 = mem1.readM(ex)
mem2 = hop.hopnet(cvecs)
r2 = mem2.readM(ex, 1000)
for i in xrange(nex):
plt.clf()
plt.imshow(cvecs[:, i].reshape((16, 16), order='F'),
cmap='gray',
vmin=0,
vmax=1,
interpolation='nearest')
plt.xticks([], [])
plt.yticks([], [])
save("vi_ex%d" % i)
plt.clf()
util.plot_io(ex.reshape((16, 16), order='F'), r1.reshape((16, 16), order='F'))
