def test_select_point_to_explore_in_two_dimensions(self):
'''
This test case just runs code on a multi-dimensional input dataset.
It doesn't test any conditions, but could presumably do so in the future.
'''
acquire(
x=a([
[-0.5, -0.5],
[-0.5, 0.5],
]),
# Fake fmap and Cmap values from another set of x's
fmap=a([
[0.03254087],
[-0.03254087],
]),
Cmap=a([
[0.07894662, -0.07894662],
[-0.07894662, 0.07894662],
]),
bounds=a([
[-1.0, 1.0],
[-1.0, 1.0],
]),
kernelfunc=default_kernel
)
def test_select_point_to_exploit(self):
# We attempt to force exploitation by covering most of the input
# space and expecting that the maximization algorithm will choose
# the point between the highest outputs, given a symmetric output function.
next_point = acquire(
x=a([
[-0.75],
[-0.25],
[0.25],
[0.75],
]),
# I got these fmap and Cmap values from running our optimizer
# on the input data with comparisons [1, 0], [1, 3], [2, 0], [2, 3].
fmap=a([
[0.08950024],
[0.21423927],
[0.21423927],
[0.08950024],
]),
Cmap=a([
[0.15672336, -0.07836168, -0.07836168, 0.0],
[-0.07836168, 0.15672336, 0.0, -0.07836168],
[-0.07836168, 0.0, 0.15672336, -0.07836168],
[0.0, -0.07836168, -0.07836168, 0.15672336],
]),
bounds=a([
[-1.0, 1.0],
]),
kernelfunc=default_kernel
)
self.assertTrue(next_point[0] > -.25)
self.assertTrue(next_point[0] < .25)
def test_select_point_to_exploit(self):
# We attempt to force exploitation by covering most of the input
# space and expecting that the maximization algorithm will choose
# the point between the highest outputs, given a symmetric output function.
next_point = acquire(
x=a([
[-0.75],
[-0.25],
[0.25],
[0.75],
]),
# I got these fmap and Cmap values from running our optimizer
# on the input data with comparisons [1, 0], [1, 3], [2, 0], [2, 3].
fmap=a([
[0.08950024],
[0.21423927],
[0.21423927],
[0.08950024],
]),
Cmap=a([
[0.15672336, -0.07836168, -0.07836168, 0.0],
[-0.07836168, 0.15672336, 0.0, -0.07836168],
[-0.07836168, 0.0, 0.15672336, -0.07836168],
[0.0, -0.07836168, -0.07836168, 0.15672336],
]),
bounds=a([
[-1.0, 1.0],
]),
kernelfunc=default_kernel)
self.assertTrue(next_point[0] > -.25)
self.assertTrue(next_point[0] < .25)
def test_select_point_to_explore_in_two_dimensions(self):
'''
This test case just runs code on a multi-dimensional input dataset.
It doesn't test any conditions, but could presumably do so in the future.
'''
acquire(
x=a([
[-0.5, -0.5],
[-0.5, 0.5],
]),
# Fake fmap and Cmap values from another set of x's
fmap=a([
[0.03254087],
[-0.03254087],
]),
Cmap=a([
[0.07894662, -0.07894662],
[-0.07894662, 0.07894662],
]),
bounds=a([
[-1.0, 1.0],
[-1.0, 1.0],
]),
kernelfunc=default_kernel)
def test_select_point_to_explore(self):
next_point = acquire(
x=a([
[-0.75],
[-0.4],
]),
# I got these fmap and Cmap values from running our optimizer
# on the input data with comparisons [0, 1]
fmap=a([
[0.03254087],
[-0.03254087],
]),
Cmap=a([
[0.07894662, -0.07894662],
[-0.07894662, 0.07894662],
]),
bounds=a([
[-1.0, 1.0],
]),
kernelfunc=default_kernel)
self.assertTrue(next_point[0] > -.4)
def test_select_point_to_explore(self):
next_point = acquire(
x=a([
[-0.75],
[-0.4],
]),
# I got these fmap and Cmap values from running our optimizer
# on the input data with comparisons [0, 1]
fmap=a([
[0.03254087],
[-0.03254087],
]),
Cmap=a([
[0.07894662, -0.07894662],
[-0.07894662, 0.07894662],
]),
bounds=a([
[-1.0, 1.0],
]),
kernelfunc=default_kernel
)
self.assertTrue(next_point[0] > -.4)
def test_select_point_to_explore_within_bounds_function(self):
next_point = acquire(
x=a([
[-0.75],
[-0.4],
]),
# See above for the source of fmap and Cmap values
fmap=a([
[0.03254087],
[-0.03254087],
]),
Cmap=a([
[0.07894662, -0.07894662],
[-0.07894662, 0.07894662],
]),
bounds=a([
[-2.0, 2.0],
]),
kernelfunc=default_kernel,
extrabounds=lambda p: p[0] > -1.0 and p[0] < 1.0,
)
self.assertTrue(next_point[0] > -.4)
self.assertTrue(next_point[0] < 1.0)
def test_select_point_to_explore_within_bounds_function(self):
next_point = acquire(
x=a([
[-0.75],
[-0.4],
]),
# See above for the source of fmap and Cmap values
fmap=a([
[0.03254087],
[-0.03254087],
]),
Cmap=a([
[0.07894662, -0.07894662],
[-0.07894662, 0.07894662],
]),
bounds=a([
[-2.0, 2.0],
]),
kernelfunc=default_kernel,
extrabounds=lambda p: p[0] > -1.0 and p[0] < 1.0,
)
self.assertTrue(next_point[0] > -.4)
self.assertTrue(next_point[0] < 1.0)
def bayesopt():
# The caller is responsible with updating comparisons.
# They should be able to just upload the x and f as they were provided by the last
# call to this URL.
f = np.array(loads(request.args.get('f', '[]')), dtype=np.float)
x = np.array(loads(request.args.get('x', '[]')), dtype=np.float)
comparisons = np.array(loads(request.args.get('comparisons', '[]')), dtype=np.int)
SIGMA = 10.0
BOUNDS = np.array([
[0.0, 4.0],
[0.0, 4.0],
[0.0, 4.0],
])
KERNELFUNC = partial(default_kernel, a=-.25)
# Create extra bounds to rule out configurations that burned
within_bounds = {}
BURNING_CONFIGURATIONS = [
(4, 0, 0),
(4, 1, 0),
(4, 0, 1),
(4, 1, 1),
(4, 1, 1),
(3, 0, 2),
(4, 0, 2),
(4, 1, 2),
(3, 0, 3),
(4, 0, 3),
(4, 1, 3),
(3, 0, 4),
(4, 0, 4),
(4, 1, 4),
]
keys = [_ for _ in itertools.product(range(5), range(5), range(5))]
for k in keys:
within_bounds[k] = (k not in BURNING_CONFIGURATIONS)
def extrabounds(p):
rounded = np.round(p).astype('int')
rounded_key = tuple(rounded)
return within_bounds[rounded_key]
# If no parameters are given, initialize the data
if f.shape == (0,):
x = np.array([
[1.0, 1.0, 1.0],
])
f = np.array([
[0.0],
])
comparisons = np.array([])
best_f_i = 0
xbest = x[0]
xnew = np.array([3.0, 3.0, 3.0])
# Predict the next point for comparison
else:
prof = cProfile.Profile()
f, Cmap = prof.runcall(
newton_rhapson, x, f, comparisons, KERNELFUNC,
compute_H, compute_g, SIGMA, maxiter=10)
prof.dump_stats('code.profile')
xnew = acquire(x, f, Cmap, BOUNDS, KERNELFUNC, extrabounds)
best_f_i = np.argmax(f)
xbest = x[best_f_i]
# Add newest points to x and f
x = np.array(x.tolist() + [xnew])
f = np.array(f.tolist() + [[0.0]])
return jsonify(**{
'x': x.tolist(),
'f': f.tolist(),
'xbest': {
'value': xbest.tolist(),
'img': get_cut_image_name(*(xbest.tolist())),
'index': best_f_i,
},
'xnew': {
'value': xnew.tolist(),
'img': get_cut_image_name(*(xnew.tolist())),
'index': len(f) - 1,
},
'comparisons': comparisons.tolist()
})
def bayesopt():
# The caller is responsible with updating comparisons.
# They should be able to just upload the x and f as they were provided by the last
# call to this URL.
f = np.array(loads(request.args.get('f', '[]')), dtype=np.float)
x = np.array(loads(request.args.get('x', '[]')), dtype=np.float)
comparisons = np.array(loads(request.args.get('comparisons', '[]')),
dtype=np.int)
SIGMA = 10.0
BOUNDS = np.array([
[0.0, 4.0],
[0.0, 4.0],
[0.0, 4.0],
])
KERNELFUNC = partial(default_kernel, a=-.25)
# Create extra bounds to rule out configurations that burned
within_bounds = {}
BURNING_CONFIGURATIONS = [
(4, 0, 0),
(4, 1, 0),
(4, 0, 1),
(4, 1, 1),
(4, 1, 1),
(3, 0, 2),
(4, 0, 2),
(4, 1, 2),
(3, 0, 3),
(4, 0, 3),
(4, 1, 3),
(3, 0, 4),
(4, 0, 4),
(4, 1, 4),
]
keys = [_ for _ in itertools.product(range(5), range(5), range(5))]
for k in keys:
within_bounds[k] = (k not in BURNING_CONFIGURATIONS)
def extrabounds(p):
rounded = np.round(p).astype('int')
rounded_key = tuple(rounded)
return within_bounds[rounded_key]
# If no parameters are given, initialize the data
if f.shape == (0, ):
x = np.array([
[1.0, 1.0, 1.0],
])
f = np.array([
[0.0],
])
comparisons = np.array([])
best_f_i = 0
xbest = x[0]
xnew = np.array([3.0, 3.0, 3.0])
# Predict the next point for comparison
else:
prof = cProfile.Profile()
f, Cmap = prof.runcall(newton_rhapson,
x,
f,
comparisons,
KERNELFUNC,
compute_H,
compute_g,
SIGMA,
maxiter=10)
prof.dump_stats('code.profile')
xnew = acquire(x, f, Cmap, BOUNDS, KERNELFUNC, extrabounds)
best_f_i = np.argmax(f)
xbest = x[best_f_i]
# Add newest points to x and f
x = np.array(x.tolist() + [xnew])
f = np.array(f.tolist() + [[0.0]])
return jsonify(
**{
'x': x.tolist(),
'f': f.tolist(),
'xbest': {
'value': xbest.tolist(),
'img': get_cut_image_name(*(xbest.tolist())),
'index': best_f_i,
},
'xnew': {
'value': xnew.tolist(),
'img': get_cut_image_name(*(xnew.tolist())),
'index': len(f) - 1,
},
'comparisons': comparisons.tolist()
})
