def test_chain():
# test LP, AD3, AD3-BB and JT on a chain.
# they should all be exact
rnd = np.random.RandomState(0)
for i in xrange(10):
forward = np.c_[np.arange(9), np.arange(1, 10)]
backward = np.c_[np.arange(1, 10), np.arange(9)]
unary_potentials = rnd.normal(size=(10, 3))
pairwise_potentials = rnd.normal(size=(3, 3))
# test that reversing edges is same as transposing pairwise potentials
y_forward = inference_lp(unary_potentials, pairwise_potentials,
forward)
y_backward = inference_lp(unary_potentials, pairwise_potentials.T,
backward)
assert_array_equal(y_forward, y_backward)
for chain in [forward, backward]:
y_lp = inference_lp(unary_potentials, pairwise_potentials, chain)
try:
from pystruct.inference import inference_dai
y_dai = inference_dai(unary_potentials, pairwise_potentials,
chain, alg='jt')
assert_array_equal(y_dai, y_lp)
except:
pass
try:
from pystruct.inference import inference_ad3
y_ad3 = inference_ad3(unary_potentials, pairwise_potentials,
chain)
y_ad3bb = inference_ad3(unary_potentials, pairwise_potentials,
chain, branch_and_bound=True)
assert_array_equal(y_ad3, y_lp)
assert_array_equal(y_ad3bb, y_lp)
except:
pass
def test_pystruct():
unaries = np.zeros((3, 5))
unaries[1, 2] = 2
pairwise = np.eye(5)
edges = np.array([[0, 1], [1, 2], [0, 2]], dtype=np.intp)
# no parameters
labels = inference_ad3(unaries, pairwise, edges)
assert_array_equal(labels, [2, 2, 2])
# exact decoding
labels_exact = inference_ad3(unaries, pairwise, edges,
branch_and_bound=True)
assert_array_equal(labels_exact, [2, 2, 2])
# request energy
labels, energy = inference_ad3(unaries, pairwise, edges,
return_energy=True)
assert_array_equal(energy, -5)
# exact decoding and request energy
labels, energy = inference_ad3(unaries, pairwise, edges,
branch_and_bound=True, return_energy=True)
assert_array_equal(energy, -5)
def test_pystruct():
unaries = np.zeros((3, 5))
unaries[1, 2] = 2
pairwise = np.eye(5)
edges = np.array([[0, 1], [1, 2], [0, 2]], dtype=np.intp)
# no parameters
labels = inference_ad3(unaries, pairwise, edges)
assert_array_equal(labels, [2, 2, 2])
# exact decoding
labels_exact = inference_ad3(unaries,
pairwise,
edges,
branch_and_bound=True)
assert_array_equal(labels_exact, [2, 2, 2])
# request energy
labels, energy = inference_ad3(unaries,
pairwise,
edges,
return_energy=True)
assert_array_equal(energy, -5)
# exact decoding and request energy
labels, energy = inference_ad3(unaries,
pairwise,
edges,
branch_and_bound=True,
return_energy=True)
assert_array_equal(energy, -5)
def test_iterative_max_product_chain():
rnd = np.random.RandomState(0)
chain = np.c_[np.arange(9), np.arange(1, 10)]
for i in xrange(10):
unary_potentials = rnd.normal(size=(10, 3))
pairwise_potentials = rnd.normal(size=(9, 3, 3))
result_ad3 = inference_ad3(unary_potentials, pairwise_potentials, chain, branch_and_bound=True)
result_mp = iterative_max_product(unary_potentials, pairwise_potentials, chain)
assert_array_equal(result_ad3, result_mp)
def test_tree_max_product_chain():
rnd = np.random.RandomState(0)
forward = np.c_[np.arange(9), np.arange(1, 10)]
backward = np.c_[np.arange(1, 10), np.arange(9)]
for i in range(10):
unary_potentials = rnd.normal(size=(10, 3))
pairwise_potentials = rnd.normal(size=(9, 3, 3))
for chain in [forward, backward]:
result_ad3 = inference_ad3(unary_potentials, pairwise_potentials, chain, branch_and_bound=True)
result_mp = inference_max_product(unary_potentials, pairwise_potentials, chain)
assert_array_equal(result_ad3, result_mp)
def test_iterative_max_product_chain():
rnd = np.random.RandomState(0)
chain = np.c_[np.arange(9), np.arange(1, 10)]
for i in xrange(10):
unary_potentials = rnd.normal(size=(10, 3))
pairwise_potentials = rnd.normal(size=(9, 3, 3))
result_ad3 = inference_ad3(unary_potentials,
pairwise_potentials,
chain,
branch_and_bound=True)
result_mp = iterative_max_product(unary_potentials,
pairwise_potentials, chain)
assert_array_equal(result_ad3, result_mp)
def test_tree_max_product_chain():
rnd = np.random.RandomState(0)
forward = np.c_[np.arange(9), np.arange(1, 10)]
backward = np.c_[np.arange(1, 10), np.arange(9)]
for i in range(10):
unary_potentials = rnd.normal(size=(10, 3))
pairwise_potentials = rnd.normal(size=(9, 3, 3))
for chain in [forward, backward]:
result_ad3 = inference_ad3(unary_potentials,
pairwise_potentials,
chain,
branch_and_bound=True)
result_mp = inference_max_product(unary_potentials,
pairwise_potentials, chain)
assert_array_equal(result_ad3, result_mp)
def predict(self, X,y=None,w=1):
p1 = self.p1
for i in p1:p1[i].value = T.shared(p1[i].value)
edges = self.edges
pdf = self._pdf(p1, X)
theta = p1['theta'].value
X = node_potn(pdf).eval()*self.w_nodes
F = edge_potn(pdf, self._copula, theta, edges,shared_copula=self.shared_copula).dimshuffle(1,0,2,3).eval()*(1-self.w_nodes)
E = edges.eval()
y_hat = np.array([inference_ad3(-x,-f,E.T) for x,f in zip(X,F)])
return y_hat
def test_iterative_max_product_tree():
try:
from scipy.sparse.csgraph import minimum_spanning_tree
except:
raise SkipTest("Not testing trees, scipy version >= 0.11 required")
rnd = np.random.RandomState(0)
for i in xrange(100):
# generate random tree using mst
graph = rnd.uniform(size=(10, 10))
tree = minimum_spanning_tree(sparse.csr_matrix(graph))
tree_edges = np.c_[tree.nonzero()]
unary_potentials = rnd.normal(size=(10, 3))
pairwise_potentials = rnd.normal(size=(9, 3, 3))
result_ad3 = inference_ad3(unary_potentials, pairwise_potentials, tree_edges, branch_and_bound=True)
result_mp = iterative_max_product(unary_potentials, pairwise_potentials, tree_edges)
assert_array_equal(result_ad3, result_mp)
def test_iterative_max_product_tree():
try:
from scipy.sparse.csgraph import minimum_spanning_tree
except:
raise SkipTest("Not testing trees, scipy version >= 0.11 required")
rnd = np.random.RandomState(0)
for i in xrange(100):
# generate random tree using mst
graph = rnd.uniform(size=(10, 10))
tree = minimum_spanning_tree(sparse.csr_matrix(graph))
tree_edges = np.c_[tree.nonzero()]
unary_potentials = rnd.normal(size=(10, 3))
pairwise_potentials = rnd.normal(size=(9, 3, 3))
result_ad3 = inference_ad3(unary_potentials,
pairwise_potentials,
tree_edges,
branch_and_bound=True)
result_mp = iterative_max_product(unary_potentials,
pairwise_potentials, tree_edges)
assert_array_equal(result_ad3, result_mp)
