def test_max_marginals():
"""
Test that max-marginals are correct.
"""
for h in hypergraphs():
w = utils.random_viterbi_potentials(h)
print w.show(h)
path = ph.best_path(h, w)
best = w.dot(path)
print "BEST"
print "\n".join(["%20s : %s"%(edge.label, w[edge]) for edge in path.edges])
print best
nt.assert_not_equal(best, 0.0)
max_marginals = ph.compute_marginals(h, w)
for node in h.nodes:
other = max_marginals[node]
nt.assert_less_equal(other, best + 1e-4)
for edge in h.edges:
other = max_marginals[edge]
nt.assert_less_equal(other, best + 1e-4)
if edge in path:
nt.assert_almost_equal(other, best)
def get_marginals(self):
if not self.potentials:
self.get_potentials()
marginal_values = \
ph.compute_marginals(self.hypergraph, self.potentials)
marginals = {}
if not self.total_potentials:
self.sum_potentials()
root_value = self.total_potentials
assert root_value > 0, "sentence is " + " ".join(self.words)
for edge in self.hypergraph.edges:
marginals[edge.id] = marginal_values[edge] / root_value
return marginals
def em(distribution_table, label_map, hypergraph, base=None, epochs=10):
base_potentials = base
if base is None:
base_potentials = np.zeros([len(hypergraph)])
ll = []
for i in range(epochs):
print "epoch:", i
potentials = ph.LogProbPotentials(hypergraph).from_array(
base_potentials + np.log(distribution_table.to_array(hypergraph, label_map))
)
print "start"
margs = ph.compute_marginals(hypergraph, potentials)
print "stop"
for node in hypergraph.nodes:
distribution_table.inc(label_map(node.label), math.exp(margs[node] - margs[hypergraph.root]))
distribution_table.reestimate()
print margs[hypergraph.root]
ll.append(margs[hypergraph.root])
return ll
def test_semirings():
for hypergraph in hypergraphs():
potentials = ph.ViterbiPotentials(hypergraph).build(lambda l: 10.0)
marg = ph.Viterbi.compute_marginals(hypergraph, potentials)
log_potentials = ph.LogViterbiPotentials(hypergraph).build(lambda l: 10.0)
potentials = ph.LogViterbiPotentials(hypergraph).build(lambda l: 10.0)
chart = ph.inside(hypergraph, log_potentials)
chart2 = ph.inside_values(hypergraph, potentials)
for node in hypergraph.nodes:
nt.assert_equal(chart[node], chart2[node])
marg = ph.LogViterbi.compute_marginals(hypergraph, log_potentials)
marg2 = ph.compute_marginals(hypergraph, potentials)
for edge in hypergraph.edges:
nt.assert_almost_equal(marg[edge], marg2[edge])
potentials = ph.Inside.Potentials(hypergraph).build(lambda l: 0.5)
chart = ph.Inside.inside(hypergraph, potentials)
potentials = ph.Inside.Potentials(hypergraph).build(lambda l: 0.5)
def test_pruning():
for h in hypergraphs():
w = utils.random_viterbi_potentials(h)
original_path = ph.best_path(h, w)
new_hyper, new_potentials = ph.prune_hypergraph(h, w, -0.99)
prune_path = ph.best_path(new_hyper, new_potentials)
assert len(original_path.edges) > 0
for edge in original_path.edges:
assert edge in prune_path
valid_path(new_hyper, prune_path)
original_score = w.dot(original_path)
print original_score
print new_potentials.dot(prune_path)
nt.assert_almost_equal(original_score,
new_potentials.dot(prune_path))
# Test pruning amount.
prune = 0.001
max_marginals = ph.compute_marginals(h, w)
new_hyper, new_potentials = ph.prune_hypergraph(h, w, prune)
assert (len(new_hyper.edges) > 0)
original_edges = {}
for edge in h.edges:
original_edges[edge.label] = edge
new_edges = {}
for edge in new_hyper.edges:
new_edges[edge.label] = edge
for name, edge in new_edges.iteritems():
orig = original_edges[name]
nt.assert_almost_equal(w[orig], new_potentials[edge])
m = max_marginals[orig]
nt.assert_greater(m, prune)
def test_posteriors():
"Check the posteriors by enumeration."
for h in hypergraphs():
w = utils.random_inside_potentials(h)
marg = ph.compute_marginals(h, w)
paths = utils.all_paths(h)
m = defaultdict(lambda: 0.0)
total_score = 0.0
for path in paths:
path_score = w.dot(path)
total_score += path_score
for edge in path:
m[edge.id] += path_score
for edge in h.edges:
nt.assert_almost_equal(
marg[edge] / marg[h.root],
m[edge.id] / total_score, places=4)
chart = ph.inside(h, w)
nt.assert_almost_equal(chart[h.root], total_score, places=4)
def em(distribution_table, label_map, hypergraph, base=None,
epochs=10):
base_potentials = base
if base is None:
base_potentials = np.zeros([len(hypergraph)])
ll = []
for i in range(epochs):
print "epoch:", i
potentials = ph.LogProbPotentials(hypergraph).from_array(
base_potentials + np.log(distribution_table.to_array(hypergraph, label_map)))
print "start"
margs = ph.compute_marginals(hypergraph, potentials)
print "stop"
for node in hypergraph.nodes:
distribution_table.inc(
label_map(node.label),
math.exp(margs[node] - margs[hypergraph.root]))
distribution_table.reestimate()
print margs[hypergraph.root]
ll.append(margs[hypergraph.root])
return ll
# In[ ]:
def build_potentials(arc):
print arc
return random.random()
potentials = ph.Potentials(hypergraph).build(build_potentials)
# phyper, ppotentials = ph.prune_hypergraph(hypergraph, potentials, 0.5)
# In[ ]:
path = ph.best_path(hypergraph, potentials)
best = potentials.dot(path)
maxmarginals = ph.compute_marginals(hypergraph, potentials)
avg = 0.0
for edge in hypergraph.edges:
avg += float(maxmarginals[edge])
avg = avg / float(len(hypergraph.edges))
thres = ((0.9) * best + (0.1) * avg)
kept = set()
for edge in hypergraph.edges:
score = float(maxmarginals[edge])
if score >= thres:
kept.add(edge.id)
# In[ ]:
