def test_mvc_terminate():
prob = graphrl.MVCProblem()
g = graphgen.ToyGraphDistribution().sample_graph()
s = graphrl.Solution(g)
s = s.add_node(0)
assert not prob.terminate(s)
s = s.add_node(3)
s = s.add_node(4)
assert not prob.terminate(s)
s = s.add_node(6)
assert prob.terminate(s)
s = s.add_node(7)
assert prob.terminate(s)
assert prob.terminate(s.subsolution_at_step(4))
assert not prob.terminate(s.subsolution_at_step(3))
assert not prob.terminate(s.subsolution_at_step(1))
def train_ep(self):
graph = self.distribution.sample_graph()
sol = graphrl.Solution(graph, device=self.dev)
time = 0
loss = 0
grad = 0
while not self.problem.terminate(sol):
newvertex = sol.pick_node(self.embedder, self.epsilon)
sol = sol.add_node(newvertex)
if time >= self.updatedelay:
self.rmem.add(
struct2vec.TrainingCase(sol, self.problem,
self.updatedelay))
if len(self.rmem) >= self.batchsize:
training = self.rmem.sample(self.batchsize)
loss += self.embedder.train(training, self.problem,
self.q_clamp)
gnorm = torch.nn.utils.clip_grad_norm_(
self.embedder.parameters(), 10)
grad += gnorm
# total_grad = torch.nn.utils.clip_grad_norm_(self.embedder.parameters(), 1)
# print(total_grad)
self.optimizer.step()
time += 1
grad /= time
return loss, grad
def test_s2v_sanity():
g = graphgen.ToyGraphDistribution().sample_graph()
s = graphrl.Solution(g)
ge = struct2vec.GraphEmbedder()
qvs, embs = ge([s])
assert qvs.size() == torch.Size((1, 8))
assert embs.size() == torch.Size((1, 8, 10))
def test_s2v_multi():
ge = struct2vec.GraphEmbedder()
gen = graphgen.ErdosRenyiDistribution()
gs = []
for _ in range(5):
gs.append(graphrl.Solution(gen.sample_graph(12)))
qvs, embs = ge(gs)
assert qvs.size() == torch.Size((5, 12))
assert embs.size() == torch.Size((5, 12, 10))
def test_view_resolution():
g = graphgen.ToyGraphDistribution().sample_graph()
s = graphrl.Solution(g)
s = s.add_node(0)
s = s.add_node(2)
s = s.add_node(3)
s = s.add_node(1)
s1 = s.subsolution_at_step(1)
assert s == s1.add_node(2).add_node(3).add_node(1)
def test_solution_subsolution():
g = graphgen.ToyGraphDistribution().sample_graph()
s = graphrl.Solution(g)
s = s.add_node(0)
s_at1 = s
s = s.add_node(2)
s_at2 = s
s = s.add_node(3)
s = s.add_node(1)
s2 = graphrl.Solution(g)
s2 = s2.add_node(0)
assert s == s
assert g == g
assert s_at1 == s2
assert s_at2 != s2
assert s_at1 != s
assert s_at2 != s
assert s_at1 == s.subsolution_at_step(1)
assert s_at2 == s.subsolution_at_step(2)
assert s_at1 != s.subsolution_at_step(2)
assert s.subsolution_at_step(0) == graphrl.Solution(g)
assert s.subsolution_at_step(1) == graphrl.Solution(g).add_node(0)
assert s.subsolution_at_step(1) != graphrl.Solution(g).add_node(2)
assert s.subsolution_at_step(2) == graphrl.Solution(g).add_node(
0).add_node(2)
i = 0
for _ in s.subsolution_at_step(1):
i += 1
assert i == 1
def test_mvc_reward():
prob = graphrl.MVCProblem()
g = graphgen.ToyGraphDistribution().sample_graph()
s = graphrl.Solution(g)
s = s.add_node(0)
assert prob.cost(s) == -1
s = s.add_node(3)
s = s.add_node(2)
assert prob.cumul_reward(s, 0, 3) == -3
def test_pick_random_node():
g = graphgen.ToyGraphDistribution().sample_graph()
s = graphrl.Solution(g)
for _ in range(100):
assert s.pick_random_node() < len(g)
s = s.add_node(0)
s = s.add_node(2)
s = s.add_node(3)
s = s.add_node(4)
s = s.add_node(5)
s = s.add_node(6)
s = s.add_node(7)
assert s.pick_random_node() == 1
s = s.add_node(1)
with pytest.raises(IndexError):
s.pick_random_node()
def test_solution_pick_node():
g = graphgen.ToyGraphDistribution().sample_graph()
s = graphrl.Solution(g)
emb = MagicMock(return_value=(torch.tensor(
[[0, 3, 0, 0, 1, 0, 2, 0]], dtype=torch.get_default_dtype()), None))
assert s.pick_node(emb, 0) == 1
s = s.add_node(1)
assert s.pick_node(emb, 0) == 6
s = s.add_node(6)
assert s.pick_node(emb, 0) == 4
s = s.add_node(4)
assert s.pick_node(emb, 0) == 0
s = s.add_node(0)
assert s.pick_node(emb, 0) == 2
s = s.add_node(2)
assert s.pick_node(emb, 0) == 3
s = s.add_node(3)
s = s.add_node(5)
s = s.add_node(7)
with pytest.raises(IndexError):
s.pick_node(emb, 0)