def test_greedy(self, graph, path=None):
self.problem.g = to_cuda(graph)
res = [self.problem.calc_S().item()]
pb = self.problem
if path is not None:
path = os.path.abspath(os.path.join(os.getcwd())) + path
vis_g(pb, name=path + str(0), topo='cut')
R = []
Reward = []
for j in range(100):
M = []
actions = pb.get_legal_actions()
for k in range(actions.shape[0]):
_, r = pb.step(actions[k], state=dc(pb.g))
M.append(r)
if max(M) <= 0:
break
if path is not None:
vis_g(pb, name=path + str(j + 1), topo='cut')
posi = [x for x in M if x > 0]
nega = [x for x in M if x <= 0]
# print('posi reward ratio:', len(posi) / len(M))
# print('posi reward avg:', sum(posi) / len(posi))
# print('nega reward avg:', sum(nega) / len(nega))
max_idx = torch.tensor(M).argmax().item()
_, r = pb.step((actions[max_idx, 0].item(), actions[max_idx,
1].item()))
R.append((actions[max_idx, 0].item(), actions[max_idx,
1].item(), r.item()))
Reward.append(r.item())
res.append(res[-1] - r.item())
return QtableKey2state(
state2QtableKey(
pb.g.ndata['label'].argmax(dim=1).cpu().numpy())), R, res
def cmpt_optimal(self, graph, path=None):
self.problem.g = to_cuda(graph)
res = [self.problem.calc_S().item()]
pb = self.problem
S = []
for j in range(280):
pb.reset_label(QtableKey2state(self.all_states[j]))
S.append(pb.calc_S())
s1 = torch.tensor(S).argmin()
res.append(S[s1].item())
if path is not None:
path = os.path.abspath(os.path.join(os.getcwd())) + path
pb.reset_label(QtableKey2state(self.all_states[s1]))
vis_g(pb, name=path, topo='cut')
return QtableKey2state(self.all_states[s1]), res
def test_dqn(self, alg, g_i, t, init_label=None, path=None):
init_state = dc(self.episodes[g_i].init_state)
label_history = init_state.ndata['label'][
self.episodes[g_i].label_perm]
self.problem.g = dc(init_state)
if init_label is None:
self.problem.reset_label(label=label_history[0].argmax(dim=1))
else:
self.problem.reset_label(label=init_label)
S = self.problem.calc_S()
print('init S:', S)
if path is not None:
path = os.path.abspath(os.path.join(os.getcwd())) + path
vis_g(self.problem, name=path + str(0), topo='cut')
for i in range(t):
actions = self.problem.get_legal_actions(
state=self.problem.g, action_type=self.action_type)
S_a_encoding, h1, h2, Q_sa = alg.forward(
to_cuda(self.problem.g),
actions.cuda(),
action_type=self.action_type,
gnn_step=3)
_, r = self.problem.step(state=self.problem.g,
action=actions[torch.argmax(Q_sa)],
action_type=self.action_type)
print(Q_sa.detach().cpu().numpy())
print('action index:', torch.argmax(Q_sa).detach().cpu().item())
print('action:',
actions[torch.argmax(Q_sa)].detach().cpu().numpy())
print('reward:', r.item())
S -= r.item()
print('kcut S:', S)
if path is not None:
vis_g(self.problem, name=path + str(i + 1), topo='cut')
torch.tensor(subopt_q_mean_win).mean() torch.tensor(subopt_q_mean_sway).mean() torch.tensor(subopt_q_mean_zero).mean() torch.tensor(subopt_q_mean_fail).mean() torch.tensor(subopt_q_mean).mean() torch.tensor(opt_q_mean).mean() # visualization gi = 0 test.episodes[gi].action_seq test.episodes[gi].reward_seq test.cmpt_optimal(test.episodes[gi].init_state, '/Analysis/eval_model/opt') test.test_greedy(test.episodes[gi].init_state, '/Analysis/eval_model/grd_') test.test_dqn(alg, gi, 10, '/Analysis/eval_model/dqn_') # visualization path = os.path.abspath(os.path.join(os.getcwd())) + '/Analysis/eval_model/test1' vis_g(problem, name=path, topo='cut') g_i = 1 test.episodes[g_i].reward_seq[:10] test.episodes[g_i]. validation_problem1[g_i][1] init_state = dc(test.episodes[g_i].init_state) label_history = init_state.ndata['label'][test.episodes[g_i].label_perm] label_history[0].argmax(dim=1)