def prune_actions(self, bg, recent_states):
batch_size = bg.batch_size
n = bg.ndata['label'].shape[0] // batch_size
batch_legal_actions = self.problem.get_legal_actions(
state=bg,
action_type=self.action_type,
action_dropout=self.action_dropout).cuda()
num_actions = batch_legal_actions.shape[0] // batch_size
forbid_action_mask = torch.zeros(batch_legal_actions.shape[0],
1).cuda()
for k in range(self.new_epi_batch_size, bg.batch_size):
cur_state = state2QtableKey(
bg.ndata['label'][k * n:(k + 1) * n, :].argmax(
dim=1).cpu().numpy()) # current state
forbid_states = set(recent_states[k - self.new_epi_batch_size])
candicate_actions = batch_legal_actions[k * num_actions:(k + 1) *
num_actions, :]
for j in range(num_actions):
cur_state_l = QtableKey2state(cur_state)
cur_state_l[candicate_actions[j][0]] ^= cur_state_l[
candicate_actions[j][1]]
cur_state_l[candicate_actions[j][1]] ^= cur_state_l[
candicate_actions[j][0]]
cur_state_l[candicate_actions[j][0]] ^= cur_state_l[
candicate_actions[j][1]]
if state2QtableKey(cur_state_l) in forbid_states:
forbid_action_mask[j + k * num_actions] += 1
batch_legal_actions = batch_legal_actions * (
1 - forbid_action_mask.int()).t().flatten().unsqueeze(1)
return batch_legal_actions
def compare_opt(self, validation_problem):
bingo = 0
for i in range(self.batch_size):
self.problem.g = self.episodes[i].init_state
end_perm = self.episodes[i].label_perm[self.end_of_episode[i]]
end_label = self.episodes[i].init_state.ndata['label'][end_perm]
dqn_result = state2QtableKey(end_label.argmax(dim=1).cpu().numpy())
opt_result = state2QtableKey(validation_problem[i][1])
grd_result = state2QtableKey(validation_problem[i][2])
if dqn_result == opt_result:
bingo += 1
print('Optimal solution hit percentage:', bingo)
return bingo
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 __init__(self, alg, problem=None, q_net='mlp', forbid_revisit=False):
if isinstance(alg, DQN):
self.alg = alg.model
else:
self.alg = alg
self.problem = problem
if isinstance(problem.g, BatchedDGLGraph):
self.n = problem.g.ndata['label'].shape[0] // problem.g.batch_size
else:
self.n = problem.g.ndata['label'].shape[0]
# if isinstance(problem, BatchedDGLGraph):
# self.problem = problem
# elif isinstance(problem, list):
# self.problem = dgl.batch(problem)
# else:
# self.problem = False
self.episodes = []
self.S = []
self.max_gain = []
self.max_gain_budget = []
self.max_gain_ratio = []
# self.action_indices = DataFrame(range(27))
self.q_net = q_net
self.forbid_revisit = forbid_revisit
self.all_states = list(
set([
state2QtableKey(x) for x in list(
itertools.permutations([0, 0, 0, 1, 1, 1, 2, 2, 2], 9))
]))
self.state_eval = []
def __init__(self, g, max_episode_len, action_type='swap'):
self.action_type = action_type
self.init_state = dc(g)
self.n = g.number_of_nodes()
self.max_episode_len = max_episode_len
self.episode_len = 0
self.action_seq = []
self.action_indices = []
self.reward_seq = []
self.q_pred = []
self.action_candidates = []
self.enc_state_seq = []
self.sub_reward_seq = []
# self.calib_reward_seq = []
if self.action_type == 'swap':
self.label_perm = torch.tensor(range(self.n)).unsqueeze(0)
self.enc_state_seq.append(
state2QtableKey(self.init_state.ndata['label'].argmax(
dim=1).cpu().numpy()))
if self.action_type == 'flip':
self.label_perm = self.init_state.ndata['label'].nonzero(
)[:, 1].unsqueeze(0)
# self.visited_state = set([''.join([str(i.item()) for i in torch.tensor(range(self.n)).unsqueeze(0)[0]])])
# self.node_visit_cnt = [0] * self.n
self.best_gain_sofar = 0
self.current_gain = 0
self.loop_start_position = 0
from tqdm import tqdm
import pickle
from toy_models.Qiter import vis_g, state2QtableKey, QtableKey2state
os.environ['CUDA_VISIBLE_DEVICES'] = '1'
problem = KCut_DGL(k=3,
m=3,
adjacent_reserve=5,
hidden_dim=16,
mode='complete')
res = np.zeros((3, 20000))
greedy_move_history = []
all_states = list(
set([
state2QtableKey(x)
for x in list(itertools.permutations([0, 0, 0, 1, 1, 1, 2, 2, 2], 9))
]))
select_problem = []
for i in range(20000):
# # brute force
problem.reset()
res[0, i] = problem.calc_S().item()
pb1 = dc(problem)
pb1.g = to_cuda(pb1.g)
S = []
for j in range(280):
pb1.reset_label(QtableKey2state(all_states[j]))
label_history[0].argmax(dim=1)
a=[test.episodes[i].loop_start_position for i in range(100)]
b=[len(test.valid_states[i]) for i in range(100)]
sum(a)
sum(b)
test.state_eval[3][1,:]
self=test
bingo = []
sway2 = []
zero = []
for i in range(100):
self.problem.g = self.episodes[i].init_state
end_perm = self.episodes[i].label_perm[self.end_of_episode[i]]
end_label = self.episodes[i].init_state.ndata['label'][end_perm]
dqn_result = state2QtableKey(end_label.argmax(dim=1).cpu().numpy())
opt_result = state2QtableKey(validation_problem1[i][1])
grd_result = state2QtableKey(validation_problem1[i][2])
if opt_result==dqn_result:
bingo.append(i)
a1 = self.episodes[i].action_seq[-1][0] * 10000 + self.episodes[i].action_seq[-1][1]
a2 = self.episodes[i].action_seq[-2][0] * 10000 + self.episodes[i].action_seq[-2][1]
if a1==a2 and a1 > 0:
sway2.append(i)
elif a1==0 and a2==0:
zero.append(i)
# hard validation set
test.run_test(problem=to_cuda(bg_hard), trial_num=1, batch_size=100, gnn_step=gnn_step,
episode_len=episode_len, explore_prob=0.0, Temperature=1e-8)
def run_test(self,
problem=None,
init_trial=10,
trial_num=1,
batch_size=100,
gnn_step=3,
episode_len=50,
explore_prob=0.1,
Temperature=1.0,
aux_model=None,
beta=0):
self.episode_len = episode_len
self.num_actions = self.problem.get_legal_actions(
action_type=self.action_type).shape[0]
if isinstance(self.problem.g, BatchedDGLGraph):
self.num_actions //= self.problem.g.batch_size
# if self.action_type == 'swap':
# self.num_actions = self.problem.N * (self.problem.N - self.problem.m) // 2 + 1
# else:
# self.num_actions = self.problem.N * (self.problem.k - 1) + 1
self.trial_num = trial_num
self.batch_size = batch_size
if problem is None:
batch_size *= trial_num
bg = self.problem.gen_batch_graph(batch_size=self.batch_size)
self.bg = to_cuda(
self.problem.gen_batch_graph(x=bg.ndata['x'].repeat(
trial_num, 1),
batch_size=batch_size))
gl = dgl.unbatch(self.bg)
test_problem = self.problem
else:
if aux_model is not None and init_trial > 1:
gg = []
gl = dgl.unbatch(problem)
for i in range(batch_size):
init_label_i = self.test_init_state(alg=self.alg,
aux=aux_model,
g_i=i,
t=episode_len,
trial_num=init_trial)
self.problem.g = gl[i]
self.problem.reset_label(label=init_label_i)
gg.append(dc(self.problem.g))
self.bg = dgl.batch(gg)
self.problem.g = self.bg
test_problem = self.problem
else:
# for validation problem set
self.bg = problem
gl = dgl.unbatch(self.bg)
self.problem.g = self.bg
test_problem = self.problem
self.action_mask = torch.tensor(
range(0, self.num_actions * batch_size, self.num_actions)).cuda()
self.S = [
test_problem.calc_S(g=gl[i]).cpu() for i in range(batch_size)
]
ep = [
EpisodeHistory(gl[i],
max_episode_len=episode_len,
action_type='swap') for i in range(batch_size)
]
self.end_of_episode = {}.fromkeys(range(batch_size), episode_len - 1)
loop_start_position = [0] * self.batch_size
loop_remain_index = set(range(self.batch_size))
self.valid_states = [[ep[i].init_state]
for i in range(self.batch_size)]
for i in tqdm(range(episode_len)):
batch_legal_actions = test_problem.get_legal_actions(
state=self.bg, action_type=self.action_type).cuda()
# if self.forbid_revisit and i > 0:
#
# previous_actions = torch.tensor([ep[k].action_seq[i-1][0] * (self.n ** 2) + ep[k].action_seq[i-1][1] for k in range(batch_size)])
# bla = (self.n ** 2) * batch_legal_actions.view(batch_size, -1, 2)[:, :, 0] + batch_legal_actions.view(batch_size, -1, 2)[:, :, 1]
# forbid_action_mask = (bla.t() == previous_actions.cuda())
# batch_legal_actions = batch_legal_actions * (1 - forbid_action_mask.int()).t().flatten().unsqueeze(1)
forbid_action_mask = torch.zeros(batch_legal_actions.shape[0],
1).cuda()
if self.forbid_revisit:
for k in range(batch_size):
cur_state = ep[k].enc_state_seq[
-1] # current state for the k-th episode
forbid_states = set(
ep[k].enc_state_seq[-1 - self.forbid_revisit:-1])
candicate_actions = batch_legal_actions[
k * self.num_actions:(k + 1) * self.num_actions, :]
for j in range(self.num_actions):
cur_state_l = QtableKey2state(cur_state)
cur_state_l[candicate_actions[j][0]] ^= cur_state_l[
candicate_actions[j][1]]
cur_state_l[candicate_actions[j][1]] ^= cur_state_l[
candicate_actions[j][0]]
cur_state_l[candicate_actions[j][0]] ^= cur_state_l[
candicate_actions[j][1]]
if state2QtableKey(cur_state_l) in forbid_states:
forbid_action_mask[j + k * self.num_actions] += 1
batch_legal_actions = batch_legal_actions * (
1 - forbid_action_mask.int()).t().flatten().unsqueeze(1)
if self.q_net == 'mlp':
# bg1 = to_cuda(self.bg)
# bg1.ndata['label'] = bg1.ndata['label'][:,[0,2,1]]
# bg2 = to_cuda(self.bg)
# bg2.ndata['label'] = bg2.ndata['label'][:,[1,0,2]]
# bg3 = to_cuda(self.bg)
# bg3.ndata['label'] = bg3.ndata['label'][:,[1,2,0]]
# bg4 = to_cuda(self.bg)
# bg4.ndata['label'] = bg4.ndata['label'][:,[2,0,1]]
# bg5 = to_cuda(self.bg)
# bg5.ndata['label'] = bg5.ndata['label'][:,[2,1,0]]
# S_a_encoding, h1, h2, Q_sa = self.alg.forward(dgl.batch([self.bg, bg1, bg2, bg3, bg4, bg5]), torch.cat([batch_legal_actions]*6),
# action_type=self.action_type, gnn_step=gnn_step)
S_a_encoding, h1, h2, Q_sa = self.alg.forward(
self.bg,
batch_legal_actions,
action_type=self.action_type,
gnn_step=gnn_step)
# Q_sa = Q_sa.view(6, -1, self.num_actions).sum(dim=0)
else:
S_a_encoding, h1, h2, Q_sa = self.alg.forward_MHA(
self.bg,
batch_legal_actions,
action_type=self.action_type,
gnn_step=gnn_step)
if beta > 0:
self.problem.sample_episode *= self.num_actions
self.problem.gen_step_batch_mask()
new_bg, _ = self.problem.step_batch(
states=dgl.batch([self.bg] * self.num_actions),
action=batch_legal_actions.view(batch_size,
self.num_actions,
2).transpose(0, 1).reshape(
-1, 2),
action_type=self.action_type,
return_sub_reward=False)
self.problem.sample_episode = self.problem.sample_episode // self.num_actions
self.problem.gen_step_batch_mask()
state_eval = self.get_state_eval_from_aux(
new_bg, self.alg, aux_model)
state_eval = state_eval.view(-1, batch_size).t().reshape(
batch_size * self.num_actions)
self.state_eval.append(
state_eval.view(batch_size, self.num_actions))
Q_sa -= state_eval * beta
terminate_episode = (Q_sa.view(-1, self.num_actions).max(
dim=1).values < 0.).nonzero().flatten().cpu().numpy()
for idx in terminate_episode:
if self.end_of_episode[idx] == episode_len - 1:
self.end_of_episode[idx] = i
best_actions = torch.multinomial(
F.softmax(Q_sa.view(-1, self.num_actions) / Temperature),
1).view(-1)
chose_actions = torch.tensor([
x if torch.rand(1) > explore_prob else torch.randint(
high=self.num_actions, size=(1, )).squeeze()
for x in best_actions
]).cuda()
chose_actions += self.action_mask
actions = batch_legal_actions[chose_actions]
new_states, rewards = self.problem.step_batch(
states=self.bg, action=actions, action_type=self.action_type)
R = [reward.item() for reward in rewards]
# enc_states = [state2QtableKey(self.bg.ndata['label'][k * self.n: (k + 1) * self.n].argmax(dim=1).cpu().numpy()) for k in range(batch_size)]
[
ep[k].write(
action=actions[k, :],
action_idx=best_actions[k] - self.action_mask[k],
reward=R[k],
q_val=Q_sa.view(-1, self.num_actions)[k, :],
actions=batch_legal_actions.view(-1, self.num_actions,
2)[k, :, :],
state_enc=None # enc_states[k]
# , sub_reward=sub_rewards[:, k].cpu().numpy()
,
sub_reward=None,
loop_start_position=loop_start_position[k])
for k in range(batch_size)
]
# # write valid sample episode for training aux model
# tmp = set()
# for k in loop_remain_index:
# if enc_states[k] in ep[k].enc_state_seq[-4: -1]:
# loop_start_position[k] = i
# tmp.add(k)
# else:
# label = ep[k].init_state.ndata['label'][ep[k].label_perm[-1]]
# l = label.argmax(dim=1)
# self.problem.g = dc(ep[k].init_state)
# self.problem.reset_label(label=l)
# self.valid_states[k].append(dc(self.problem.g))
# loop_remain_index = loop_remain_index.difference(tmp)
self.episodes = ep
def run_cascade_episode(self, action_type='swap', gnn_step=3):
sum_r = 0
new_graphs = [
to_cuda(
self.problem.gen_batch_graph(
batch_size=self.new_epi_batch_size))
]
new_graphs.extend(
list(
itertools.chain(*[[
tpl.s1 for tpl in self.cascade_replay_buffer[i]
[-self.new_epi_batch_size:]
] for i in range(self.buf_epi_len - 1)])))
bg = to_cuda(dgl.batch(new_graphs))
batch_size = self.new_epi_batch_size * self.buf_epi_len
num_actions = self.problem.get_legal_actions(
action_type=action_type,
action_dropout=self.action_dropout).shape[0]
# recent_states = list(itertools.chain(
# *[[tpl.recent_states for tpl in self.cascade_replay_buffer[t][-self.new_epi_batch_size:]] for t in
# range(self.buf_epi_len - 1)]))
#
# batch_legal_actions = self.prune_actions(bg, recent_states)
batch_legal_actions = self.problem.get_legal_actions(
state=bg,
action_type=self.action_type,
action_dropout=self.action_dropout).cuda()
# epsilon greedy strategy
# TODO: multi-gpu parallelization
_, _, _, Q_sa = self.model(dc(bg),
batch_legal_actions,
action_type=action_type,
gnn_step=gnn_step)
# TODO: can alter explore strength according to kcut_valueS
kcut_valueS = self.problem.calc_batchS(bg=bg)
chosen_actions = self.sample_actions_from_q(Q_sa,
num_actions,
batch_size,
Temperature=self.eps)
actions = batch_legal_actions[chosen_actions]
# update bg inplace and calculate batch rewards
g0 = [g for g in dgl.unbatch(dc(bg))] # current_state
_, rewards, sub_rewards = self.problem.step_batch(
states=bg, action=actions, return_sub_reward=True)
g1 = [g for g in dgl.unbatch(dc(bg))] # after_state
# kcut_valueS_ = self.problem.calc_batchS(bg=bg)
# print('assert=0:', kcut_valueS_ - kcut_valueS + rewards)
# rewards_ = rewards.reshape(self.buf_epi_len, -1)
# posi_r = rewards_ >= 0
# nege_r = rewards_ < 0
# scaled_rewards = (rewards_-torch.mean(rewards_, dim=1).t().unsqueeze(1)) / torch.std(rewards_, dim=1).t().unsqueeze(1)
# scaled_rewards = scaled_rewards * posi_r + rewards_ * nege_r
#
# scaled_rewards = scaled_rewards.view(-1)
# update episode best S
best_S = torch.cat([kcut_valueS[:self.new_epi_batch_size].cpu() \
, torch.tensor([[np.min([kcut_valueS[j+t*self.new_epi_batch_size], self.cascade_replay_buffer[t - 1][j-self.new_epi_batch_size].best_S]) for j in range(self.new_epi_batch_size)] for t in range(1, self.buf_epi_len)]).flatten()])
# update recent states
recent_states = [[
state2QtableKey(g.ndata['label'].argmax(dim=1).cpu().numpy())
] for g in g0[:self.new_epi_batch_size]]
recent_states.extend(
list(
itertools.chain(*[[
self.cascade_replay_buffer[t - 1]
[j - self.new_epi_batch_size].recent_states + [
state2QtableKey(g0[j + t * self.new_epi_batch_size].
ndata['label'].argmax(
dim=1).cpu().numpy())
] for j in range(self.new_epi_batch_size)
] for t in range(1, self.buf_epi_len)])))
memory_len = 10
recent_states = [state[-memory_len:] for state in recent_states]
forbid_actions = []
[
self.cascade_replay_buffer[t].extend([
sars(g0[j + t * self.new_epi_batch_size],
actions[j + t * self.new_epi_batch_size],
rewards[j + t * self.new_epi_batch_size],
sub_rewards[:, j + t * self.new_epi_batch_size],
g1[j + t * self.new_epi_batch_size],
kcut_valueS[j + t * self.new_epi_batch_size],
best_S[j + t * self.new_epi_batch_size], t,
recent_states[j + t * self.new_epi_batch_size],
forbid_actions) for j in range(self.new_epi_batch_size)
]) for t in range(self.buf_epi_len)
]
self.cascade_buffer_kcut_value = torch.cat([
self.cascade_buffer_kcut_value,
torch.abs(kcut_valueS.detach().cpu().view(self.buf_epi_len,
self.new_epi_batch_size))
],
dim=1).detach()
if self.priority_sampling:
# compute prioritized weights
batch_legal_actions = self.problem.get_legal_actions(
state=bg,
action_type=action_type,
action_dropout=self.action_dropout).cuda()
_, _, _, Q_sa_next = self.model(dc(bg),
batch_legal_actions,
action_type=action_type,
gnn_step=gnn_step)
delta = Q_sa[chosen_actions] - (
rewards +
self.gamma * Q_sa_next.view(-1, num_actions).max(dim=1).values)
# delta = (Q_sa[chosen_actions] - (rewards + self.gamma * Q_sa_next.view(-1, num_actions).max(dim=1).values)) / torch.clamp(torch.abs(Q_sa[chosen_actions]),0.1)
self.cascade_replay_buffer_weight = torch.cat([
self.cascade_replay_buffer_weight,
torch.abs(delta.detach().cpu().view(self.buf_epi_len,
self.new_epi_batch_size))
],
dim=1).detach()
# print(self.cascade_replay_buffer_weight)
R = [reward.item() for reward in rewards]
sum_r += sum(R)
self.log.add_item('tot_return', sum_r)
return R
