def validate(vali_set, model):
N_JOBS = vali_set[0][0].shape[0]
N_MACHINES = vali_set[0][0].shape[1]
from JSSP_Env import SJSSP
from mb_agg import g_pool_cal
from agent_utils import sample_select_action
from agent_utils import greedy_select_action
import numpy as np
import torch
from Params import configs
env = SJSSP(n_j=N_JOBS, n_m=N_MACHINES)
device = torch.device(configs.device)
g_pool_step = g_pool_cal(
graph_pool_type=configs.graph_pool_type,
batch_size=torch.Size([1, env.number_of_tasks, env.number_of_tasks]),
n_nodes=env.number_of_tasks,
device=device)
make_spans = []
# rollout using model
for data in vali_set:
adj, fea, candidate, mask = env.reset(data)
rewards = -env.initQuality
while True:
fea_tensor = torch.from_numpy(np.copy(fea)).to(device)
adj_tensor = torch.from_numpy(np.copy(adj)).to(device).to_sparse()
candidate_tensor = torch.from_numpy(np.copy(candidate)).to(device)
mask_tensor = torch.from_numpy(np.copy(mask)).to(device)
with torch.no_grad():
pi, _ = model(x=fea_tensor,
graph_pool=g_pool_step,
padded_nei=None,
adj=adj_tensor,
candidate=candidate_tensor.unsqueeze(0),
mask=mask_tensor.unsqueeze(0))
# action = sample_select_action(pi, candidate)
action = greedy_select_action(pi, candidate)
adj, fea, reward, done, candidate, mask = env.step(action.item())
rewards += reward
if done:
break
make_spans.append(rewards - env.posRewards)
# print(rewards - env.posRewards)
return np.array(make_spans)
SEED = 11
np.random.seed(SEED)
env = SJSSP(n_j=n_j, n_m=n_m)
# rollout env random action
t1 = time.time()
data = uni_instance_gen(n_j=n_j, n_m=n_m, low=low, high=high)
dur = np.array([[83, 65, 3], [69, 42, 64], [27, 27, 18]])
mch = np.array([[3, 2, 1], [1, 2, 3], [2, 1, 3]])
# data = (dur, mch)
print('Dur')
print(data[0])
print('Mach')
print(data[-1])
print()
_, _, omega, mask = env.reset(data)
# print('Init end time')
# print(env.LBs)
# print()
rewards = [-env.initQuality]
while True:
action = np.random.choice(omega[~mask])
# print('action:', action)
adj, _, reward, done, omega, mask = env.step(action)
rewards.append(reward)
# print('ET after action:\n', env.LBs)
# print(fea)
# print()
if env.done():
break
t2 = time.time()
g_pool_step = g_pool_cal(graph_pool_type=configs.graph_pool_type,
batch_size=torch.Size(
[1, env.number_of_tasks, env.number_of_tasks]),
n_nodes=env.number_of_tasks,
device=device)
dataLoaded = np.load('./BenchDataNmpy/' + benchmark + str(N_JOBS_P) + 'x' +
str(N_MACHINES_P) + '.npy')
dataset = []
for i in range(dataLoaded.shape[0]):
dataset.append((dataLoaded[i][0], dataLoaded[i][1]))
result = []
t1 = time.time()
for i, data in enumerate(dataset):
adj, fea, candidate, mask = env.reset(data)
ep_reward = -env.max_endTime
while True:
# Running policy_old:
fea_tensor = torch.from_numpy(np.copy(fea)).to(device)
adj_tensor = torch.from_numpy(np.copy(adj)).to(device).to_sparse()
candidate_tensor = torch.from_numpy(np.copy(candidate)).to(device)
mask_tensor = torch.from_numpy(np.copy(mask)).to(device)
with torch.no_grad():
pi, _ = ppo.policy(x=fea_tensor,
graph_pool=g_pool_step,
padded_nei=None,
adj=adj_tensor,
candidate=candidate_tensor.unsqueeze(0),
mask=mask_tensor.unsqueeze(0))