# for i, _batch in eval_loader:
# if use_cuda:
# _batch = _batch.cuda()
# R, log_prob, actions = model(_batch, argmax=True)
# for j, chosen in enumerate(actions.cpu().numpy()):
# order = np.zeros_like(chosen)
# for p in range(args.num_tasks):
# order[chosen[p]] = args.num_tasks - p - 1
# if use_cuda:
# ret.append(test_module(_batch[j].cpu().numpy(), args.num_procs, order, use_deadline, False))
# else:
# ret.append(test_module(_batch[j].numpy(), args.num_procs, order, use_deadline, False))
# print("[Before training][RL model generates %d]" % (np.sum(ret)))
linear_model = LinearSolver(args.num_procs, args.num_tasks,
args.use_deadline, use_cuda)
# TRAIN LOOP
if use_cuda:
linear_model = linear_model.to("cuda:0")
rl_model = rl_model.to("cuda:0")
linear_model = linear_model.train()
criterion = nn.MSELoss()
optimizer = optim.Adam(linear_model.parameters(), lr=5e-3)
start = time.time()
for epoch in range(args.num_epochs):
loss_ = 0
avg_hit = []
for batch_idx, (_, sample_batch) in enumerate(train_loader):
else:
ret.append(
test_module(_batch[j].numpy(), args.num_procs, order,
use_deadline, False))
rl_order = order
print("[Before training][RL model generates %d]" % (np.sum(ret)))
temp_fname = "LIN-p%d-t%d-d%d-l[%s, %s].torchmodel" % (
args.num_procs, args.num_tasks, int(use_deadline), args.range_l,
args.range_r)
try:
model = torch.load("linearmodels/" + temp_fname).cuda()
except:
raise AssertionError("Loading Error")
linear_model = LinearSolver(args.num_procs, args.num_tasks,
args.use_deadline, args.use_cuda)
linear_model.load_state_dict(model.state_dict())
linear_model = linear_model.to("cuda:0")
# EVALUATE
linear_model.eval()
lin_ret = []
for i, _batch in eval_loader:
if use_cuda:
_batch = _batch.to("cuda:0")
ev_linear_score = linear_model(_batch)
_, ev_linear_score_idx = torch.sort(ev_linear_score, descending=True)
np_linear_score = ev_linear_score_idx.cpu().detach().numpy()
for j, chosen in enumerate(np_linear_score):
# for i, _batch in eval_loader:
# if use_cuda:
# _batch = _batch.cuda()
# R, log_prob, actions = model(_batch, argmax=True)
# for j, chosen in enumerate(actions.cpu().numpy()):
# order = np.zeros_like(chosen)
# for p in range(args.num_tasks):
# order[chosen[p]] = args.num_tasks - p - 1
# if use_cuda:
# ret.append(test_module(_batch[j].cpu().numpy(), args.num_procs, order, use_deadline, False))
# else:
# ret.append(test_module(_batch[j].numpy(), args.num_procs, order, use_deadline, False))
#
# print("[Before training][RL model generates %d]" % (np.sum(ret)))
linear_model = LinearSolver(args.num_procs, args.num_tasks,
args.use_deadline, args.use_cuda)
# linear_name = "LIN-p%d-t%d-d%d-l[%s, %s].torchmodel" % (args.num_procs, args.num_tasks, int(use_deadline), args.range_l, args.range_r)
# with open("../Pandamodels/linearmodels/" + linear_name, "rb") as f:
# tmp = torch.load(f)
# linear_model.load_state_dict(tmp.state_dict())
# 밑에꺼 그냥 주석제거하고 트레이닝 하면됨.
# ## TRAIN LOOP ##
if args.use_cuda:
linear_model = linear_model.to("cuda:0")
rl_model = rl_model.to("cuda:0")
linear_model = linear_model.train()
optimizer = optim.Adam(linear_model.parameters(), lr=1e-4)
test_dataset,
batch_size=args.batch_size,
shuffle=False,
pin_memory=True
)
eval_loader = DataLoader(
test_dataset,
batch_size=args.batch_size,
shuffle=False,
pin_memory=True
)
dist_file_name = "LIN-p%d-t%d-d%d-l[%s, %s]" % (
args.num_procs, args.num_tasks, int(use_deadline), args.range_l, args.range_r)
Distilation = LinearSolver(args.num_procs, args.num_tasks,
args.use_deadline, False)
with open("../Pandamodels/linearmodels/" + dist_file_name + ".torchmodel", "rb") as f:
tmp = torch.load(f)
Distilation.load_state_dict(tmp.state_dict())
Distilation.eval()
start = time.time()
distil_ret = []
for i, batch in eval_loader:
linear_score = Distilation(batch)
gumbel_score = sample_gumbel(linear_score, sampling_number=5) # [batch_size x num_gumbel_sample x num_tasks]
gumbel_rank = get_rank(gumbel_score) # [batch_size x num_gumbel_sample x num_tasks]
val = 0
for j, order in enumerate(gumbel_rank): # j : ~batch size
tmp_ret = []