print("[before training][OPA generates %d]" % opares)
load_fname = "globalRL-p%d-t%d-d%d-l" % (args.num_procs, args.num_tasks,
args.use_deadline)
tmp = torch.load("../Pandamodels/globalrlmodels/" + load_fname +
".torchmodel").cuda()
rl_model = Solver(args.num_procs,
args.embedding_size,
args.hidden_size,
args.num_tasks,
use_deadline=False,
use_cuda=True)
rl_model.load_state_dict(tmp.state_dict())
if args.use_cuda:
rl_model.cuda()
"""Freeze the weight of the global reinforcement model"""
freezing_param_name = ["init_w", "embedding", "mha"]
for name, param in rl_model.named_parameters():
if name.split(".")[1] in freezing_param_name:
param.requires_grad = False
"""Evaluate global model before the training"""
rl_model.eval()
ret = []
for i, batch in eval_loader:
if use_cuda:
batch = batch.cuda()
R, log_prob, actions = rl_model(batch, argmax=True)
for j, chosen in enumerate(actions.cpu().numpy()):
order = np.zeros_like(chosen)
for k in range(args.num_tasks):
print("[before training][OPA generates %d]" % opares)
temp_fname = "localRL-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)
model = torch.load("../Pandamodels/localrlmodels/" + temp_fname).cuda()
rl_model = Solver(args.num_procs,
args.embedding_size,
args.hidden_size,
args.num_tasks,
use_deadline=False,
use_cuda=True)
rl_model.load_state_dict(model.state_dict())
if use_cuda:
model = model.cuda()
rl_model = rl_model.cuda()
rl_model = rl_model.eval()
ret = []
# 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:
pin_memory=use_pin_memory)
eval_loader = DataLoader(test_dataset,
batch_size=args.num_test_dataset,
shuffle=False)
# Calculating heuristics
model = Solver(args.num_procs,
args.embedding_size,
args.hidden_size,
args.num_tasks,
use_deadline=use_deadline)
if args.use_cuda:
model = model.cuda()
# Train loop
moving_avg = torch.zeros(args.num_train_dataset)
if args.use_cuda:
moving_avg = moving_avg.cuda()
#generating first baseline
cc = 1
for (indices, sample_batch) in tqdm(train_data_loader):
if args.use_cuda:
sample_batch = sample_batch.cuda()
rewards, _, _ = model(sample_batch)
print(rewards)
moving_avg[indices] = rewards.float()
model.eval()
ret = []
print("No previous model!")
model = Solver(args.num_procs,
args.embedding_size,
args.hidden_size,
args.num_tasks,
use_deadline=False,
use_cuda=use_cuda)
bl_model = Solver(args.num_procs,
args.embedding_size,
args.hidden_size,
args.num_tasks,
use_deadline=False,
use_cuda=use_cuda)
bl_model.load_state_dict(model.state_dict())
if use_cuda:
model = model.cuda()
bl_model = bl_model.cuda()
bl_model = bl_model.eval()
def wrap(x):
_sample, num_proc, use_deadline = x
return heu.OPA(
_sample, num_proc, None,
use_deadline) # 여기서 OPA에 테스트 안넘겼는데 까고 들어가보면 DA어쩌구 테스트를 사용함.
with ProcessPoolExecutor(max_workers=4) as executor:
inputs = []
res_opa = np.zeros(len(test_dataset), dtype=int).tolist()
for i, sample in test_dataset:
#ret = heu.OPA(sample, args.num_procs, heu.test_DA_LC, use_deadline)
def kl_div(n_step):
util_range = get_util_range(args.num_procs)
trsets = []
tesets = []
on = False
for util in util_range:
on = False
if util == args.range_l:
on = True
if on:
if positive:
load_file_name = "../Pandadata/tr/%d-%d/positive/%s"
else:
load_file_name = "../Pandadata/tr/%d-%d/%s"
with open(load_file_name % (args.num_procs, args.num_tasks, util), 'rb') as f:
ts = pickle.load(f)
trsets.append(ts)
with open("../Pandadata/te/%d-%d/%s" % (args.num_procs, args.num_tasks, util), 'rb') as f:
ts = pickle.load(f)
tesets.append(ts)
if util == args.range_r:
break
train_dataset = Datasets(trsets)
test_dataset = Datasets(tesets)
train_dataset.setlen(args.num_train_dataset)
test_dataset.setlen(args.num_test_dataset)
train_loader = DataLoader(
train_dataset,
batch_size=args.batch_size,
shuffle=True,
pin_memory=True
)
test_loader = DataLoader(
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
)
temp_fname = "localRL-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)
model = torch.load("../Pandamodels/localrlmodels/" + temp_fname).cuda()
rl_model = Solver(
args.num_procs,
args.embedding_size,
args.hidden_size,
args.num_tasks,
use_deadline=False,
use_cuda=True, ret_embedded_vector=True,
)
rl_model.load_state_dict(model.state_dict())
if use_cuda:
model = model.cuda()
rl_model = rl_model.cuda()
rl_model = rl_model.eval()
if use_cuda:
rl_model = rl_model.to("cuda:0")
ss = np.array(list(range(32)))
ss2 = np.array(list(reversed(range(32))))
guide = torch.LongTensor(np.array([ss, ss2], dtype=np.int32)).cuda()
for epoch in range(args.num_epochs):
loss_ = 0
avg_hit = []
for batch_idx, (_, sample_batch) in enumerate(train_loader):
sample_batch = sample_batch[:2, :, :]
_, actions, distributions = rl_model(sample_batch, guide=guide)
break
break
a = distributions[0][5].detach().cpu().numpy()
b = distributions[1][5].detach().cpu().numpy()
print(0.5 * np.sum(np.abs(a - b)))
exit(0)
kl_div = 0
KL_calc = torch.nn.KLDivLoss(reduction="batchmean")
actions = actions.squeeze()
# Timestep +1
if n_step == 1:
for t in range(args.num_tasks - 1):
previous_distribution = distributions[t].squeeze()
sampled_task = actions[t]
previous_distribution[sampled_task] = 0
# renormalized_distribution = previous_distribution
renormalized_distribution = torch.log(previous_distribution / torch.sum(previous_distribution))
next_distribution = distributions[t+1]
kl_div += KL_calc(renormalized_distribution, next_distribution)
# kl_div += torch.nn.KLDivLoss(size_average=False)(renormalized_distribution, next_distribution)
return kl_div / (args.num_tasks-1)
# Timestep +3
elif n_step == 3:
for t in range(args.num_tasks - 3):
prev_distribution = distributions[t].squeeze()
first_sampled_mask = actions[t]
second_sampled_mask = actions[t+1]
third_sampled_mask = actions[t+2]
prev_distribution[first_sampled_mask] = 0
prev_distribution = prev_distribution / torch.sum(prev_distribution)
prev_distribution[second_sampled_mask] = 0
prev_distribution = prev_distribution / torch.sum(prev_distribution)
prev_distribution = prev_distribution.detach().cpu().numpy()
# renormalized_distribution = torch.log(prev_distribution)
rl_next_distribution = distributions[t+2]
rl_next_distribution = rl_next_distribution.detach().cpu().numpy()
kl_div += np.sum(np.abs(prev_distribution - rl_next_distribution))
# kl_div += KL_calc(renormalized_distribution, rl_next_distribution)
return kl_div / (args.num_tasks-3)
# Timestep +5
else:
for t in range(args.num_tasks - 5):
prev_distribution = distributions[t].squeeze()
first_sampled_mask = actions[t]
second_sampled_mask = actions[t+1]
third_sampled_mask = actions[t+2]
fourth_sampled_mask = actions[t+3]
fifth_sampled_mask = actions[t+4]
prev_distribution[first_sampled_mask] = 0
prev_distribution = prev_distribution / torch.sum(prev_distribution)
prev_distribution[second_sampled_mask] = 0
prev_distribution = prev_distribution / torch.sum(prev_distribution)
prev_distribution[third_sampled_mask] = 0
prev_distribution = prev_distribution / torch.sum(prev_distribution)
prev_distribution[fourth_sampled_mask] = 0
prev_distribution = prev_distribution / torch.sum(prev_distribution)
prev_distribution[fifth_sampled_mask] = 0
prev_distribution = prev_distribution / torch.sum(prev_distribution)
renormalized_distribution = torch.log(prev_distribution)
rl_next_distribution = distributions[t+5]
kl_div += KL_calc(renormalized_distribution, rl_next_distribution)
return kl_div / (args.num_tasks - 5)
