def prepare_optimizee(args, sgd_in_names, obs_shape, hidden_size, actor_critic, current_optimizee_step, prev_optimizee_step):
prev_optimizee_step += current_optimizee_step
current_optimizee_step = 0
##### Vgg16 #####
vgg = vgg16(pretrained=True)
model = FCN_Vgg(n_class=args.num_class)
model.copy_params_from_vgg16(vgg)
###################
# Setup optimizer
sgd_in = [
{'params': get_params(model, ["conv1_1", "conv1_2"]), 'lr': args.lr},
{'params': get_params(model, ["conv2_1", "conv2_2"]), 'lr': args.lr},
{'params': get_params(model, ["conv3_1", "conv3_2", "conv3_3"]), 'lr': args.lr},
{'params': get_params(model, ["conv4_1", "conv4_2", "conv4_3"]), 'lr': args.lr},
{'params': get_params(model, ["conv5_1", "conv5_2", "conv5_3"]), 'lr': args.lr},
{'params': get_params(model, ["fc6", "fc7"]), 'lr': args.lr},
{'params': get_params(model, ["score_fr", "score_pool3", "score_pool4", "upscore2", "upscore8", "upscore_pool4"]), 'lr': args.lr},
]
optimizer = torch.optim.SGD(sgd_in, lr=args.lr, momentum=args.momentum, weight_decay=args.weight_decay)
# Specify which GPUs to use
model = model.cuda()
model.eval()
return model, optimizer, current_optimizee_step, prev_optimizee_step
def get_architecture(arch: str) -> torch.nn.Module:
""" Return a neural network (with random weights)
:param arch: the architecture - should be in the ARCHITECTURES list above
:return: a Pytorch module
"""
if arch == 'lenet':
model = lenet()
elif arch == 'alexnet':
model = alexnet()
elif arch == 'resnet20':
model = resnet20()
elif arch == 'resnt26':
model = resnet26()
elif arch == 'resnet32':
model = resnet32()
elif arch == 'resnet110':
model = resnet110()
elif arch == 'densenet':
model = densenet_BC_cifar(depth=100, k=12)
elif arch == 'vgg16':
model = vgg16()
elif arch == 'vgg19':
model = vgg19()
else:
raise ValueError('arch not in ARCHITECTURES')
return model
def main():
# set GPU ID
os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu
cudnn.benchmark = True
# check save path
save_path = args.save_path
if not os.path.exists(save_path):
os.makedirs(save_path)
# make dataloader
train_loader, test_loader, \
test_onehot, test_label = dataset.get_loader(args.data,
args.data_path,
args.batch_size)
# set num_class
if args.data == 'cifar100':
num_class = 100
else:
num_class = 10
# set num_classes
model_dict = {
"num_classes": num_class,
}
# set network
if args.model == 'res':
model = resnet.resnet110(**model_dict).cuda()
elif args.model == 'dense':
model = densenet_BC.DenseNet3(depth=100,
num_classes=num_class,
growth_rate=12,
reduction=0.5,
bottleneck=True,
dropRate=0.0).cuda()
elif args.model == 'vgg':
model = vgg.vgg16(**model_dict).cuda()
# set criterion
cls_criterion = nn.CrossEntropyLoss().cuda()
# make logger
result_logger = utils.Logger(os.path.join(save_path, 'result.log'))
# load pretrained model
model_state_dict = torch.load(os.path.join(args.save_path,
'{0}.pth'.format(args.file_name)))
model.load_state_dict(model_state_dict)
# calc measure
acc, aurc, eaurc, aupr, fpr, ece, nll, brier = metrics.calc_metrics(test_loader,
test_label,
test_onehot,
model,
cls_criterion)
# result write
result_logger.write([acc,aurc*1000,eaurc*1000,aupr*100,fpr*100,ece*100,nll*10,brier*100])
def __init__(self, device):
super().__init__()
self.faster = fr.FasterRCNN()
self.lower_module = lm.lower_module(device)
self.upper_module = um.upper_module()
self.vgg = vgg.vgg16()
self.transform = transforms.Compose([
transforms.ToTensor(),
])
self.predict = nn.Linear(c.Concat_vec_len * 2, c.HOIlen)
with open('data/train_bbox.json', 'r') as f:
self.train_bboxes = json.load(f)
with open('data/test_bbox.json', 'r') as f:
self.test_bboxes = json.load(f)
def main():
# set GPU ID
os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu
cudnn.benchmark = True
# check save path
save_path = args.save_path
if not os.path.exists(save_path):
os.makedirs(save_path)
# make dataloader
train_loader, test_loader, \
test_onehot, test_label = dataset.get_loader(args.data,
args.data_path,
args.batch_size)
# set num_class
if args.data == 'cifar100':
num_class = 100
else:
num_class = 10
# set num_classes
model_dict = {
"num_classes": num_class,
}
# set model
if args.model == 'res':
model = resnet.resnet110(**model_dict).cuda()
elif args.model == 'dense':
model = densenet_BC.DenseNet3(depth=100,
num_classes=num_class,
growth_rate=12,
reduction=0.5,
bottleneck=True,
dropRate=0.0).cuda()
elif args.model == 'vgg':
model = vgg.vgg16(**model_dict).cuda()
# set criterion
cls_criterion = nn.CrossEntropyLoss().cuda()
ranking_criterion = nn.MarginRankingLoss(margin=0.0).cuda()
# set optimizer (default:sgd)
optimizer = optim.SGD(model.parameters(),
lr=0.1,
momentum=0.9,
weight_decay=0.0001,
nesterov=False)
# set scheduler
scheduler = MultiStepLR(optimizer, milestones=[150, 250], gamma=0.1)
# make logger
train_logger = utils_orig.Logger(os.path.join(save_path, 'train.log'))
result_logger = utils_orig.Logger(os.path.join(save_path, 'result.log'))
# make History Class
correctness_history = crl_utils.History(len(train_loader.dataset))
# start Train
for epoch in range(1, args.epochs + 1):
scheduler.step()
train.train(train_loader, model, cls_criterion, ranking_criterion,
optimizer, epoch, correctness_history, train_logger, args)
# save model
if epoch == args.epochs:
torch.save(model.state_dict(),
os.path.join(save_path, 'model.pth'))
# finish train
# calc measure
acc, aurc, eaurc, aupr, fpr, ece, nll, brier = metrics.calc_metrics(
test_loader, test_label, test_onehot, model, cls_criterion)
# result write
result_logger.write([
acc, aurc * 1000, eaurc * 1000, aupr * 100, fpr * 100, ece * 100,
nll * 10, brier * 100
])
import model.vgg as vgg
import cv2
from torchvision import transforms
model = vgg.vgg16().cuda(2)
transform = transforms.Compose([
transforms.ToTensor(),
])
data = cv2.imread('input_img/000000035005.jpg')
data = transform(data).cuda(2)
print(data.shape)
print(model(data.unsqueeze_(0)))
def main():
args = get_args()
pid = os.getpid()
device = torch.device("cuda:0" if args.cuda else "cpu")
if args.cuda and torch.cuda.is_available() and args.cuda_deterministic:
torch.backends.cudnn.benchmark = False
torch.backends.cudnn.deterministic = True
best_mIoU = 0
#### preparation ###########################################
from config_seg import config as data_setting
data_setting.batch_size = args.batch_size
# if args.src_list is not None:
# data_setting.train_source = args.src_list
# if args.tgt_list is not None:
# data_setting.eval_source = args.tgt_list
train_loader = get_train_loader(data_setting, GTA5, test=False)
train_loader_iter = iter(train_loader)
current_optimizee_step, prev_optimizee_step = 0, 0
model_old = None
if args.lwf:
# create a fixed model copy for Life-long learning
##### Vgg16 #####
model_old = vgg16(pretrained=True)
###################
model_old.eval()
model_old.to(device)
############################################################
### Agent Settings ########################################
RANDOM = False # False | True | 'init'
action_space = np.arange(0, 1.1, 0.1)
# action_space = np.arange(0, 3); granularity = 0.01
obs_avg = True
_window_size = 1
window_size = 1 if obs_avg else _window_size
window_shrink_size = 20 # larger: controller will be updated more frequently w.r.t. optimizee_step
sgd_in_names = ["conv1", "conv2", "conv3", "conv4", "conv5", "FC", "fc_new"]
coord_size = len(sgd_in_names)
ob_name_lstm = ["loss", "loss_kl", "step", "fc_mean", "fc_std"]
ob_name_scalar = []
obs_shape = (len(ob_name_lstm) * window_size + len(ob_name_scalar) + coord_size, )
_hidden_size = 20
hidden_size = _hidden_size * len(ob_name_lstm)
actor_critic = Policy(coord_size, input_size=(len(ob_name_lstm), len(ob_name_scalar)), action_space=len(action_space), hidden_size=_hidden_size, window_size=window_size)
actor_critic.to(device)
actor_critic.eval()
partial = torch.load("./pretrained/policy_vgg16_segmentation.pth", map_location=lambda storage, loc: storage)
state = actor_critic.state_dict()
pretrained_dict = {k: v for k, v in partial.items()}
state.update(pretrained_dict)
actor_critic.load_state_dict(state)
if args.algo == 'reinforce':
agent = algo.REINFORCE(
actor_critic,
args.entropy_coef,
lr=args.lr_meta,
eps=args.eps,
alpha=args.alpha,
max_grad_norm=args.max_grad_norm)
elif args.algo == 'a2c':
agent = algo.A2C_ACKTR(
actor_critic,
args.value_loss_coef,
args.entropy_coef,
lr=args.lr_meta,
eps=args.eps,
alpha=args.alpha,
max_grad_norm=args.max_grad_norm)
elif args.algo == 'ppo':
agent = algo.PPO(
actor_critic,
args.clip_param,
args.ppo_epoch,
args.num_mini_batch,
args.value_loss_coef,
args.entropy_coef,
lr=args.lr_meta,
eps=args.eps,
max_grad_norm=args.max_grad_norm)
elif args.algo == 'acktr':
agent = algo.A2C_ACKTR(actor_critic, args.value_loss_coef, args.entropy_coef, acktr=True)
################################################################
_min_iter = 20
# reset optmizee
model, optimizer, current_optimizee_step, prev_optimizee_step = prepare_optimizee(args, sgd_in_names, obs_shape, hidden_size, actor_critic, current_optimizee_step, prev_optimizee_step)
##### Logging ###########################
# Log outputs
if RANDOM:
args.name = "Random_GTA5_Min%diter.Step%d.Window%d_batch%d_Epoch%d_LR%.1e.warmpoly_lwf.%d"%\
(_min_iter, args.num_steps, window_shrink_size, args.batch_size, args.epochs, args.lr, args.lwf)
else:
args.name = "metatrain_GTA5_%s.SGD.Gamma%.1f.LRmeta.%.1e.Hidden%d.Loss.avg.exp.Earlystop.%d.Min%diter.Step%d.Window%d_batch%d_Epoch%d_LR%.1e.warmpoly_lwf.%d"%\
(args.algo, args.gamma, args.lr_meta, _hidden_size, args.early_stop, _min_iter, args.num_steps, window_shrink_size, args.batch_size, args.epochs, args.lr, args.lwf)
if args.resume:
args.name += "_resumed"
if args.cuda and torch.cuda.is_available() and args.cuda_deterministic:
torch.backends.cudnn.benchmark = False
torch.backends.cudnn.deterministic = True
# Log outputs
directory = "runs/%s/"%(args.name)
if not os.path.exists(directory):
os.makedirs(directory)
filename = directory + 'train.log'
for handler in logging.root.handlers[:]:
logging.root.removeHandler(handler)
rootLogger = logging.getLogger()
logFormatter = logging.Formatter("%(asctime)s [%(levelname)-5.5s] %(message)s")
fileHandler = logging.FileHandler(filename)
fileHandler.setFormatter(logFormatter)
rootLogger.addHandler(fileHandler)
consoleHandler = logging.StreamHandler()
consoleHandler.setFormatter(logFormatter)
rootLogger.addHandler(consoleHandler)
rootLogger.setLevel(logging.INFO)
writer = SummaryWriter(directory)
###########################################
threds = 1
evaluator = SegEvaluator(Cityscapes(data_setting, 'val', None), args.num_class, np.array([0.485, 0.456, 0.406]),
np.array([0.229, 0.224, 0.225]), model, [1, ], False, devices=0, config=data_setting, threds=threds,
verbose=False, save_path=None, show_image=False)
epoch_size = len(train_loader)
total_steps = epoch_size*args.epochs
bar_format = '{desc}[{elapsed}<{remaining},{rate_fmt}]'
pbar = tqdm(range(int(epoch_size*args.epochs)), file=sys.stdout, bar_format=bar_format, ncols=100)
_window_size = max(_min_iter, current_optimizee_step + prev_optimizee_step // window_shrink_size)
train_loader_iter, obs, loss, loss_kl, fc_mean, fc_std = train_step(args, _window_size, train_loader_iter, train_loader, model, optimizer, obs_avg, args.lr, pbar, current_optimizee_step + prev_optimizee_step, total_steps, model_old=model_old)
writer.add_scalar("loss/ce", loss, current_optimizee_step + prev_optimizee_step)
writer.add_scalar("loss/kl", loss_kl, current_optimizee_step + prev_optimizee_step)
writer.add_scalar("loss/total", loss + loss_kl, current_optimizee_step + prev_optimizee_step)
writer.add_scalar("fc/mean", fc_mean, current_optimizee_step + prev_optimizee_step)
writer.add_scalar("fc/std", fc_std, current_optimizee_step + prev_optimizee_step)
current_optimizee_step += _window_size
prev_obs = obs.unsqueeze(0)
prev_hidden = torch.zeros(actor_critic.net.num_recurrent_layers, 1, hidden_size).cuda()
for epoch in range(args.epochs):
print("\n===== Epoch %d / %d ====="%(epoch+1, args.epochs))
print("============= " + args.name + " ================")
print("============= PID: " + str(pid) + " ================")
while current_optimizee_step < epoch_size:
# Sample actions
with torch.no_grad():
if not RANDOM:
value, action, action_log_prob, recurrent_hidden_states, distribution = actor_critic.act(prev_obs, prev_hidden, deterministic=False)
action = action.squeeze(0)
action_log_prob = action_log_prob.squeeze(0)
value = value.squeeze(0)
for idx in range(len(action)):
writer.add_scalar("action/%s"%(sgd_in_names[idx]), action[idx], current_optimizee_step + prev_optimizee_step)
writer.add_scalar("entropy/%s"%(sgd_in_names[idx]), distribution.distributions[idx].entropy(), current_optimizee_step + prev_optimizee_step)
optimizer.param_groups[idx]['lr'] = float(action_space[action[idx]]) * args.lr
writer.add_scalar("LR/%s"%(sgd_in_names[idx]), optimizer.param_groups[idx]['lr'], current_optimizee_step + prev_optimizee_step)
else:
if RANDOM is True or RANDOM == 'init':
for idx in range(coord_size):
optimizer.param_groups[idx]['lr'] = float(choice(action_space)) * args.lr
if RANDOM == 'init':
RANDOM = 'done'
for idx in range(coord_size):
writer.add_scalar("LR/%s"%sgd_in_names[idx], optimizer.param_groups[idx]['lr'], current_optimizee_step + prev_optimizee_step)
# Obser reward and next obs
_window_size = max(_min_iter, current_optimizee_step + prev_optimizee_step // window_shrink_size)
_window_size = min(_window_size, epoch_size - current_optimizee_step)
train_loader_iter, obs, loss, loss_kl, fc_mean, fc_std = train_step(args, _window_size, train_loader_iter, train_loader, model, optimizer, obs_avg, args.lr, pbar, current_optimizee_step + prev_optimizee_step, total_steps, model_old=model_old)
writer.add_scalar("loss/ce", loss, current_optimizee_step + prev_optimizee_step)
writer.add_scalar("loss/kl", loss_kl, current_optimizee_step + prev_optimizee_step)
writer.add_scalar("loss/total", loss + loss_kl, current_optimizee_step + prev_optimizee_step)
writer.add_scalar("fc/mean", fc_mean, current_optimizee_step + prev_optimizee_step)
writer.add_scalar("fc/std", fc_std, current_optimizee_step + prev_optimizee_step)
current_optimizee_step += _window_size
prev_obs = obs.unsqueeze(0)
if not RANDOM: prev_hidden = recurrent_hidden_states
prev_optimizee_step += current_optimizee_step
current_optimizee_step = 0
# evaluate on validation set
torch.cuda.empty_cache()
mIoU = validate(evaluator, model)
writer.add_scalar("mIoU", mIoU, epoch)
# remember best prec@1 and save checkpoint
is_best = mIoU > best_mIoU
best_mIoU = max(mIoU, best_mIoU)
save_checkpoint(args.name, {
'epoch': epoch + 1,
'state_dict': model.state_dict(),
'best_mIoU': best_mIoU,
}, is_best)
logging.info('Best accuracy: {mIoU:.3f}'.format(mIoU=best_mIoU))
def main():
global args, best_mIoU
args = parser.parse_args()
pid = os.getpid()
# Log outputs
args.name = "GTA5_Vgg16_batch%d_512x512_Poly_LR%.1e_1to%.1f_all_lwf.%d_epoch%d" % (
args.batch_size, args.lr, args.factor, args.lwf, args.epochs)
if args.resume:
args.name += "_resumed"
directory = "runs/%s/" % (args.name)
if not os.path.exists(directory):
os.makedirs(directory)
filename = directory + 'train.log'
for handler in logging.root.handlers[:]:
logging.root.removeHandler(handler)
rootLogger = logging.getLogger()
logFormatter = logging.Formatter(
"%(asctime)s [%(levelname)-5.5s] %(message)s")
fileHandler = logging.FileHandler(filename)
fileHandler.setFormatter(logFormatter)
rootLogger.addHandler(fileHandler)
consoleHandler = logging.StreamHandler()
consoleHandler.setFormatter(logFormatter)
rootLogger.addHandler(consoleHandler)
rootLogger.setLevel(logging.INFO)
writer = SummaryWriter(directory)
from config_seg import config as data_setting
data_setting.batch_size = args.batch_size
train_loader = get_train_loader(data_setting, GTA5, test=False)
##### Vgg16 #####
vgg = vgg16(pretrained=True)
model = FCN_Vgg(n_class=args.num_class)
model.copy_params_from_vgg16(vgg)
###################
threds = 1
evaluator = SegEvaluator(Cityscapes(data_setting, 'val', None),
args.num_class,
np.array([0.485, 0.456, 0.406]),
np.array([0.229, 0.224, 0.225]),
model, [
1,
],
False,
devices=args.gpus,
config=data_setting,
threds=threds,
verbose=False,
save_path=None,
show_image=False)
# Setup optimizer
##### Vgg16 #####
sgd_in = [
{
'params': get_params(model, ["conv1_1", "conv1_2"]),
'lr': args.factor * args.lr
},
{
'params': get_params(model, ["conv2_1", "conv2_2"]),
'lr': args.factor * args.lr
},
{
'params': get_params(model, ["conv3_1", "conv3_2", "conv3_3"]),
'lr': args.factor * args.lr
},
{
'params': get_params(model, ["conv4_1", "conv4_2", "conv4_3"]),
'lr': args.factor * args.lr
},
{
'params': get_params(model, ["conv5_1", "conv5_2", "conv5_3"]),
'lr': args.factor * args.lr
},
{
'params': get_params(model, ["fc6", "fc7"]),
'lr': args.factor * args.lr
},
{
'params':
get_params(model, [
"score_fr", "score_pool3", "score_pool4", "upscore2",
"upscore8", "upscore_pool4"
]),
'lr':
args.lr
},
]
base_lrs = [group['lr'] for group in sgd_in]
optimizer = torch.optim.SGD(sgd_in,
lr=args.lr,
momentum=args.momentum,
weight_decay=args.weight_decay)
# Optionally resume from a checkpoint
if args.resume:
if os.path.isfile(args.resume):
print("=> loading checkpoint '{}'".format(args.resume))
checkpoint = torch.load(args.resume)
args.start_epoch = checkpoint['epoch']
best_mIoU = checkpoint['best_mIoU']
model.load_state_dict(checkpoint['state_dict'])
print("=> loaded checkpoint '{}' (epoch {})".format(
args.resume, checkpoint['epoch']))
else:
print("=ImageClassdata> no checkpoint found at '{}'".format(
args.resume))
model = model.cuda()
model_old = None
if args.lwf > 0:
# create a fixed model copy for Life-long learning
model_old = vgg16(pretrained=True)
###################
for param in model_old.parameters():
param.requires_grad = False
model_old.eval()
model_old.cuda()
if args.evaluate:
mIoU = validate(evaluator, model)
print(mIoU)
# Main training loop
iter_max = args.epochs * math.ceil(len(train_loader) / args.iter_size)
iter_stat = IterNums(iter_max)
for epoch in range(args.start_epoch, args.epochs):
logging.info("============= " + args.name + " ================")
logging.info("============= PID: " + str(pid) + " ================")
logging.info("Epoch: %d" % (epoch + 1))
# train for one epoch
train(args,
train_loader,
model,
optimizer,
base_lrs,
iter_stat,
epoch,
writer,
model_old=model_old,
adjust_lr=epoch < args.epochs)
# evaluate on validation set
torch.cuda.empty_cache()
mIoU = validate(evaluator, model)
writer.add_scalar("mIoU", mIoU, epoch)
# remember best mIoU and save checkpoint
is_best = mIoU > best_mIoU
best_mIoU = max(mIoU, best_mIoU)
save_checkpoint(
directory, {
'epoch': epoch + 1,
'state_dict': model.state_dict(),
'best_mIoU': best_mIoU,
}, is_best)
logging.info('Best accuracy: {mIoU:.3f}'.format(mIoU=best_mIoU))
def main():
# set GPU ID
os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu
cudnn.benchmark = True
# check save path
save_path = args.save_path
if not os.path.exists(save_path):
os.makedirs(save_path)
# set num_class
if args.data == 'cifar100':
num_class = 100
else:
num_class = 10
# set num_classes
model_dict = {
"num_classes": num_class,
}
_, test_loader, \
test_onehot, test_label = dataset.get_loader(args.data,
args.data_path,
args.batch_size)
train_set = dataset.get_dataset(args.data, args.data_path, mode='train')
unlabeled_pool = dataset.get_dataset(args.data,
args.data_path,
mode='unlabeled')
num_train = len(train_set)
indices = list(range(num_train))
random.shuffle(indices)
labeled_set = indices[:args.initial_budget]
unlabeled_set = indices[args.initial_budget:]
labeled_dataloader = DataLoader(train_set,
sampler=SubsetRandomSampler(labeled_set),
batch_size=args.batch_size,
drop_last=True)
now = datetime.datetime.now()
formatedDate = now.strftime('%Y%m%d_%H_%M_')
result_logger = utils.Logger(
os.path.join(args.save_path, formatedDate + 'result.log'))
arguments = []
for key, val in (args.__dict__.items()):
arguments.append("{} : {}\n".format(key, val))
result_logger.write(arguments)
result_logger = utils.Logger(
os.path.join(args.save_path, formatedDate + 'result.log'))
# make logger
train_logger = utils.Logger(
os.path.join(save_path, formatedDate + 'train.log'))
test_epoch_logger = utils.Logger(
os.path.join(save_path, formatedDate + 'test_epoch.log'))
current_train = len(labeled_set)
while (current_train < args.max_budget + 1):
# set model
if args.model == 'res':
model = resnet.ResNet152(**model_dict).cuda()
elif args.model == 'dense':
model = densenet_BC.DenseNet3(depth=100,
num_classes=num_class,
growth_rate=12,
reduction=0.5,
bottleneck=True,
dropRate=0.0).cuda()
elif args.model == 'vgg':
model = vgg.vgg16(**model_dict).cuda()
# set criterion
cls_criterion = nn.CrossEntropyLoss().cuda()
ranking_criterion = nn.MarginRankingLoss(margin=0.0).cuda()
# set optimizer (default:sgd)
optimizer = optim.SGD(model.parameters(),
lr=args.lr,
momentum=0.9,
weight_decay=0.0005,
nesterov=False)
# set scheduler
scheduler = MultiStepLR(optimizer, milestones=[120, 160], gamma=0.1)
# make History Class
correctness_history = crl_utils.History(len(
labeled_dataloader.dataset))
# start Train
for epoch in range(1, args.epochs + 1):
train.train(labeled_dataloader, model, cls_criterion,
ranking_criterion, optimizer, epoch,
correctness_history, train_logger, args)
test_acc, test_loss = metrics.evaluate(test_loader, model,
cls_criterion, args.budget,
epoch, test_epoch_logger)
scheduler.step()
# save model
if epoch == args.epochs:
torch.save(model.state_dict(),
os.path.join(save_path, 'model.pth'))
# finish train
# calc measure
acc, aurc, eaurc, aupr, fpr, ece, nll, brier = metrics.calc_metrics(
test_loader, test_label, test_onehot, model, cls_criterion)
# result write
result_logger.write([
current_train, test_acc, aurc * 1000, eaurc * 1000, aupr * 100,
fpr * 100, ece * 100, nll * 10, brier * 100
])
random.shuffle(unlabeled_set)
subset = unlabeled_set[:args.subset]
unlabeled_poolloader = DataLoader(
unlabeled_pool,
sampler=SubsetSequentialSampler(subset),
batch_size=args.batch_size,
drop_last=False)
all_confidence = get_confidence(model, unlabeled_poolloader)
print(len(all_confidence))
arg = np.argsort(all_confidence)
labeled_set = list(
set(labeled_set) | set(np.array(unlabeled_set)[arg][:args.budget]))
unlabeled_set = list(set(unlabeled_set) - set(labeled_set))
current_train = len(labeled_set)
#unlabeled_set = list(torch.tensor(unlabeled_set)[arg][args.budget:].numpy()) \
# + unlabeled_set[args.subset:]
print("after acquistiion")
print('current labeled :', len(labeled_set))
print('current unlabeled :', len(unlabeled_set))
labeled_dataloader = DataLoader(
train_set,
sampler=SubsetRandomSampler(labeled_set),
batch_size=args.batch_size,
drop_last=True)
def main():
file_name = "./flood_graph/150_250/128/500/ji_sort/1_conf/sample-wised/default/{}/".format(
args.b)
start = time.time()
# set GPU ID
os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu
cudnn.benchmark = True
# check save path
save_path = file_name
# save_path = args.save_path
if not os.path.exists(save_path):
os.makedirs(save_path)
# make dataloader
if args.valid == True:
train_loader, valid_loader, test_loader, test_onehot, test_label = dataset.get_valid_loader(
args.data, args.data_path, args.batch_size)
else:
train_loader, train_onehot, train_label, test_loader, test_onehot, test_label = dataset.get_loader(
args.data, args.data_path, args.batch_size)
# set num_class
if args.data == 'cifar100':
num_class = 100
else:
num_class = 10
# set num_classes
model_dict = {
"num_classes": num_class,
}
# set model
if args.model == 'res':
model = resnet.resnet110(**model_dict).cuda()
elif args.model == 'dense':
model = densenet_BC.DenseNet3(depth=100,
num_classes=num_class,
growth_rate=12,
reduction=0.5,
bottleneck=True,
dropRate=0.0).cuda()
elif args.model == 'vgg':
model = vgg.vgg16(**model_dict).cuda()
# set criterion
if args.loss == 'MS':
cls_criterion = losses.MultiSimilarityLoss().cuda()
elif args.loss == 'Contrastive':
cls_criterion = losses.ContrastiveLoss().cuda()
elif args.loss == 'Triplet':
cls_criterion = losses.TripletLoss().cuda()
elif args.loss == 'NPair':
cls_criterion = losses.NPairLoss().cuda()
elif args.loss == 'Focal':
cls_criterion = losses.FocalLoss(gamma=3.0).cuda()
else:
if args.mode == 0:
cls_criterion = nn.CrossEntropyLoss().cuda()
else:
cls_criterion = nn.CrossEntropyLoss(reduction="none").cuda()
ranking_criterion = nn.MarginRankingLoss(margin=0.0).cuda()
# set optimizer (default:sgd)
optimizer = optim.SGD(
model.parameters(),
lr=0.1,
momentum=0.9,
weight_decay=5e-4,
# weight_decay=0.0001,
nesterov=False)
# optimizer = optim.SGD(model.parameters(),
# lr=float(args.lr),
# momentum=0.9,
# weight_decay=args.weight_decay,
# nesterov=False)
# set scheduler
# scheduler = MultiStepLR(optimizer,
# milestones=[500, 750],
# gamma=0.1)
scheduler = MultiStepLR(optimizer, milestones=[150, 250], gamma=0.1)
# scheduler = torch.optim.lr_scheduler.StepLR(optimizer, step_size=args.lr_decay_step, gamma=args.lr_decay_gamma)
# make logger
train_logger = utils.Logger(os.path.join(save_path, 'train.log'))
result_logger = utils.Logger(os.path.join(save_path, 'result.log'))
# make History Class
correctness_history = crl_utils.History(len(train_loader.dataset))
## define matrix
if args.data == 'cifar':
matrix_idx_confidence = [[_] for _ in range(50000)]
matrix_idx_iscorrect = [[_] for _ in range(50000)]
else:
matrix_idx_confidence = [[_] for _ in range(73257)]
matrix_idx_iscorrect = [[_] for _ in range(73257)]
# write csv
#'''
import csv
f = open('{}/logs_{}_{}.txt'.format(file_name, args.b, args.epochs),
'w',
newline='')
f.write("location = {}\n\n".format(file_name) + str(args))
f0 = open('{}/Test_confidence_{}_{}.csv'.format(file_name, args.b,
args.epochs),
'w',
newline='')
# f0 = open('./baseline_graph/150_250/128/500/Test_confidence_{}_{}.csv'.format(args.b, args.epochs), 'w', newline='')
# f0 = open('./CRL_graph/150_250/Test_confidence_{}_{}.csv'.format(args.b, args.epochs), 'w', newline='')
wr_conf_test = csv.writer(f0)
header = [_ for _ in range(args.epochs + 1)]
header[0] = 'Epoch'
wr_conf_test.writerows([header])
f1 = open('{}/Train_confidence_{}_{}.csv'.format(file_name, args.b,
args.epochs),
'w',
newline='')
# f1 = open('./baseline_graph/150_250/128/500/Train_confidence_{}_{}.csv'.format(args.b, args.epochs), 'w', newline='')
# f1 = open('./CRL_graph/150_250/Train_confidence_{}_{}.csv'.format(args.b, args.epochs), 'w', newline='')
wr = csv.writer(f1)
header = [_ for _ in range(args.epochs + 1)]
header[0] = 'Epoch'
wr.writerows([header])
f2 = open('{}/Train_Flood_{}_{}_{}.csv'.format(file_name, args.data,
args.b, args.epochs),
'w',
newline='')
# f2 = open('./baseline_graph/150_250/128/500/Train_Base_{}_{}_{}.csv'.format(args.data, args.b, args.epochs), 'w', newline='')
# f2 = open('./CRL_graph/150_250/Train_Flood_{}_{}_{}.csv'.format(args.data, args.b, args.epochs), 'w', newline='')
wr_train = csv.writer(f2)
header = [_ for _ in range(args.epochs + 1)]
header[0] = 'Epoch'
wr_train.writerows([header])
f3 = open('{}/Test_Flood_{}_{}_{}.csv'.format(file_name, args.data, args.b,
args.epochs),
'w',
newline='')
# f3 = open('./baseline_graph/150_250/128/500/Test_Base_{}_{}_{}.csv'.format(args.data, args.b, args.epochs), 'w', newline='')
# f3 = open('./CRL_graph/150_250/Test_Flood_{}_{}_{}.csv'.format(args.data, args.b, args.epochs), 'w', newline='')
wr_test = csv.writer(f3)
header = [_ for _ in range(args.epochs + 1)]
header[0] = 'Epoch'
wr_test.writerows([header])
#'''
# start Train
best_valid_acc = 0
test_ece_report = []
test_acc_report = []
test_nll_report = []
test_over_con99_report = []
test_e99_report = []
test_cls_loss_report = []
train_ece_report = []
train_acc_report = []
train_nll_report = []
train_over_con99_report = []
train_e99_report = []
train_cls_loss_report = []
train_rank_loss_report = []
train_total_loss_report = []
for epoch in range(1, args.epochs + 1):
scheduler.step()
matrix_idx_confidence, matrix_idx_iscorrect, idx, iscorrect, confidence, target, cls_loss_tr, rank_loss_tr, batch_correctness, total_confidence, total_correctness = \
train.train(matrix_idx_confidence, matrix_idx_iscorrect, train_loader,
model,
wr,
cls_criterion,
ranking_criterion,
optimizer,
epoch,
correctness_history,
train_logger,
args)
if args.rank_weight != 0.0:
print("RANK ", rank_loss_tr)
total_loss_tr = cls_loss_tr + rank_loss_tr
if args.valid == True:
idx, iscorrect, confidence, target, cls_loss_val, acc = train.valid(
valid_loader, model, cls_criterion, ranking_criterion,
optimizer, epoch, correctness_history, train_logger, args)
if acc > best_valid_acc:
best_valid_acc = acc
print("*** Update Best Acc ***")
# save model
if epoch == args.epochs:
torch.save(model.state_dict(),
os.path.join(save_path, 'model.pth'))
print("########### Train ###########")
acc_tr, aurc_tr, eaurc_tr, aupr_tr, fpr_tr, ece_tr, nll_tr, brier_tr, E99_tr, over_99_tr, cls_loss_tr = metrics.calc_metrics(
train_loader, train_label, train_onehot, model, cls_criterion,
args)
if args.sort == True and epoch == 260:
#if args.sort == True:
train_loader = dataset.sort_get_loader(
args.data, args.data_path, args.batch_size, idx,
np.array(target), iscorrect,
batch_correctness, total_confidence, total_correctness,
np.array(confidence), epoch, args)
train_acc_report.append(acc_tr)
train_nll_report.append(nll_tr * 10)
train_ece_report.append(ece_tr)
train_over_con99_report.append(over_99_tr)
train_e99_report.append(E99_tr)
train_cls_loss_report.append(cls_loss_tr)
if args.rank_weight != 0.0:
train_total_loss_report.append(total_loss_tr)
train_rank_loss_report.append(rank_loss_tr)
print("CLS ", cls_loss_tr)
# finish train
print("########### Test ###########")
# calc measure
acc_te, aurc_te, eaurc_te, aupr_te, fpr_te, ece_te, nll_te, brier_te, E99_te, over_99_te, cls_loss_te = metrics.calc_metrics(
test_loader, test_label, test_onehot, model, cls_criterion, args)
test_ece_report.append(ece_te)
test_acc_report.append(acc_te)
test_nll_report.append(nll_te * 10)
test_over_con99_report.append(over_99_te)
test_e99_report.append(E99_te)
test_cls_loss_report.append(cls_loss_te)
print("CLS ", cls_loss_te)
print("############################")
# for idx in matrix_idx_confidence:
# wr.writerow(idx)
#'''
# draw graph
df = pd.DataFrame()
df['epoch'] = [i for i in range(1, args.epochs + 1)]
df['test_ece'] = test_ece_report
df['train_ece'] = train_ece_report
fig_loss = plt.figure(figsize=(35, 35))
fig_loss.set_facecolor('white')
ax = fig_loss.add_subplot()
ax.plot(df['epoch'],
df['test_ece'],
df['epoch'],
df['train_ece'],
linewidth=10)
ax.legend(['Test', 'Train'], loc=2, prop={'size': 60})
plt.title('[FL] ECE per epoch', fontsize=80)
# plt.title('[BASE] ECE per epoch', fontsize=80)
# plt.title('[CRL] ECE per epoch', fontsize=80)
plt.xlabel('Epoch', fontsize=70)
plt.ylabel('ECE', fontsize=70)
plt.ylim([0, 1])
plt.setp(ax.get_xticklabels(), fontsize=30)
plt.setp(ax.get_yticklabels(), fontsize=30)
plt.savefig('{}/{}_{}_ECE_lr_{}.png'.format(file_name, args.model, args.b,
args.epochs))
# plt.savefig('./baseline_graph/150_250/128/500/{}_{}_ECE_lr_{}.png'.format(args.model, args.b, args.epochs))
# plt.savefig('./CRL_graph/150_250/{}_{}_ECE_lr_{}.png'.format(args.model, args.b, args.epochs))
df2 = pd.DataFrame()
df2['epoch'] = [i for i in range(1, args.epochs + 1)]
df2['test_acc'] = test_acc_report
df2['train_acc'] = train_acc_report
fig_acc = plt.figure(figsize=(35, 35))
fig_acc.set_facecolor('white')
ax = fig_acc.add_subplot()
ax.plot(df2['epoch'],
df2['test_acc'],
df2['epoch'],
df2['train_acc'],
linewidth=10)
ax.legend(['Test', 'Train'], loc=2, prop={'size': 60})
plt.title('[FL] Accuracy per epoch', fontsize=80)
# plt.title('[BASE] Accuracy per epoch', fontsize=80)
# plt.title('[CRL] Accuracy per epoch', fontsize=80)
plt.xlabel('Epoch', fontsize=70)
plt.ylabel('Accuracy', fontsize=70)
plt.ylim([0, 100])
plt.setp(ax.get_xticklabels(), fontsize=30)
plt.setp(ax.get_yticklabels(), fontsize=30)
plt.savefig('{}/{}_{}_acc_lr_{}.png'.format(file_name, args.model, args.b,
args.epochs))
# plt.savefig('./baseline_graph/150_250/128/500/{}_{}_acc_lr_{}.png'.format(args.model, args.b, args.epochs))
# plt.savefig('./CRL_graph/150_250/{}_{}_acc_lr_{}.png'.format(args.model, args.b, args.epochs))
df3 = pd.DataFrame()
df3['epoch'] = [i for i in range(1, args.epochs + 1)]
df3['test_nll'] = test_nll_report
df3['train_nll'] = train_nll_report
fig_acc = plt.figure(figsize=(35, 35))
fig_acc.set_facecolor('white')
ax = fig_acc.add_subplot()
ax.plot(df3['epoch'],
df3['test_nll'],
df3['epoch'],
df3['train_nll'],
linewidth=10)
ax.legend(['Test', 'Train'], loc=2, prop={'size': 60})
plt.title('[FL] NLL per epoch', fontsize=80)
# plt.title('[BASE] NLL per epoch', fontsize=80)
# plt.title('[CRL] NLL per epoch', fontsize=80)
plt.xlabel('Epoch', fontsize=70)
plt.ylabel('NLL', fontsize=70)
plt.ylim([0, 45])
plt.setp(ax.get_xticklabels(), fontsize=30)
plt.setp(ax.get_yticklabels(), fontsize=30)
plt.savefig('{}/{}_{}_nll_lr_{}.png'.format(file_name, args.model, args.b,
args.epochs))
# plt.savefig('./baseline_graph/150_250/128/500/{}_{}_nll_lr_{}.png'.format(args.model, args.b, args.epochs))
# plt.savefig('./CRL_graph/150_250/{}_{}_nll_lr_{}.png'.format(args.model, args.b, args.epochs))
df4 = pd.DataFrame()
df4['epoch'] = [i for i in range(1, args.epochs + 1)]
df4['test_over_con99'] = test_over_con99_report
df4['train_over_con99'] = train_over_con99_report
fig_acc = plt.figure(figsize=(35, 35))
fig_acc.set_facecolor('white')
ax = fig_acc.add_subplot()
ax.plot(df4['epoch'],
df4['test_over_con99'],
df4['epoch'],
df4['train_over_con99'],
linewidth=10)
ax.legend(['Test', 'Train'], loc=2, prop={'size': 60})
plt.title('[FL] Over conf99 per epoch', fontsize=80)
# plt.title('[BASE] Over conf99 per epoch', fontsize=80)
# plt.title('[CRL] Over conf99 per epoch', fontsize=80)
plt.xlabel('Epoch', fontsize=70)
plt.ylabel('Over con99', fontsize=70)
if args.data == 'cifar10' or args.data == 'cifar100':
plt.ylim([0, 50000])
else:
plt.ylim([0, 73257])
plt.setp(ax.get_xticklabels(), fontsize=30)
plt.setp(ax.get_yticklabels(), fontsize=30)
plt.savefig('{}/{}_{}_over_conf99_lr_{}.png'.format(
file_name, args.model, args.b, args.epochs))
# plt.savefig('./baseline_graph/150_250/128/500/{}_{}_over_conf99_lr_{}.png'.format(args.model, args.b, args.epochs))
# plt.savefig('./CRL_graph/150_250/{}_{}_over_conf99_lr_{}.png'.format(args.model, args.b, args.epochs))
df5 = pd.DataFrame()
df5['epoch'] = [i for i in range(1, args.epochs + 1)]
df5['test_e99'] = test_e99_report
df5['train_e99'] = train_e99_report
fig_acc = plt.figure(figsize=(35, 35))
fig_acc.set_facecolor('white')
ax = fig_acc.add_subplot()
ax.plot(df5['epoch'],
df5['test_e99'],
df5['epoch'],
df5['train_e99'],
linewidth=10)
ax.legend(['Test', 'Train'], loc=2, prop={'size': 60})
plt.title('[FL] E99 per epoch', fontsize=80)
# plt.title('[BASE] E99 per epoch', fontsize=80)
# plt.title('[CRL] E99 per epoch', fontsize=80)
plt.xlabel('Epoch', fontsize=70)
plt.ylabel('E99', fontsize=70)
plt.ylim([0, 0.2])
plt.setp(ax.get_xticklabels(), fontsize=30)
plt.setp(ax.get_yticklabels(), fontsize=30)
plt.savefig('{}/{}_{}_E99_flood_lr_{}.png'.format(file_name, args.model,
args.b, args.epochs))
# plt.savefig('./baseline_graph/150_250/128/500/{}_{}_E99_flood_lr_{}.png'.format(args.model, args.b, args.epochs))
# plt.savefig('./CRL_graph/150_250/{}_{}_E99_flood_lr_{}.png'.format(args.model, args.b, args.epochs))
df5 = pd.DataFrame()
df5['epoch'] = [i for i in range(1, args.epochs + 1)]
df5['test_cls_loss'] = test_cls_loss_report
df5['train_cls_loss'] = train_cls_loss_report
fig_acc = plt.figure(figsize=(35, 35))
fig_acc.set_facecolor('white')
ax = fig_acc.add_subplot()
ax.plot(df5['epoch'],
df5['test_cls_loss'],
df5['epoch'],
df5['train_cls_loss'],
linewidth=10)
ax.legend(['Test', 'Train'], loc=2, prop={'size': 60})
plt.title('[FL] CLS_loss per epoch', fontsize=80)
# plt.title('[BASE] CLS_loss per epoch', fontsize=80)
# plt.title('[CRL] CLS_loss per epoch', fontsize=80)
plt.xlabel('Epoch', fontsize=70)
plt.ylabel('Loss', fontsize=70)
plt.ylim([0, 5])
plt.setp(ax.get_xticklabels(), fontsize=30)
plt.setp(ax.get_yticklabels(), fontsize=30)
plt.savefig('{}/{}_{}_cls_loss_flood_lr_{}.png'.format(
file_name, args.model, args.b, args.epochs))
# plt.savefig('./baseline_graph/150_250/128/500/{}_{}_cls_loss_flood_lr_{}.png'.format(args.model, args.b, args.epochs))
# plt.savefig('./CRL_graph/150_250/{}_{}_cls_loss_flood_lr_{}.png'.format(args.model, args.b, args.epochs))
if args.rank_weight != 0.0:
df6 = pd.DataFrame()
df6['epoch'] = [i for i in range(1, args.epochs + 1)]
df6['train_cls_loss'] = train_cls_loss_report
df6['train_rank_loss'] = train_rank_loss_report
df6['train_total_loss'] = train_total_loss_report
fig_acc = plt.figure(figsize=(35, 35))
fig_acc.set_facecolor('white')
ax = fig_acc.add_subplot()
ax.plot(df6['epoch'],
df6['train_cls_loss'],
df6['epoch'],
df6['train_rank_loss'],
df6['epoch'],
df6['train_total_loss'],
linewidth=10)
ax.legend(['CLS', 'Rank', 'Total'], loc=2, prop={'size': 60})
plt.title('[FL] CLS_loss per epoch', fontsize=80)
plt.xlabel('Epoch', fontsize=70)
plt.ylabel('Loss', fontsize=70)
# plt.ylim([0, 5])
plt.setp(ax.get_xticklabels(), fontsize=30)
plt.setp(ax.get_yticklabels(), fontsize=30)
plt.savefig(
'./CRL_graph/150_250/{}_{}_cls_loss_flood_lr_{}.png'.format(
args.model, args.b, args.epochs))
test_acc_report.insert(0, 'ACC')
test_ece_report.insert(0, 'ECE')
test_nll_report.insert(0, 'NLL')
test_over_con99_report.insert(0, 'Over_conf99')
test_e99_report.insert(0, 'E99')
test_cls_loss_report.insert(0, 'CLS')
wr_test.writerow(test_acc_report)
wr_test.writerow(test_ece_report)
wr_test.writerow(test_nll_report)
wr_test.writerow(test_over_con99_report)
wr_test.writerow(test_e99_report)
wr_test.writerow(test_cls_loss_report)
train_acc_report.insert(0, 'ACC')
train_ece_report.insert(0, 'ECE')
train_nll_report.insert(0, 'NLL')
train_over_con99_report.insert(0, 'Over_conf99')
train_e99_report.insert(0, 'E99')
train_cls_loss_report.insert(0, 'CLS')
wr_train.writerow(train_acc_report)
wr_train.writerow(train_ece_report)
wr_train.writerow(train_nll_report)
wr_train.writerow(train_over_con99_report)
wr_train.writerow(train_e99_report)
wr_train.writerow(train_cls_loss_report)
if args.rank_weight != 0.0:
train_rank_loss_report.insert(0, 'Rank')
train_total_loss_report.insert(0, 'Total')
wr_train.writerow(train_rank_loss_report)
wr_train.writerow(train_total_loss_report)
#'''
# result write
result_logger.write([
acc_te, aurc_te * 1000, eaurc_te * 1000, aupr_te * 100, fpr_te * 100,
ece_te * 100, nll_te * 10, brier_te * 100, E99_te * 100
])
if args.valid == True:
print("Best Valid Acc : {}".format(acc))
print("Flood Level: {}".format(args.b))
print("Sort : {}".format(args.sort))
print("Sort Mode : {}".format(args.sort_mode))
print("TIME : ", time.time() - start)
