def get_lr_scheduler(optimizer, config):
name = config.type
args = copy_and_delete(config.toDict(), 'type')
args = copy_and_delete(args, 'scheme')
lr_scheduler = get_if_implemented(torch.optim.lr_scheduler, name)
if lr_scheduler is None:
try:
lr_scheduler = get_module('./schedulers', name)
except:
pass
if lr_scheduler is None:
fcn = get_function('schedulers.functional', name)
assert fcn, "No functional implementation of {} was found".format(name)
fcn_wrapper = fcn(**args)
lr_scheduler = torch.optim.lr_scheduler.LambdaLR(
optimizer, [fcn_wrapper])
else:
lr_scheduler = lr_scheduler(optimizer, **args)
return lr_scheduler
def train(config):
## set pre-process
prep_dict = {}
prep_config = config["prep"]
prep_dict["source"] = prep.image_train(**config["prep"]['params'])
prep_dict["target"] = prep.image_train(**config["prep"]['params'])
if prep_config["test_10crop"]:
prep_dict["test"] = prep.image_test_10crop(**config["prep"]['params'])
else:
prep_dict["test"] = prep.image_test(**config["prep"]['params'])
## prepare data
dsets = {}
dset_loaders = {}
data_config = config["data"]
train_bs = data_config["source"]["batch_size"]
test_bs = data_config["test"]["batch_size"]
dsets["source"] = ImageList(open(data_config["source"]["list_path"]).readlines(), \
transform=prep_dict["source"])
dset_loaders["source"] = DataLoader(dsets["source"], batch_size=train_bs, \
shuffle=True, num_workers=4, drop_last=True)
dsets["target"] = ImageList(open(data_config["target"]["list_path"]).readlines(), \
transform=prep_dict["target"])
dset_loaders["target"] = DataLoader(dsets["target"], batch_size=train_bs, \
shuffle=True, num_workers=4, drop_last=True)
if prep_config["test_10crop"]:
for i in range(10):
dsets["test"] = [ImageList(open(data_config["test"]["list_path"]).readlines(), \
transform=prep_dict["test"][i]) for i in range(10)]
dset_loaders["test"] = [DataLoader(dset, batch_size=test_bs, \
shuffle=False, num_workers=4) for dset in dsets['test']]
else:
dsets["test"] = ImageList(open(data_config["test"]["list_path"]).readlines(), \
transform=prep_dict["test"])
dset_loaders["test"] = DataLoader(dsets["test"], batch_size=test_bs, \
shuffle=False, num_workers=4)
class_num = config["network"]["params"]["class_num"]
## set base network
net_config = config["network"]
base_network = net_config["name"](**net_config["params"])
base_network = base_network.cuda() #加载基础网络结构
parameter_list = base_network.get_parameters()
## set optimizer
optimizer_config = config["optimizer"]
optimizer = optimizer_config["type"](parameter_list, \
**(optimizer_config["optim_params"]))
crit = LabelSmoothingLoss(smoothing=0.1, classes=31) #标签平滑操作
param_lr = []
for param_group in optimizer.param_groups:
param_lr.append(param_group["lr"])
schedule_param = optimizer_config["lr_param"]
lr_scheduler = lr_schedule.schedule_dict[optimizer_config["lr_type"]]
gpus = config['gpu'].split(',')
if len(gpus) > 1:
base_network = nn.DataParallel(base_network,
device_ids=[int(i) for i in gpus])
## train
len_train_source = len(dset_loaders["source"])
len_train_target = len(dset_loaders["target"])
best_acc = 0.0
start_time = time.time()
for i in range(config["num_iterations"]):
if i % config["test_interval"] == config["test_interval"] - 1:
# 在这里进行测试的工作
base_network.train(False)
temp_acc = image_classification_test(dset_loaders,
base_network,
test_10crop=False)
temp_model = nn.Sequential(base_network)
if temp_acc > best_acc:
best_acc = temp_acc
best_model = temp_model
log_str = "iter: {:05d}, precision: {:.5f}".format(i, temp_acc)
config["out_file"].write(log_str + "\n")
config["out_file"].flush()
print(log_str)
end_time = time.time()
print('iter {} cost time {:.4f} sec.'.format(
i, end_time - start_time)) # 打印时间间隔
start_time = time.time()
if i % config["snapshot_interval"] == 0:
torch.save(nn.Sequential(base_network), osp.join(config["output_path"], \
"iter_{:05d}_model.pth.tar".format(i)))
## train one iter
base_network.train(True) # 训练模式
optimizer = lr_scheduler(optimizer, i, **schedule_param)
optimizer.zero_grad()
if i % len_train_source == 0:
iter_source = iter(dset_loaders["source"])
if i % len_train_target == 0:
iter_target = iter(dset_loaders["target"])
inputs_source, labels_source = iter_source.next()
inputs_target, labels_target = iter_target.next()
inputs_source, inputs_target, labels_source = inputs_source.cuda(
), inputs_target.cuda(), labels_source.cuda()
features_source, outputs_source = base_network(inputs_source)
features_target, outputs_target = base_network(inputs_target)
#目标域的熵正则化操作
t_logit = outputs_target
t_prob = F.softmax(t_logit)
t_entropy_loss = get_entropy_loss(t_prob) # 计算目标域的熵的损失
entropy_loss = config['ENT_w'] * (t_entropy_loss)
# classifier_loss = nn.CrossEntropyLoss()(outputs_source, labels_source) # 源域的分类损失
classifier_loss = crit(outputs_source, labels_source)
total_loss = classifier_loss + entropy_loss
if i % config["test_interval"] == config["test_interval"] - 1:
print('total loss: {:.4f}, classifier loss: {:.4f}'.format(
total_loss.item(), classifier_loss.item()))
total_loss.backward()
optimizer.step()
if (i + 1) % 1000 == 0:
model_path = os.path.join(
os.path.dirname(config['save_model_name']), 'temp_model.pth')
save_model(base_network, model_path)
model_path = config['save_model_name']
best_model_path = os.path.join(os.path.dirname(model_path),
'best_model.pth')
save_model(base_network, model_path)
save_model(best_model, best_model_path)
torch.save(best_model, osp.join(config["output_path"],
"best_model.pth.tar"))
return best_acc
def train(config):
## set pre-process
prep_dict = {}
prep_config = config["prep"]
prep_dict["source"] = prep.image_train(**config["prep"]['params'])
prep_dict["target"] = prep.image_train(**config["prep"]['params'])
if prep_config["test_10crop"]:
prep_dict["test"] = prep.image_test_10crop(**config["prep"]['params'])
else:
prep_dict["test"] = prep.image_test(**config["prep"]['params'])
## prepare data
dsets = {}
dset_loaders = {}
data_config = config["data"]
train_bs = data_config["source"]["batch_size"]
test_bs = data_config["test"]["batch_size"]
dsets["source"] = ImageList(open(data_config["source"]["list_path"]).readlines(), \
transform=prep_dict["source"])
dset_loaders["source"] = DataLoader(dsets["source"], batch_size=train_bs, \
shuffle=True, num_workers=4, drop_last=True)
dsets["target"] = ImageList(open(data_config["target"]["list_path"]).readlines(), \
transform=prep_dict["target"])
dset_loaders["target"] = DataLoader(dsets["target"], batch_size=train_bs, \
shuffle=True, num_workers=4, drop_last=True)
if prep_config["test_10crop"]:
for i in range(10):
dsets["test"] = [ImageList(open(data_config["test"]["list_path"]).readlines(), \
transform=prep_dict["test"][i]) for i in range(10)]
dset_loaders["test"] = [DataLoader(dset, batch_size=test_bs, \
shuffle=False, num_workers=4) for dset in dsets['test']]
else:
dsets["test"] = ImageList(open(data_config["test"]["list_path"]).readlines(), \
transform=prep_dict["test"])
dset_loaders["test"] = DataLoader(dsets["test"], batch_size=test_bs, \
shuffle=False, num_workers=4)
class_num = config["network"]["params"]["class_num"]
crit = LabelSmoothingLoss(smoothing=0.05, classes=class_num)#标签平滑操作
## set base network
net_config = config["network"]
base_network = net_config["name"](**net_config["params"])
base_network = base_network.cuda() # 加载基础网络结构
## add additional network for some methods
if config["loss"]["random"]:
random_layer = network.RandomLayer([base_network.output_num(), class_num], config["loss"]["random_dim"])
ad_net = network.AdversarialNetwork(config["loss"]["random_dim"], 1024)
else:
random_layer = None
ad_net = network.AdversarialNetwork(base_network.output_num() * class_num, 1024) # 对抗网络结构
if config["loss"]["random"]:
random_layer.cuda()
ad_net = ad_net.cuda()
parameter_list = base_network.get_parameters() + ad_net.get_parameters()
## set optimizer
optimizer_config = config["optimizer"]
optimizer = optimizer_config["type"](parameter_list, \
**(optimizer_config["optim_params"]))
#中心损失函数
criterion_centor=CenterLoss(num_classes=class_num,feat_dim=256,use_gpu=True)
optimizer_centerloss=torch.optim.SGD(criterion_centor.parameters(),lr=config['lr'])
param_lr = []
for param_group in optimizer.param_groups:
param_lr.append(param_group["lr"])
schedule_param = optimizer_config["lr_param"]
lr_scheduler = lr_schedule.schedule_dict[optimizer_config["lr_type"]]
gpus = config['gpu'].split(',')
if len(gpus) > 1:
ad_net = nn.DataParallel(ad_net, device_ids=[int(i) for i in gpus])
base_network = nn.DataParallel(base_network, device_ids=[int(i) for i in gpus])
## train
len_train_source = len(dset_loaders["source"])
len_train_target = len(dset_loaders["target"])
transfer_loss_value = classifier_loss_value = total_loss_value = 0.0
best_acc = 0.0
start_time = time.time()
for i in range(config["num_iterations"]):
if i % config["test_interval"] == config["test_interval"] - 1:
# 在这里进行测试的工作
base_network.train(False)
temp_acc = image_classification_test(dset_loaders, \
base_network, test_10crop=prep_config["test_10crop"])
temp_model = nn.Sequential(base_network)
if temp_acc > best_acc:
best_acc = temp_acc
best_model = temp_model
log_str = "iter: {:05d}, precision: {:.5f}".format(i, temp_acc)
config["out_file"].write(log_str + "\n")
config["out_file"].flush()
print(log_str)
end_time = time.time()
print('iter {} cost time {:.4f} sec.'.format(i, end_time - start_time)) # 打印时间间隔
start_time = time.time()
if i % config["snapshot_interval"] == 0:
torch.save(nn.Sequential(base_network), osp.join(config["output_path"], \
"iter_{:05d}_model.pth.tar".format(i)))
loss_params = config["loss"]
## train one iter
base_network.train(True) # 训练模式
ad_net.train(True)
optimizer = lr_scheduler(optimizer, i, **schedule_param)
# optimizer_centerloss=lr_scheduler(optimizer_centerloss, i, **schedule_param)
optimizer.zero_grad()
optimizer_centerloss.zero_grad()
if i % len_train_source == 0:
iter_source = iter(dset_loaders["source"])
if i % len_train_target == 0:
iter_target = iter(dset_loaders["target"])
inputs_source, labels_source = iter_source.next()
inputs_target, labels_target = iter_target.next()
inputs_source, inputs_target, labels_source = inputs_source.cuda(), inputs_target.cuda(), labels_source.cuda()
features_source, outputs_source = base_network(inputs_source)
features_target, outputs_target = base_network(inputs_target)
features = torch.cat((features_source, features_target), dim=0)
outputs = torch.cat((outputs_source, outputs_target), dim=0)
softmax_out = nn.Softmax(dim=1)(outputs)
if config['method'] == 'CDAN+E':
entropy = loss.Entropy(softmax_out)
transfer_loss = loss.CDAN([features, softmax_out], ad_net, entropy, network.calc_coeff(i), random_layer)
elif config['method'] == 'CDAN':
transfer_loss = loss.CDAN([features, softmax_out], ad_net, None, None, random_layer)
elif config['method'] == 'DANN':
transfer_loss = loss.DANN(features, ad_net)
else:
raise ValueError('Method cannot be recognized.')
# classifier_loss = nn.CrossEntropyLoss()(outputs_source, labels_source) # 源域的分类损失
classifier_loss = crit(outputs_source, labels_source) # 源域的分类损失，标签平滑操作
# 目标域的熵正则化操作
outputs_target=outputs[len(inputs_source):,:]
t_logit = outputs_target
t_prob = torch.softmax(t_logit,dim=1)
t_entropy_loss = get_entropy_loss(t_prob) # 计算目标域的熵的损失
entropy_loss = 0.05 * (t_entropy_loss)
# 计算中心损失函数
loss_centor = criterion_centor(features_source, labels_source) # 中心损失计算
total_loss = loss_params["trade_off"] * transfer_loss + classifier_loss + config['centor_w']*loss_centor
if i % config["test_interval"] == config["test_interval"] - 1:
print('total loss: {:.4f}, transfer loss: {:.4f}, classifier loss: {:.4f}, centor loss: {:.4f}'.format(
total_loss.item(),transfer_loss.item(),classifier_loss.item(),config['centor_w']*loss_centor.item()
))
total_loss.backward()
optimizer.step()
# by doing so, weight_cent would not impact on the learning of centers
for param in criterion_centor.parameters():
param.grad.data *= (1. / config['centor_w'])
optimizer_centerloss.step()
torch.save(best_model, osp.join(config["output_path"], "best_model.pth.tar"))
return best_acc
def main():
config = vars(parse_args())
now = datetime.datetime.now()
if config["name"] is None:
if config["deep_supervision"]:
config["name"] = "%s_%s_wDS_%s" % (
config["dataset"],
config["arch"],
now.strftime("%Y%m%d_%H%M%S"),
)
else:
config["name"] = "%s_%s_woDS_%s" % (
config["dataset"],
config["arch"],
now.strftime("%Y%m%d_%H%M%S"),
)
output_path = os.path.join(cfg.UNET_RESULTS_DIR, config["name"])
try:
os.makedirs(output_path, exist_ok=True)
except Exception as e:
print(e)
models_path = os.path.join(output_path, "models")
os.mkdir(models_path)
with open(os.path.join(models_path, "config.yml"), "w") as f:
yaml.dump(config, f)
print("-" * 20)
for key in config:
print("%s: %s" % (key, config[key]))
print("-" * 20)
# Tensorboad 用のログを記録するディレクトリパス
log_dir = os.path.join(output_path, "log")
os.mkdir(log_dir)
writer = SummaryWriter(log_dir=log_dir)
# define loss function(criterion)
if config["loss"] == "BCEWithLogitsLoss":
criterion = nn.BCEWithLogitsLoss().cuda()
else:
criterion = losses.__dict__[config["loss"]]().cuda()
cudnn.benchmark = True
# create model
print("=> creating model %s" % config["arch"])
model = archs.__dict__[config["arch"]](config["num_classes"],
config["input_channels"],
config["deep_supervision"])
model = model.cuda()
# モデルを TensorBorad で表示するため，ログに保存
# image = torch.randn(1, 3, 2224, 224)
# writer.add_graph(model, image)
params = filter(lambda p: p.requires_grad, model.parameters())
if config["optimizer"] == "Adam":
optimizer = optim.Adam(params,
lr=config["lr"],
weight_decay=config["weight_decay"])
elif config["optimizer"] == "SGD":
optimizer = optim.SGD(
params,
lr=config["lr"],
momentum=config["momentum"],
nesterov=config["nesterov"],
weight_decay=config["weight_decay"],
)
else:
raise NotImplementedError
# scheduler
if config["scheduler"] == "CosineAnnealingLR":
scheduler = lr_scheduler.CosineAnnealingLR(optimizer=optimizer,
T_max=config["epochs"],
eta_min=config["min_lr"])
elif config["scheduler"] == "ReduceLROnPlateau":
scheduler = lr_scheduler(
optimizer=optimizer,
factor=config["factor"],
patience=config["patience"],
verbose=1,
min_lr=config["min_lr"],
)
elif config["scheduler"] == "MultiStepLR":
scheduler = lr_scheduler.MultiStepLR(
optimizer=optimizer,
milestones=[int(e) for e in config["milestones"].split(",")],
gamma=config["gamma"],
)
elif config["scheduler"] == "ConstantLR":
scheduler = None
else:
raise NotImplementedError
# Data loading code
if config["dataset"] == "dsb2018_96":
input_dir = cfg.DSB2018_96_DIR
img_ids = glob(
os.path.join(input_dir, "images", "*" + config["img_ext"]))
img_ids = [os.path.splitext(os.path.basename(p))[0] for p in img_ids]
train_img_ids, val_img_ids = train_test_split(img_ids,
test_size=0.2,
random_state=41)
train_transform = Compose([
transforms.RandomRotate90(),
transforms.Flip(),
OneOf(
[
transforms.HueSaturationValue(),
transforms.RandomBrightness(),
transforms.RandomContrast(),
],
p=1,
),
transforms.Resize(config["input_h"], config["input_w"]),
transforms.Normalize(),
])
val_transform = Compose([
transforms.Resize(config["input_h"], config["input_w"]),
transforms.Normalize(),
])
train_dataset = Dataset(
img_ids=train_img_ids,
img_dir=os.path.join(input_dir, "images"),
mask_dir=os.path.join(input_dir, "masks"),
img_ext=config["img_ext"],
mask_ext=config["mask_ext"],
num_classes=config["num_classes"],
transform=train_transform,
)
val_dataset = Dataset(
img_ids=val_img_ids,
img_dir=os.path.join(input_dir, "images"),
mask_dir=os.path.join(input_dir, "masks"),
img_ext=config["img_ext"],
mask_ext=config["mask_ext"],
num_classes=config["num_classes"],
transform=val_transform,
)
train_loader = torch.utils.data.DataLoader(
train_dataset,
batch_size=config["batch_size"],
shuffle=True,
num_workers=config["num_workers"],
drop_last=True,
)
val_loader = torch.utils.data.DataLoader(
val_dataset,
batch_size=config["batch_size"],
shuffle=False,
num_workers=config["num_workers"],
drop_last=False,
)
log = OrderedDict([
("epoch", []),
("lr", []),
("loss", []),
("iou", []),
("val_loss", []),
("val_iou", []),
])
best_iou = 0
trigger = 0
for epoch in range(config["epochs"] + 1):
print("Epoch [%d/%d]" % (epoch, config["epochs"]))
# train for one epoch
train_log = train(config, train_loader, model, criterion, optimizer)
# evaluate on validation set
val_log = validate(config, val_loader, model, criterion)
if config["scheduler"] == "CosineAnnealingLR":
scheduler.step()
elif config["scheduler"] == "ReduceLROnPlateau":
scheduler.step(val_log["loss"])
print("loss %.4f - iou %.4f - val_loss %.4f - val_iou %.4f" %
(train_log["loss"], train_log["iou"], val_log["loss"],
val_log["iou"]))
log["epoch"].append(epoch)
log["lr"].append(config["lr"])
log["loss"].append(train_log["loss"])
log["iou"].append(train_log["iou"])
log["val_loss"].append(val_log["loss"])
log["val_iou"].append(val_log["iou"])
# Tensorboard用のデータ
writer.add_scalar("training loss", train_log["loss"], epoch)
writer.add_scalar("validation loss", val_log["loss"], epoch)
pd.DataFrame(log).to_csv("%s/log.csv" % (log_dir), index=False)
if epoch == 0:
best_loss = val_log["loss"]
trigger += 1
# Best Model Save
# if val_log['iou'] > best_iou:
if (val_log["iou"] > best_iou) and (val_log["loss"] <= best_loss):
torch.save(model.state_dict(), "%s/model.pth" % (models_path))
best_iou = val_log["iou"]
best_loss = val_log["loss"]
print("=> saved best model")
trigger = 0
# early stopping
if (config["early_stopping"] >= 0 and
trigger >= config["early_stopping"]) or val_log["loss"] < 1e-4:
print("=> early stopping")
break
torch.cuda.empty_cache()
# summary writer を必要としない場合，close()メソッドを呼び出す
writer.close()