def test_train_gdl(self):
dataset = SegmentationDummyDataset(batch_size=32)
pspnet = PSPNet(epochs=1,
dataset=dataset,
frontend_name='resnet101',
generarized_dice_loss={'alpha': 0.01})
history = pspnet.train()
ok_('loss' in history)
def main():
global args, logger
args = get_parser()
check(args)
logger = get_logger()
os.environ["CUDA_VISIBLE_DEVICES"] = ','.join(str(x) for x in args.test_gpu)
logger.info(args)
logger.info("=> creating model ...")
logger.info("Classes: {}".format(args.classes))
value_scale = 255
mean = [0.485, 0.456, 0.406]
mean = [item * value_scale for item in mean]
std = [0.229, 0.224, 0.225]
std = [item * value_scale for item in std]
gray_folder = os.path.join(args.save_folder, 'gray')
color_folder = os.path.join(args.save_folder, 'color')
test_transform = transform.Compose([transform.ToTensor()])
test_data = dataset.SemData(split=args.split, data_root=args.data_root, data_list=args.test_list, transform=test_transform)
index_start = args.index_start
if args.index_step == 0:
index_end = len(test_data.data_list)
else:
index_end = min(index_start + args.index_step, len(test_data.data_list))
test_data.data_list = test_data.data_list[index_start:index_end]
test_loader = torch.utils.data.DataLoader(test_data, batch_size=1, shuffle=False, num_workers=args.workers, pin_memory=True)
colors = np.loadtxt(args.colors_path).astype('uint8')
names = [line.rstrip('\n') for line in open(args.names_path)]
if not args.has_prediction:
if args.arch == 'psp':
from model.pspnet import PSPNet
model = PSPNet(layers=args.layers, classes=args.classes, zoom_factor=args.zoom_factor, pretrained=False)
elif args.arch == 'psa':
from model.psanet import PSANet
model = PSANet(layers=args.layers, classes=args.classes, zoom_factor=args.zoom_factor, compact=args.compact,
shrink_factor=args.shrink_factor, mask_h=args.mask_h, mask_w=args.mask_w,
normalization_factor=args.normalization_factor, psa_softmax=args.psa_softmax, pretrained=False)
logger.info(model)
model = torch.nn.DataParallel(model).cuda()
cudnn.benchmark = True
if os.path.isfile(args.model_path):
logger.info("=> loading checkpoint '{}'".format(args.model_path))
checkpoint = torch.load(args.model_path)
model.load_state_dict(checkpoint['state_dict'], strict=False)
logger.info("=> loaded checkpoint '{}'".format(args.model_path))
else:
raise RuntimeError("=> no checkpoint found at '{}'".format(args.model_path))
test(test_loader, test_data.data_list, model, args.classes, mean, std, args.base_size, args.test_h, args.test_w, args.scales, gray_folder, color_folder, colors)
if args.split != 'test':
cal_acc(test_data.data_list, gray_folder, args.classes, names)
def __setupSegmentation(self):
if torch.cuda.device_count() == 0:
device = torch.device("cpu")
else:
device = torch.device("cuda")
model = PSPNet(layers=self.args.layers, classes=self.args.classes, zoom_factor=self.args.zoom_factor, pretrained=False)
model = torch.nn.DataParallel(model)
cudnn.benchmark = True
if os.path.isfile(self.args.model_path):
checkpoint = torch.load(self.args.model_path, map_location=torch.device('cpu'))
model.load_state_dict(checkpoint['state_dict'], strict=False)
return model
else:
raise RuntimeError("=> no checkpoint found at '{}'".format(self.args.model_path))
def __init__(self, tag='teacher', layers=50, bins=(1, 2, 3, 6), dropout=0.1, classes=19, zoom_factor=8, temperature = 1, alpha = 0., use_ppm=True, criterion=nn.CrossEntropyLoss(ignore_index=255), pretrained=True):
super(KDNet, self).__init__()
assert layers in [18, 50]
assert 2048 % len(bins) == 0
assert classes > 1
assert zoom_factor in [1, 2, 4, 8]
self.zoom_factor = zoom_factor
self.use_ppm = use_ppm
self.criterion = criterion
self.temperature = temperature
self.alpha = alpha
self.student_net = PSPNet(tag='student', layers=18, bins=(1, 2, 3, 6), dropout=0.1, classes=19, zoom_factor=8, use_ppm=True, criterion=nn.CrossEntropyLoss(ignore_index=255), pretrained=True)
self.teacher_loader = teacher_loader()
def __init__(self, config_file=CONFIG_FILE):
# Load Parameters
self.args_ = config.load_cfg_from_cfg_file(config_file)
self.logger_ = get_logger()
os.environ["CUDA_VISIBLE_DEVICES"] = ','.join(
str(x) for x in self.args_.test_gpu)
value_scale = 255
mean = [0.485, 0.456, 0.406]
self.mean_ = [item * value_scale for item in mean]
std = [0.229, 0.224, 0.225]
self.std_ = [item * value_scale for item in std]
# self.colors_ = np.loadtxt(self.args_.colors_path).astype('uint8')
# Load Model
if self.args_.arch == 'psp':
from model.pspnet import PSPNet
self.model_ = PSPNet(layers=self.args_.layers,
classes=self.args_.classes,
zoom_factor=self.args_.zoom_factor,
pretrained=False)
elif self.args_.arch == 'psa':
from model.psanet import PSANet
self.model_ = PSANet(
layers=self.args_.layers,
classes=self.args_.classes,
zoom_factor=self.args_.zoom_factor,
compact=self.args_.compact,
shrink_factor=self.args_.shrink_factor,
mask_h=self.args_.mask_h,
mask_w=self.args_.mask_w,
normalization_factor=self.args_.normalization_factor,
psa_softmax=self.args_.psa_softmax,
pretrained=False)
self.model_ = torch.nn.DataParallel(self.model_).cuda()
cudnn.benchmark = True
if os.path.isfile(self.args_.model_path):
self.logger_ = get_logger().info(
"=> loading checkpoint '{}'".format(self.args_.model_path))
checkpoint = torch.load(self.args_.model_path)
self.model_.load_state_dict(checkpoint['state_dict'], strict=False)
self.logger_ = get_logger().info(
"=> loaded checkpoint '{}'".format(self.args_.model_path))
else:
raise RuntimeError("=> no checkpoint found at '{}'".format(
self.args_.model_path))
def main():
global args, logger
args = get_parser()
check(args)
logger = get_logger()
os.environ["CUDA_VISIBLE_DEVICES"] = ','.join(str(x) for x in args.test_gpu)
logger.info(args)
logger.info("=> creating model ...")
logger.info("Classes: {}".format(args.classes))
value_scale = 255
mean = [0.485, 0.456, 0.406]
mean = [item * value_scale for item in mean]
std = [0.229, 0.224, 0.225]
std = [item * value_scale for item in std]
colors = np.loadtxt(args.colors_path).astype('uint8')
if args.arch == 'psp':
from model.pspnet import PSPNet
model = PSPNet(layers=args.layers, classes=args.classes, zoom_factor=args.zoom_factor, pretrained=False)
elif args.arch == 'psa':
from model.psanet import PSANet
model = PSANet(layers=args.layers, classes=args.classes, zoom_factor=args.zoom_factor, compact=args.compact,
shrink_factor=args.shrink_factor, mask_h=args.mask_h, mask_w=args.mask_w,
normalization_factor=args.normalization_factor, psa_softmax=args.psa_softmax, pretrained=False)
logger.info(model)
model = torch.nn.DataParallel(model).cuda()
cudnn.benchmark = False #True
if os.path.isfile(args.model_path):
logger.info("=> loading checkpoint '{}'".format(args.model_path))
checkpoint = torch.load(args.model_path)
model.load_state_dict(checkpoint['state_dict'], strict=False)
logger.info("=> loaded checkpoint '{}'".format(args.model_path))
else:
raise RuntimeError("=> no checkpoint found at '{}'".format(args.model_path))
with open(args.image) as f:
image_files = f.read().splitlines()
for file in image_files:
image = file.split()
image = os.path.join('/scratch/mw3706/dim/Deep_Image_Matting_Reproduce/pspnet/data/portrait/', image[0])
test(model.eval(), image, args.classes, mean, std, args.base_size, args.test_h, args.test_w, args.scales, colors)
if (args.image).split('/')[-1] == 'training.txt'
train_label_list = os.listdir('/scratch/mw3706/dim/Deep_Image_Matting_Reproduce/pspnet/data/portrait/label/train_label')
with open('/scratch/mw3706/dim/Deep_Image_Matting_Reproduce/pspnet/data/portrait/label/training.txt', 'w') as f:
for label in train_label_list:
f.write('/scratch/mw3706/dim/Deep_Image_Matting_Reproduce/pspnet/data/portrait/label/train_label/'+label+'\n')
else:
val_label_list = os.listdir('/scratch/mw3706/dim/Deep_Image_Matting_Reproduce/pspnet/data/portrait/label/val_label/')
with open('/scratch/mw3706/dim/Deep_Image_Matting_Reproduce/pspnet/data/portrait/label/validation.txt', 'w') as f:
for label in val_label_list:
f.write('/scratch/mw3706/dim/Deep_Image_Matting_Reproduce/pspnet/data/portrait/label/val_label/'+label+'\n')
def main():
global args, logger
args = get_parser()
check(args)
logger = get_logger()
os.environ["CUDA_VISIBLE_DEVICES"] = ','.join(
str(x) for x in args.test_gpu)
logger.info(args)
logger.info("=> creating model ...")
logger.info("Classes: {}".format(args.classes))
value_scale = 255
mean = [0.485, 0.456, 0.406]
mean = [item * value_scale for item in mean]
std = [0.229, 0.224, 0.225]
std = [item * value_scale for item in std]
colors = np.loadtxt(args.colors_path).astype('uint8')
if args.arch == 'psp':
from model.pspnet import PSPNet
model = PSPNet(layers=args.layers,
classes=args.classes,
zoom_factor=args.zoom_factor,
pretrained=False)
elif args.arch == 'psa':
from model.psanet import PSANet
model = PSANet(layers=args.layers,
classes=args.classes,
zoom_factor=args.zoom_factor,
compact=args.compact,
shrink_factor=args.shrink_factor,
mask_h=args.mask_h,
mask_w=args.mask_w,
normalization_factor=args.normalization_factor,
psa_softmax=args.psa_softmax,
pretrained=False)
logger.info(model)
model = torch.nn.DataParallel(model).cuda()
cudnn.benchmark = True
if os.path.isfile(args.model_path):
logger.info("=> loading checkpoint '{}'".format(args.model_path))
checkpoint = torch.load(args.model_path)
model.load_state_dict(checkpoint['state_dict'], strict=False)
logger.info("=> loaded checkpoint '{}'".format(args.model_path))
else:
raise RuntimeError("=> no checkpoint found at '{}'".format(
args.model_path))
paths = glob.glob(args.image + '/scene*/color/*00.jpg')
for path in paths:
test(model.eval(), path, args.classes, mean, std, args.base_size,
args.test_h, args.test_w, args.scales, colors)
def get_model(args, criterion, BatchNorm):
"""
Args:
-
Returns:
-
"""
if args.arch == 'psp':
from model.pspnet import PSPNet
model = PSPNet(layers=args.layers, classes=args.classes, zoom_factor=args.zoom_factor, criterion=criterion, BatchNorm=BatchNorm, network_name=args.network_name)
elif args.arch == 'hrnet':
from model.seg_hrnet import get_configured_hrnet
# note apex batchnorm is hardcoded
model = get_configured_hrnet(args.classes)
elif args.arch == 'hrnet_ocr':
from model.seg_hrnet_ocr import get_configured_hrnet_ocr
model = get_configured_hrnet_ocr(args.classes)
return model
def main_worker(gpu, ngpus_per_node, argss):
global args
args = argss
## step.1 设置分布式相关参数
# 1.1 分布式初始化
if args.distributed:
if args.dist_url == "env://" and args.rank == -1:
args.rank = int(os.environ["RANK"])
if args.multiprocessing_distributed:
args.rank = args.rank * ngpus_per_node + gpu
dist.init_process_group(backend=args.dist_backend,
init_method=args.dist_url,
world_size=args.world_size,
rank=args.rank) # 分布式初始化
## step.2 构建网络
# ---------------------------------------------- 根据实际情况自己写 ---------------------------------------------#
criterion = nn.CrossEntropyLoss(
ignore_index=args.ignore_label) # 交叉熵损失函数, 根据情况自己修改
if args.arch == 'psp':
from model.pspnet import PSPNet
model = PSPNet(layers=args.layers,
classes=args.classes,
zoom_factor=args.zoom_factor,
criterion=criterion)
modules_ori = [
model.layer0, model.layer1, model.layer2, model.layer3,
model.layer4
]
modules_new = [model.ppm, model.cls, model.aux]
elif args.arch == 'psa':
from model.psanet import PSANet
model = PSANet(layers=args.layers,
classes=args.classes,
zoom_factor=args.zoom_factor,
psa_type=args.psa_type,
compact=args.compact,
shrink_factor=args.shrink_factor,
mask_h=args.mask_h,
mask_w=args.mask_w,
normalization_factor=args.normalization_factor,
psa_softmax=args.psa_softmax,
criterion=criterion)
modules_ori = [
model.layer0, model.layer1, model.layer2, model.layer3,
model.layer4
]
modules_new = [model.psa, model.cls, model.aux]
# ---------------------------------------------------- END ---------------------------------------------------#
## step.3 设置优化器
params_list = [] # 模型参数列表
for module in modules_ori:
params_list.append(dict(params=module.parameters(),
lr=args.base_lr)) # 原来backbone网络 学习率 0.01
for module in modules_new:
params_list.append(
dict(params=module.parameters(),
lr=args.base_lr * 10)) # 新加入预测网络 学习率 0.1
args.index_split = 5
optimizer = torch.optim.SGD(params_list,
lr=args.base_lr,
momentum=args.momentum,
weight_decay=args.weight_decay) # SGD优化器
# 3.x 设置sync_bn from torch.nn.SyncBatchNorm
if args.sync_bn:
model = nn.SyncBatchNorm.convert_sync_batchnorm(model)
## step.4 多线程分布式工作
# 4.1 判断是否是在主进程中, 如果在进行如下程序
if main_process():
global logger, writer
logger = get_logger() # 设置logger
writer = SummaryWriter(args.save_path) # 设置writer
logger.info(args) # 输出参数列表
logger.info("=> creating model ...")
logger.info("Classes: {}".format(args.classes))
logger.info(model) # 输出网络列表
# 4.2 分布式工作
if args.distributed:
torch.cuda.set_device(gpu) # 指定编号为gpu的那一张显卡
args.batch_size = int(args.batch_size /
ngpus_per_node) # 每张卡的训练的batch size
args.batch_size_val = int(args.batch_size_val /
ngpus_per_node) # 每张卡的评测的batch size
args.workers = int(
(args.workers + ngpus_per_node - 1) / ngpus_per_node) # 每张卡工作的数目
model = torch.nn.parallel.DistributedDataParallel(
model.cuda(), device_ids=[gpu]) # 加载torch分布式
else:
model = torch.nn.DataParallel(model.cuda()) # 数据并行
## step.5 加载网络权重
# 5.1 直接加载网络预权重
if args.weight:
if os.path.isfile(args.weight):
if main_process():
logger.info("=> loading weight '{}'".format(args.weight))
checkpoint = torch.load(args.weight)
model.load_state_dict(checkpoint['state_dict'])
if main_process():
logger.info("=> loaded weight '{}'".format(args.weight))
else:
if main_process():
logger.info("=> no weight found at '{}'".format(args.weight))
# 5.2 加载上次没训练完的模型权重
if args.resume:
if os.path.isfile(args.resume):
if main_process():
logger.info("=> loading checkpoint '{}'".format(args.resume))
# checkpoint = torch.load(args.resume)
checkpoint = torch.load(
args.resume, map_location=lambda storage, loc: storage.cuda())
args.start_epoch = checkpoint['epoch']
model.load_state_dict(checkpoint['state_dict'])
optimizer.load_state_dict(checkpoint['optimizer'])
if main_process():
logger.info("=> loaded checkpoint '{}' (epoch {})".format(
args.resume, checkpoint['epoch']))
else:
if main_process():
logger.info("=> no checkpoint found at '{}'".format(
args.resume))
## step.7 设置数据loader
# 7.1 loader参数设置
value_scale = 255
mean = [0.485, 0.456, 0.406]
mean = [item * value_scale for item in mean]
std = [0.229, 0.224, 0.225]
std = [item * value_scale for item in std]
train_transform = transform.Compose([
transform.RandScale([args.scale_min, args.scale_max]),
transform.RandRotate([args.rotate_min, args.rotate_max],
padding=mean,
ignore_label=args.ignore_label),
transform.RandomGaussianBlur(),
transform.RandomHorizontalFlip(),
transform.Crop([args.train_h, args.train_w],
crop_type='rand',
padding=mean,
ignore_label=args.ignore_label),
transform.ToTensor(),
transform.Normalize(mean=mean, std=std)
]) # 组合数据预处理
# 7.2 训练数据, 可以根据需要自己修改或写
# ---------------------------------------------- 根据实际情况自己写 ---------------------------------------------#
train_data = dataset.SemData(split='train',
data_root=args.data_root,
data_list=args.train_list,
transform=train_transform)
# ---------------------------------------------------- END ---------------------------------------------------#
if args.distributed:
train_sampler = torch.utils.data.distributed.DistributedSampler(
train_data) # 分布式下数据loader
else:
train_sampler = None
train_loader = torch.utils.data.DataLoader(train_data,
batch_size=args.batch_size,
shuffle=(train_sampler is None),
num_workers=args.workers,
pin_memory=True,
sampler=train_sampler,
drop_last=True)
if args.evaluate: # evaluate数据
val_transform = transform.Compose([
transform.Crop([args.train_h, args.train_w],
crop_type='center',
padding=mean,
ignore_label=args.ignore_label),
transform.ToTensor(),
transform.Normalize(mean=mean, std=std)
])
val_data = dataset.SemData(split='val',
data_root=args.data_root,
data_list=args.val_list,
transform=val_transform)
if args.distributed:
val_sampler = torch.utils.data.distributed.DistributedSampler(
val_data)
else:
val_sampler = None
val_loader = torch.utils.data.DataLoader(
val_data,
batch_size=args.batch_size_val,
shuffle=False,
num_workers=args.workers,
pin_memory=True,
sampler=val_sampler)
## step.8 主循环
for epoch in range(args.start_epoch, args.epochs):
epoch_log = epoch + 1
if args.distributed:
train_sampler.set_epoch(epoch)
# 8.1 训练函数
# ---------------------------------------------- 根据实际情况自己写 ---------------------------------------------#
loss_train, mIoU_train, mAcc_train, allAcc_train = train(
train_loader, model, optimizer, epoch)
# ---------------------------------------------------- END ---------------------------------------------------#
if main_process():
writer.add_scalar('loss_train', loss_train, epoch_log)
writer.add_scalar('mIoU_train', mIoU_train, epoch_log)
writer.add_scalar('mAcc_train', mAcc_train, epoch_log)
writer.add_scalar('allAcc_train', allAcc_train, epoch_log)
# 8.2 保存checkpoint
if (epoch_log % args.save_freq == 0) and main_process():
filename = args.save_path + '/train_epoch_' + str(
epoch_log) + '.pth'
logger.info('Saving checkpoint to: ' + filename)
torch.save(
{
'epoch': epoch_log,
'state_dict': model.state_dict(),
'optimizer': optimizer.state_dict()
}, filename)
if epoch_log / args.save_freq > 2:
deletename = args.save_path + '/train_epoch_' + str(
epoch_log - args.save_freq * 2) + '.pth'
os.remove(deletename)
# 训练一个epoch之后evaluate
if args.evaluate:
loss_val, mIoU_val, mAcc_val, allAcc_val = validate(
val_loader, model, criterion)
if main_process():
writer.add_scalar('loss_val', loss_val, epoch_log)
writer.add_scalar('mIoU_val', mIoU_val, epoch_log)
writer.add_scalar('mAcc_val', mAcc_val, epoch_log)
writer.add_scalar('allAcc_val', allAcc_val, epoch_log)
def main(
config_name,
weights_url='https://github.com/deepparrot/semseg/releases/download/0.1/pspnet50-ade20k.pth',
weights_name='pspnet50-ade20k.pth'):
args = config.load_cfg_from_cfg_file(config_name)
check(args)
os.environ["CUDA_VISIBLE_DEVICES"] = ','.join(
str(x) for x in args.test_gpu)
value_scale = 255
mean = [0.485, 0.456, 0.406]
mean = [item * value_scale for item in mean]
std = [0.229, 0.224, 0.225]
std = [item * value_scale for item in std]
gray_folder = os.path.join(args.save_folder, 'gray')
color_folder = os.path.join(args.save_folder, 'color')
args.data_root = './.data/vision/ade20k'
args.val_list = './.data/vision/ade20k/validation.txt'
args.test_list = './.data/vision/ade20k/validation.txt'
print(args.data_root)
test_transform = transform.Compose([transform.ToTensor()])
test_data = dataset.SemData(split=args.split,
data_root=args.data_root,
data_list=args.test_list,
transform=test_transform)
index_start = args.index_start
if args.index_step == 0:
index_end = len(test_data.data_list)
else:
index_end = min(index_start + args.index_step,
len(test_data.data_list))
test_data.data_list = test_data.data_list[index_start:index_end]
test_loader = torch.utils.data.DataLoader(test_data,
batch_size=1,
shuffle=False,
num_workers=args.workers,
pin_memory=True)
colors = np.loadtxt(args.colors_path).astype('uint8')
names = []
if not args.has_prediction:
if args.arch == 'psp':
from model.pspnet import PSPNet
model = PSPNet(layers=args.layers,
classes=args.classes,
zoom_factor=args.zoom_factor,
pretrained=False)
elif args.arch == 'psa':
from model.psanet import PSANet
model = PSANet(layers=args.layers,
classes=args.classes,
zoom_factor=args.zoom_factor,
compact=args.compact,
shrink_factor=args.shrink_factor,
mask_h=args.mask_h,
mask_w=args.mask_w,
normalization_factor=args.normalization_factor,
psa_softmax=args.psa_softmax,
pretrained=False)
model = torch.nn.DataParallel(model).cuda()
cudnn.benchmark = True
local_checkpoint, _ = urllib.request.urlretrieve(
weights_url, weights_name)
if os.path.isfile(local_checkpoint):
checkpoint = torch.load(local_checkpoint)
model.load_state_dict(checkpoint['state_dict'], strict=False)
else:
raise RuntimeError(
"=> no checkpoint found at '{}'".format(local_checkpoint))
test(test_loader, test_data.data_list, model, args.classes, mean, std,
args.base_size, args.test_h, args.test_w, args.scales,
gray_folder, color_folder, colors)
if args.split != 'test':
cal_acc(test_data.data_list, gray_folder, args.classes, names)
def main_worker(gpu, ngpus_per_node, argss):
global args
args = argss
if args.distributed:
if args.dist_url == "env://" and args.rank == -1:
args.rank = int(os.environ["RANK"])
if args.multiprocessing_distributed:
args.rank = args.rank * ngpus_per_node + gpu
dist.init_process_group(backend=args.dist_backend,
init_method=args.dist_url,
world_size=args.world_size,
rank=args.rank)
criterion = nn.CrossEntropyLoss(ignore_index=args.ignore_label)
if args.arch == 'psp':
from model.pspnet import PSPNet
model = PSPNet(layers=args.layers,
classes=args.classes,
zoom_factor=args.zoom_factor,
criterion=criterion,
args=args)
modules_ori = [
model.layer0, model.layer1, model.layer2, model.layer3,
model.layer4
]
modules_new = [model.ppm, model.cls, model.aux]
elif args.arch == 'psa':
from model.psanet import PSANet
model = PSANet(layers=args.layers,
classes=args.classes,
zoom_factor=args.zoom_factor,
psa_type=args.psa_type,
compact=args.compact,
shrink_factor=args.shrink_factor,
mask_h=args.mask_h,
mask_w=args.mask_w,
normalization_factor=args.normalization_factor,
psa_softmax=args.psa_softmax,
criterion=criterion)
modules_ori = [
model.layer0, model.layer1, model.layer2, model.layer3,
model.layer4
]
modules_new = [model.psa, model.cls, model.aux]
params_list = []
for module in modules_ori:
params_list.append(dict(params=module.parameters(), lr=args.base_lr))
for module in modules_new:
params_list.append(
dict(params=module.parameters(), lr=args.base_lr * 10))
args.index_split = 5
optimizer = torch.optim.SGD(params_list,
lr=args.base_lr,
momentum=args.momentum,
weight_decay=args.weight_decay)
if args.sync_bn:
model = nn.SyncBatchNorm.convert_sync_batchnorm(model)
if main_process():
global logger, writer
logger = get_logger()
writer = SummaryWriter(args.save_path)
logger.info(args)
logger.info("=> creating model ...")
logger.info("Classes: {}".format(args.classes))
logger.info(model)
else:
logger = None
if args.distributed:
torch.cuda.set_device(gpu)
args.batch_size = int(args.batch_size / ngpus_per_node)
args.batch_size_val = int(args.batch_size_val / ngpus_per_node)
args.workers = int(
(args.workers + ngpus_per_node - 1) / ngpus_per_node)
model = torch.nn.parallel.DistributedDataParallel(model.cuda(),
device_ids=[gpu])
else:
model = torch.nn.DataParallel(model.cuda())
if args.weight:
if os.path.isfile(args.weight):
if main_process():
logger.info("=> loading weight '{}'".format(args.weight))
checkpoint = torch.load(args.weight)
model.load_state_dict(checkpoint['state_dict'])
if main_process():
logger.info("=> loaded weight '{}'".format(args.weight))
else:
if main_process():
logger.info("=> no weight found at '{}'".format(args.weight))
if args.resume != 'none':
if os.path.isfile(args.resume):
if main_process():
logger.info("=> loading checkpoint '{}'".format(args.resume))
# checkpoint = torch.load(args.resume)
checkpoint = torch.load(
args.resume, map_location=lambda storage, loc: storage.cuda())
args.start_epoch = checkpoint['epoch']
# model.load_state_dict(checkpoint['state_dict'])
# optimizer.load_state_dict(checkpoint['optimizer'])
# print(checkpoint['optimizer'].keys())
if args.if_remove_cls:
if main_process():
logger.info(
'=====!!!!!!!===== Remove cls layer in resuming...')
checkpoint['state_dict'] = {
x: checkpoint['state_dict'][x]
for x in checkpoint['state_dict'].keys()
if ('module.cls' not in x and 'module.aux' not in x)
}
# checkpoint['optimizer'] = {x: checkpoint['optimizer'][x] for x in checkpoint['optimizer'].keys() if ('module.cls' not in x and 'module.aux' not in x)}
# if main_process():
# print('----', checkpoint['state_dict'].keys())
# print('----', checkpoint['optimizer'].keys())
# print('----1', checkpoint['optimizer']['state'].keys())
model.load_state_dict(checkpoint['state_dict'], strict=False)
if not args.if_remove_cls:
optimizer.load_state_dict(checkpoint['optimizer'])
if main_process():
logger.info("=> loaded checkpoint '{}' (epoch {})".format(
args.resume, checkpoint['epoch']))
else:
if main_process():
logger.info("=> no checkpoint found at '{}'".format(
args.resume))
value_scale = 255
mean = [0.485, 0.456, 0.406]
mean = [item * value_scale for item in mean]
std = [0.229, 0.224, 0.225]
std = [item * value_scale for item in std]
transform_list_train = []
if args.resize:
transform_list_train.append(
transform.Resize((args.resize_h, args.resize_w)))
transform_list_train += [
transform.RandScale([args.scale_min, args.scale_max]),
transform.RandRotate([args.rotate_min, args.rotate_max],
padding=mean,
ignore_label=args.ignore_label),
transform.RandomGaussianBlur(),
transform.RandomHorizontalFlip(),
transform.Crop([args.train_h, args.train_w],
crop_type='rand',
padding=mean,
ignore_label=args.ignore_label),
transform.ToTensor(),
transform.Normalize(mean=mean, std=std)
]
train_transform = transform.Compose(transform_list_train)
train_data = dataset.SemData(split='val',
data_root=args.data_root,
data_list=args.train_list,
transform=train_transform,
logger=logger,
is_master=main_process(),
args=args)
if args.distributed:
train_sampler = torch.utils.data.distributed.DistributedSampler(
train_data)
else:
train_sampler = None
train_loader = torch.utils.data.DataLoader(train_data,
batch_size=args.batch_size,
shuffle=(train_sampler is None),
num_workers=args.workers,
pin_memory=True,
sampler=train_sampler,
drop_last=True)
if args.evaluate:
transform_list_val = []
if args.resize:
transform_list_val.append(
transform.Resize((args.resize_h, args.resize_w)))
transform_list_val += [
transform.Crop([args.train_h, args.train_w],
crop_type='center',
padding=mean,
ignore_label=args.ignore_label),
transform.ToTensor(),
transform.Normalize(mean=mean, std=std)
]
val_transform = transform.Compose(transform_list_val)
val_data = dataset.SemData(split='val',
data_root=args.data_root,
data_list=args.val_list,
transform=val_transform,
is_master=main_process(),
args=args)
args.read_image = val_data.read_image
if args.distributed:
val_sampler = torch.utils.data.distributed.DistributedSampler(
val_data)
else:
val_sampler = None
val_loader = torch.utils.data.DataLoader(
val_data,
batch_size=args.batch_size_val,
shuffle=False,
num_workers=args.workers,
pin_memory=True,
sampler=val_sampler)
for epoch in range(args.start_epoch, args.epochs):
epoch_log = epoch + 1
# if args.evaluate and args.val_every_iter == -1:
# # logger.info('Validating.....')
# loss_val, mIoU_val, mAcc_val, allAcc_val, return_dict = validate(val_loader, model, criterion, args)
# if main_process():
# writer.add_scalar('VAL/loss_val', loss_val, epoch_log)
# writer.add_scalar('VAL/mIoU_val', mIoU_val, epoch_log)
# writer.add_scalar('VAL/mAcc_val', mAcc_val, epoch_log)
# writer.add_scalar('VAL/allAcc_val', allAcc_val, epoch_log)
# for sample_idx in range(len(return_dict['image_name_list'])):
# writer.add_text('VAL-image_name/%d'%sample_idx, return_dict['image_name_list'][sample_idx], epoch)
# writer.add_image('VAL-image/%d'%sample_idx, return_dict['im_list'][sample_idx], epoch, dataformats='HWC')
# writer.add_image('VAL-color_label/%d'%sample_idx, return_dict['color_GT_list'][sample_idx], epoch, dataformats='HWC')
# writer.add_image('VAL-color_pred/%d'%sample_idx, return_dict['color_pred_list'][sample_idx], epoch, dataformats='HWC')
if args.distributed:
train_sampler.set_epoch(epoch)
loss_train, mIoU_train, mAcc_train, allAcc_train = train(
train_loader, model, optimizer, epoch, epoch_log, val_loader,
criterion)
if main_process():
writer.add_scalar('TRAIN/loss_train', loss_train, epoch_log)
writer.add_scalar('TRAIN/mIoU_train', mIoU_train, epoch_log)
writer.add_scalar('TRAIN/mAcc_train', mAcc_train, epoch_log)
writer.add_scalar('TRAIN/allAcc_train', allAcc_train, epoch_log)
def main_worker(local_rank, ngpus_per_node, argss):
global args
args = argss
dist.init_process_group(backend=args.dist_backend)
teacher_model = None
if args.teacher_model_path:
teacher_model = PSPNet(layers=args.teacher_layers, classes=args.classes, zoom_factor=args.zoom_factor)
kd_path = 'alpha_' + str(args.alpha) + '_Temp_' + str(args.temperature)
args.save_path = os.path.join(args.save_path, kd_path)
if not os.path.exists(args.save_path):
os.mkdir(args.save_path)
if args.arch == 'psp':
model = PSPNet(layers=args.layers, classes=args.classes, zoom_factor=args.zoom_factor)
modules_ori = [model.layer0, model.layer1, model.layer2, model.layer3, model.layer4]
modules_new = [model.ppm, model.cls, model.aux]
elif args.arch == 'bise_v1':
model = BiseNet(num_classes=args.classes)
modules_ori = [model.sp, model.cp]
modules_new = [model.ffm, model.conv_out, model.conv_out16, model.conv_out32]
params_list = []
for module in modules_ori:
params_list.append(dict(params=module.parameters(), lr=args.base_lr))
for module in modules_new:
params_list.append(dict(params=module.parameters(), lr=args.base_lr * 10))
args.index_split = 5
optimizer = torch.optim.SGD(params_list, lr=args.base_lr, momentum=args.momentum, weight_decay=args.weight_decay)
if args.sync_bn:
model = nn.SyncBatchNorm.convert_sync_batchnorm(model)
if teacher_model is not None:
teacher_model = nn.SyncBatchNorm.convert_sync_batchnorm(teacher_model)
if main_process():
global logger, writer
logger = get_logger()
writer = SummaryWriter(args.save_path) # tensorboardX
logger.info(args)
logger.info("=> creating model ...")
logger.info("Classes: {}".format(args.classes))
logger.info(model)
if teacher_model is not None:
logger.info(teacher_model)
if args.distributed:
torch.cuda.set_device(local_rank)
args.batch_size = int(args.batch_size / ngpus_per_node)
args.batch_size_val = int(args.batch_size_val / ngpus_per_node)
args.workers = int((args.workers + ngpus_per_node - 1) / ngpus_per_node)
model = torch.nn.parallel.DistributedDataParallel(model.cuda(), device_ids=[local_rank])
if teacher_model is not None:
teacher_model = torch.nn.parallel.DistributedDataParallel(teacher_model.cuda(), device_ids=[local_rank])
else:
model = torch.nn.DataParallel(model.cuda())
if teacher_model is not None:
teacher_model = torch.nn.DataParallel(teacher_model.cuda())
if teacher_model is not None:
checkpoint = torch.load(args.teacher_model_path, map_location=lambda storage, loc: storage.cuda())
teacher_model.load_state_dict(checkpoint['state_dict'], strict=False)
print("=> loading teacher checkpoint '{}'".format(args.teacher_model_path))
criterion = nn.CrossEntropyLoss(ignore_index=args.ignore_label).cuda(local_rank)
kd_criterion = None
if teacher_model is not None:
kd_criterion = KDLoss(ignore_index=args.ignore_label).cuda(local_rank)
if args.weight:
if os.path.isfile(args.weight):
if main_process():
logger.info("=> loading weight: '{}'".format(args.weight))
checkpoint = torch.load(args.weight)
model.load_state_dict(checkpoint['state_dict'])
if main_process():
logger.info("=> loaded weight '{}'".format(args.weight))
else:
if main_process():
logger.info("=> mp weight found at '{}'".format(args.weight))
best_mIoU_val = 0.0
if args.resume:
if os.path.isfile(args.resume):
if main_process():
logger.info("=> loading checkpoint '{}'".format(args.resume))
# Load all tensors onto GPU
checkpoint = torch.load(args.resume, map_location=lambda storage, loc: storage.cuda())
args.start_epoch = checkpoint['epoch']
model.load_state_dict(checkpoint['state_dict'])
optimizer.load_state_dict(checkpoint['optimizer'])
best_mIoU_val = checkpoint['best_mIoU_val']
if main_process():
logger.info("=> loaded checkpoint '{}' (epoch {})".format(args.resume, checkpoint['point']))
else:
if main_process():
logger.info("=> no checkpoint found at '{}'".format(args.resume))
value_scale = 255
mean = [0.485, 0.456, 0.406]
mean = [item * value_scale for item in mean]
std = [0.229, 0.224, 0.225]
std = [item * value_scale for item in std]
train_transform = transform.Compose([
transform.RandScale([args.scale_min, args.scale_max]),
transform.RandRotate([args.rotate_min, args.rotate_max], padding=mean, ignore_label=args.ignore_label),
transform.RandomGaussianBlur(),
transform.RandomHorizontalFlip(),
transform.Crop([args.train_h, args.train_w], crop_type='rand', padding=mean, ignore_label=args.ignore_label),
transform.ToTensor(),
transform.Normalize(mean=mean, std=std)])
train_data = dataset.SemData(split='train', data_root=args.data_root, data_list=args.train_list, transform=train_transform)
if args.distributed:
train_sampler = torch.utils.data.distributed.DistributedSampler(train_data)
else:
train_sampler = None
train_loader = torch.utils.data.DataLoader(train_data, batch_size=args.batch_size, shuffle=(train_sampler is None), num_workers=args.workers, pin_memory=True, sampler=train_sampler, drop_last=True)
if args.evaluate:
val_transform = transform.Compose([
transform.Crop([args.train_h, args.train_w], crop_type='center', padding=mean, ignore_label=args.ignore_label),
transform.ToTensor(),
transform.Normalize(mean=mean, std=std)])
val_data = dataset.SemData(split='val', data_root=args.data_root, data_list=args.val_list, transform=val_transform)
if args.distributed:
val_sampler = torch.utils.data.distributed.DistributedSampler(val_data)
else:
val_sampler = None
val_loader = torch.utils.data.DataLoader(val_data, batch_size=args.batch_size_val, shuffle=False, num_workers=args.workers, pin_memory=True, sampler=val_sampler)
for epoch in range(args.start_epoch, args.epochs):
epoch_log = epoch + 1
if args.distributed:
# Use .set_epoch() method to reshuffle the dataset partition at every iteration
train_sampler.set_epoch(epoch)
loss_train, mIoU_train, mAcc_train, allAcc_train = train(local_rank, train_loader, model, teacher_model, criterion, kd_criterion, optimizer, epoch)
if main_process():
writer.add_scalar('loss_train', loss_train, epoch_log)
writer.add_scalar('mIoU_train', mIoU_train, epoch_log)
writer.add_scalar('mAcc_train', mAcc_train, epoch_log)
writer.add_scalar('allAcc_train', allAcc_train, epoch_log)
is_best = False
if args.evaluate:
loss_val, mIoU_val, mAcc_val, allAcc_val = validate(local_rank, val_loader, model, criterion)
if main_process():
writer.add_scalar('loss_val', loss_val, epoch_log)
writer.add_scalar('mIoU_val', mIoU_val, epoch_log)
writer.add_scalar('mAcc_val', mAcc_val, epoch_log)
writer.add_scalar('allAcc_val', allAcc_val, epoch_log)
if best_mIoU_val < mIoU_val:
is_best = True
best_mIoU_val = mIoU_val
logger.info('==>The best val mIoU: %.3f' % (best_mIoU_val))
if (epoch_log % args.save_freq == 0) and main_process():
save_checkpoint(
{
'epoch': epoch_log,
'state_dict': model.state_dict(),
'optimizer': optimizer.state_dict(),
'best_mIoU_val': best_mIoU_val
},
is_best,
args.save_path
)
if is_best:
logger.info('Saving checkpoint to:' + args.save_path + '/best.pth with mIoU: ' + str(best_mIoU_val) )
else:
logger.info('Saving checkpoint to:' + args.save_path + '/last.pth with mIoU: ' + str(mIoU_val) )
if main_process():
writer.close() # it must close the writer, otherwise it will appear the EOFError!
logger.info('==>Training done!\nBest mIoU: %.3f' % (best_mIoU_val))
def main():
global args, logger
args = get_parser()
if args.test_in_nyu_label_space:
args.colors_path = 'nyu/nyu_colors.txt'
args.names_path = 'nyu/nyu_names.txt'
if args.if_cluster:
args.data_root = args.data_root_cluster
args.project_path = args.project_path_cluster
args.data_config_path = 'data'
for key in ['train_list', 'val_list', 'test_list', 'colors_path', 'names_path']:
args[key] = os.path.join(args.data_config_path, args[key])
for key in ['save_path', 'model_path', 'save_folder']:
args[key] = os.path.join(args.project_path, args[key])
# for key in ['save_path', 'model_path', 'save_folder']:
# args[key] = args[key] % args.exp_name
check(args)
logger = get_logger()
os.environ["CUDA_VISIBLE_DEVICES"] = ','.join(str(x) for x in args.test_gpu)
logger.info(args)
logger.info("=> creating model ...")
logger.info("Classes: {}".format(args.classes))
value_scale = 255
mean = [0.485, 0.456, 0.406]
mean = [item * value_scale for item in mean]
std = [0.229, 0.224, 0.225]
std = [item * value_scale for item in std]
gray_folder = os.path.join(args.save_folder, 'gray')
color_folder = os.path.join(args.save_folder, 'color')
transform_list_test = []
if args.resize:
transform_list_test.append(transform.Resize((args.resize_h_test, args.resize_w_test)))
transform_list_test += [
transform.Crop([args.test_h, args.test_w], crop_type='center', padding=mean, ignore_label=args.ignore_label),
transform.ToTensor(),
transform.Normalize(mean=mean, std=std)
]
test_transform = transform.Compose(transform_list_test)
test_data = dataset.SemData(split=args.split, data_root=args.data_root, data_list=args.test_list, transform=test_transform, is_master=True, args=args)
# test_data = dataset.SemData(split='val', data_root=args.data_root, data_list=args.val_list, transform=test_transform, is_master=True, args=args)
index_start = args.index_start
if args.index_step == 0:
index_end = len(test_data.data_list)
else:
index_end = min(index_start + args.index_step, len(test_data.data_list))
test_data.data_list = test_data.data_list[index_start:index_end]
test_loader = torch.utils.data.DataLoader(test_data, batch_size=1, shuffle=False, num_workers=args.workers, pin_memory=True)
colors = np.loadtxt(args.colors_path).astype('uint8')
names = [line.rstrip('\n') for line in open(args.names_path)]
args.read_image = test_data.read_image
if not args.has_prediction:
if args.arch == 'psp':
from model.pspnet import PSPNet
model = PSPNet(layers=args.layers, classes=args.classes, zoom_factor=args.zoom_factor, pretrained=False)
elif args.arch == 'psa':
from model.psanet import PSANet
model = PSANet(layers=args.layers, classes=args.classes, zoom_factor=args.zoom_factor, compact=args.compact,
shrink_factor=args.shrink_factor, mask_h=args.mask_h, mask_w=args.mask_w,
normalization_factor=args.normalization_factor, psa_softmax=args.psa_softmax, pretrained=False)
logger.info(model)
model = torch.nn.DataParallel(model).cuda()
cudnn.benchmark = True
if os.path.isfile(args.model_path):
logger.info("=> loading checkpoint '{}'".format(args.model_path))
checkpoint = torch.load(args.model_path)
model.load_state_dict(checkpoint['state_dict'], strict=True)
logger.info("=> loaded checkpoint '{}'".format(args.model_path))
else:
raise RuntimeError("=> no checkpoint found at '{}'".format(args.model_path))
pred_path_list, target_path_list = test(test_loader, test_data.data_list, model, args.classes, mean, std, args.base_size, args.test_h, args.test_w, args.scales, gray_folder, color_folder, colors)
if args.split != 'test' or (args.split == 'test' and args.test_has_gt):
cal_acc(test_data.data_list, gray_folder, args.classes, names, pred_path_list=pred_path_list, target_path_list=target_path_list)
def main():
global args, logger
args = get_parser()
check(args)
logger = get_logger()
os.environ["CUDA_VISIBLE_DEVICES"] = ','.join(str(x) for x in args.test_gpu)
logger.info(args)
logger.info("=> creating model ...")
logger.info("Classes: {}".format(args.classes))
value_scale = 255
mean = [0.485, 0.456, 0.406]
mean = [item * value_scale for item in mean]
std = [0.229, 0.224, 0.225]
std = [item * value_scale for item in std]
test_h = int(args.base_h * args.scale + 1)
test_w = int(args.base_w * args.scale + 1)
if args.teacher_model_path:
kd_path = 'alpha_' + str(args.alpha) + '_Temp_' + str(args.temperature)
kd_save = kd_path + '/val/ss'
args.save_folder = os.path.join(args.save_folder, kd_save)
args.save_path = os.path.join(args.save_path, kd_path)
args.model_path = os.path.join(args.save_path, args.model_path)
gray_folder = os.path.join(args.save_folder, 'gray')
color_folder = os.path.join(args.save_folder, 'color')
test_transform = transform.Compose([transform.ToTensor()])
test_data = dataset.SemData(split='test', data_root=args.data_root, data_list=args.test_list, transform=test_transform)
index_start = args.index_start
if args.index_step == 0:
index_end = len(test_data.data_list)
else:
index_end = min(index_start + args.index_step, len(test_data.data_list))
test_data.data_list = test_data.data_list[index_start:index_end]
test_loader = torch.utils.data.DataLoader(test_data, batch_size=1, shuffle=False, num_workers=4, pin_memory=True)
colors = np.loadtxt(args.colors_path).astype('uint8')
names = [line.rstrip('\n') for line in open(args.names_path)]
if not args.has_prediction:
if args.arch == 'psp':
from model.pspnet import PSPNet
model = PSPNet(layers=args.layers, classes=args.classes, zoom_factor=args.zoom_factor)
elif args.arch == 'nonlocal':
from model.nonlocal_net import Nonlocal
model = Nonlocal(layers=args.layers, classes=args.classes, zoom_factor=args.zoom_factor)
elif args.arch == 'sanet':
from model.sanet import SANet
model = SANet(layers=args.layers, classes=args.classes, zoom_factor=args.zoom_factor)
elif args.arch == 'bise_v1':
from model.bisenet_v1 import BiseNet
model = BiseNet(layers=args.layers, classes=args.classes, with_sp=args.with_sp)
elif args.arch == 'fanet':
from model.fanet import FANet
model = FANet(layers=args.layers, classes=args.classes)
logger.info(model)
model = torch.nn.DataParallel(model).cuda()
cudnn.benchmark = True
if os.path.isfile(args.model_path):
logger.info("=> loading checkpoint '{}'".format(args.model_path))
checkpoint = torch.load(args.model_path)
model.load_state_dict(checkpoint['state_dict'])
logger.info("=> loaded checkpoint '{}'".format(args.model_path))
else:
raise RuntimeError("=> no checkpoint found at '{}'".format(args.model_path))
test(test_loader, test_data.data_list, model, args.classes, mean, std, args.base_size, test_h, test_w, args.scales, gray_folder, color_folder, colors)
if args.split != 'test':
cal_acc(test_data.data_list, gray_folder, args.classes, names)
def main():
models = {
'resnet101':
lambda: PSPNet(n_classes=21,
sizes=(1, 2, 3, 6),
psp_size=2048,
deep_features_size=1024,
backend='resnet101')
}
h, w = map(int, args.input_size.split(','))
input_size = (h, w)
cudnn.enabled = True
# create network
model = models['resnet101']()
# load pretrained parameters
if args.restore_from[:4] == 'http':
#saved_state_dict = torch.load(args.restore_from)
saved_state_dict = model_zoo.load_url(args.restore_from)
else:
saved_state_dict = torch.load(args.restore_from)
# only copy the params that exist in current model (caffe-like)
new_params = model.state_dict().copy()
for name, param in new_params.items():
print(name)
if name in saved_state_dict and param.size(
) == saved_state_dict[name].size():
new_params[name].copy_(saved_state_dict[name])
print('copy {}'.format(name))
model.load_state_dict(new_params)
model.train()
model = nn.DataParallel(model)
model.cuda()
cudnn.benchmark = True
# init D
# model_D = FCDiscriminator(num_classes=args.num_classes)
# if args.restore_from_D is not None:
# model_D.load_state_dict(torch.load(args.restore_from_D))
#
# model_D = nn.DataParallel(model_D)
# model_D.train()
# model_D.cuda()
if not os.path.exists(args.snapshot_dir):
os.makedirs(args.snapshot_dir)
train_dataset = VOCDataSet(args.data_dir,
args.data_list,
crop_size=input_size,
scale=args.random_scale,
mirror=args.random_mirror,
mean=IMG_MEAN)
train_dataset_size = len(train_dataset)
train_gt_dataset = VOCGTDataSet(args.data_dir,
args.data_list,
crop_size=input_size,
scale=args.random_scale,
mirror=args.random_mirror,
mean=IMG_MEAN)
if args.partial_data == 0:
trainloader = data.DataLoader(train_dataset,
batch_size=args.batch_size,
shuffle=True,
num_workers=5,
pin_memory=True)
trainloader_gt = data.DataLoader(train_gt_dataset,
batch_size=args.batch_size,
shuffle=True,
num_workers=5,
pin_memory=True)
else:
#sample partial data
partial_size = int(args.partial_data * train_dataset_size)
trainloader_iter = enumerate(trainloader)
trainloader_gt_iter = enumerate(trainloader_gt)
# implement model.optim_parameters(args) to handle different models' lr setting
# optimizer for segmentation network
optimizer = optim.SGD(model.parameters(),
lr=args.learning_rate,
momentum=args.momentum,
weight_decay=args.weight_decay)
optimizer.zero_grad()
# optimizer for discriminator network
# optimizer_D = optim.Adam(model_D.parameters(), lr=args.learning_rate_D, betas=(0.9,0.99))
# optimizer_D.zero_grad()
# loss/ bilinear upsampling
bce_loss = BCEWithLogitsLoss2d()
interp = nn.Upsample(size=(input_size[1], input_size[0]), mode='bilinear')
if version.parse(torch.__version__) >= version.parse('0.4.0'):
interp = nn.Upsample(size=(input_size[1], input_size[0]),
mode='bilinear',
align_corners=True)
else:
interp = nn.Upsample(size=(input_size[1], input_size[0]),
mode='bilinear')
# labels for adversarial training
pred_label = 0
gt_label = 1
seg_criterion = NLL2d().cuda()
cls_criterion = nn.BCEWithLogitsLoss(weight=None)
for i_iter in range(args.num_steps):
loss_seg_value = 0
loss_adv_pred_value = 0
loss_D_value = 0
loss_semi_value = 0
loss_semi_adv_value = 0
optimizer.zero_grad()
adjust_learning_rate(optimizer, i_iter)
# optimizer_D.zero_grad()
# adjust_learning_rate_D(optimizer_D, i_iter)
for sub_i in range(args.iter_size):
# train G
# train with source
try:
_, batch = trainloader_iter.next()
except:
trainloader_iter = enumerate(trainloader)
_, batch = trainloader_iter.next()
images, labels, _, _, y_cls = batch
labels = Variable(labels.long()).cuda()
y_cls = Variable(y_cls.float()).cuda()
images = Variable(images).cuda()
#ignore_mask = (labels.numpy() == 255)
out, out_cls = model(images)
seg_loss, cls_loss = seg_criterion(out, labels), cls_criterion(
out_cls, y_cls)
loss = seg_loss + cls_loss
# proper normalization
loss = loss / args.iter_size
loss.backward()
loss_seg_value += seg_loss.data.cpu().numpy()[0] / args.iter_size
optimizer.step()
# optimizer_D.step()
print('exp = {}'.format(args.snapshot_dir))
print(
'iter = {0:8d}/{1:8d}, loss_seg = {2:.3f}, loss_adv_p = {3:.3f}, loss_D = {4:.3f}, loss_semi = {5:.3f}, loss_semi_adv = {6:.3f}'
.format(i_iter, args.num_steps, loss_seg_value,
loss_adv_pred_value, loss_D_value, loss_semi_value,
loss_semi_adv_value))
if i_iter >= args.num_steps - 1:
print('save model ...')
torch.save(
model.state_dict(),
osp.join(
args.snapshot_dir, 'VOC_' +
os.path.abspath(__file__).split('/')[-1].split('.')[0] +
'_' + str(args.num_steps) + '.pth'))
#torch.save(model_D.state_dict(),osp.join(args.snapshot_dir, 'VOC_'+os.path.abspath(__file__).split('/')[-1].split('.')[0]+'_'+str(args.num_steps)+'_D.pth'))
break
if i_iter % args.save_pred_every == 0 and i_iter != 0:
print('taking snapshot ...')
torch.save(
model.state_dict(),
osp.join(
args.snapshot_dir, 'VOC_' +
os.path.abspath(__file__).split('/')[-1].split('.')[0] +
'_' + str(i_iter) + '.pth'))
#torch.save(model_D.state_dict(),osp.join(args.snapshot_dir, 'VOC_'+os.path.abspath(__file__).split('/')[-1].split('.')[0]+'_'+str(i_iter)+'_D.pth'))
end = timeit.default_timer()
print(end - start, 'seconds')
#-*-coding:utf-8-*-
import torch
import torch.nn as nn
from model.pspnet import PSPNet
import torch.nn.functional as F
from torchsummary import summary
import numpy as np
psp_models = {
'resnet18':
lambda: PSPNet(sizes=(1, 2, 3, 6),
psp_size=512,
deep_features_size=256,
backend='resnet18')
}
# 之前的pspnet模型，用于语义分割
class ModifiedResnet(nn.Module):
def __init__(self):
super(ModifiedResnet, self).__init__()
self.model = psp_models['resnet18'.lower()]()
# self.model = nn.DataParallel(self.model) # 数据并行，用于多GPU的调用
def forward(self, x):
return self.model(x)
class PoseNetFeat(nn.Module):
def __init__(self, num_points):
'''
def test_train(self):
dataset = SegmentationDummyDataset(batch_size=32)
pspnet = PSPNet(epochs=1, dataset=dataset, frontend_name='resnet101')
history = pspnet.train()
ok_('loss' in history)
def main_worker(gpu, ngpus_per_node, argss):
global args
args = argss
if args.sync_bn:
if args.multiprocessing_distributed:
BatchNorm = apex.parallel.SyncBatchNorm
else:
from lib.sync_bn.modules import BatchNorm2d
BatchNorm = BatchNorm2d
else:
BatchNorm = nn.BatchNorm2d
if args.distributed:
if args.dist_url == "env://" and args.rank == -1:
args.rank = int(os.environ["RANK"])
if args.multiprocessing_distributed:
args.rank = args.rank * ngpus_per_node + gpu
dist.init_process_group(backend=args.dist_backend, init_method=args.dist_url, world_size=args.world_size, rank=args.rank)
criterion = nn.CrossEntropyLoss(ignore_index=args.ignore_label)
if args.arch == 'psp':
from model.pspnet import PSPNet
model = PSPNet(layers=args.layers, classes=args.classes, zoom_factor=args.zoom_factor, criterion=criterion, BatchNorm=BatchNorm)
modules_ori = [model.layer0, model.layer1, model.layer2, model.layer3, model.layer4]
modules_new = [model.ppm, model.cls, model.aux]
elif args.arch == 'psa':
from model.psanet import PSANet
model = PSANet(layers=args.layers, classes=args.classes, zoom_factor=args.zoom_factor, psa_type=args.psa_type,
compact=args.compact, shrink_factor=args.shrink_factor, mask_h=args.mask_h, mask_w=args.mask_w,
normalization_factor=args.normalization_factor, psa_softmax=args.psa_softmax,
criterion=criterion,
BatchNorm=BatchNorm)
modules_ori = [model.layer0, model.layer1, model.layer2, model.layer3, model.layer4]
modules_new = [model.psa, model.cls, model.aux]
params_list = []
for module in modules_ori:
params_list.append(dict(params=module.parameters(), lr=args.base_lr))
for module in modules_new:
params_list.append(dict(params=module.parameters(), lr=args.base_lr * 10))
args.index_split = 5
optimizer = torch.optim.SGD(params_list, lr=args.base_lr, momentum=args.momentum, weight_decay=args.weight_decay)
if main_process():
global logger, writer
logger = get_logger()
writer = SummaryWriter(args.save_path)
logger.info(args)
logger.info("=> creating model ...")
logger.info("Classes: {}".format(args.classes))
logger.info(model)
if args.distributed:
torch.cuda.set_device(gpu)
args.batch_size = int(args.batch_size / ngpus_per_node)
args.batch_size_val = int(args.batch_size_val / ngpus_per_node)
args.workers = int(args.workers / ngpus_per_node)
if args.use_apex:
model, optimizer = apex.amp.initialize(model.cuda(), optimizer, opt_level=args.opt_level, keep_batchnorm_fp32=args.keep_batchnorm_fp32, loss_scale=args.loss_scale)
model = apex.parallel.DistributedDataParallel(model)
else:
model = torch.nn.parallel.DistributedDataParallel(model.cuda(), device_ids=[gpu])
else:
model = torch.nn.DataParallel(model.cuda())
if args.weight:
if os.path.isfile(args.weight):
if main_process():
logger.info("=> loading weight '{}'".format(args.weight))
checkpoint = torch.load(args.weight)
model.load_state_dict(checkpoint['state_dict'])
if main_process():
logger.info("=> loaded weight '{}'".format(args.weight))
else:
if main_process():
logger.info("=> no weight found at '{}'".format(args.weight))
if args.resume:
if os.path.isfile(args.resume):
if main_process():
logger.info("=> loading checkpoint '{}'".format(args.resume))
# checkpoint = torch.load(args.resume)
checkpoint = torch.load(args.resume, map_location=lambda storage, loc: storage.cuda())
args.start_epoch = checkpoint['epoch']
model.load_state_dict(checkpoint['state_dict'])
optimizer.load_state_dict(checkpoint['optimizer'])
if main_process():
logger.info("=> loaded checkpoint '{}' (epoch {})".format(args.resume, checkpoint['epoch']))
else:
if main_process():
logger.info("=> no checkpoint found at '{}'".format(args.resume))
value_scale = 255
mean = [0.485, 0.456, 0.406]
mean = [item * value_scale for item in mean]
std = [0.229, 0.224, 0.225]
std = [item * value_scale for item in std]
train_transform = transform.Compose([
transform.RandScale([args.scale_min, args.scale_max]),
transform.RandRotate([args.rotate_min, args.rotate_max], padding=mean, ignore_label=args.ignore_label),
transform.RandomGaussianBlur(),
transform.RandomHorizontalFlip(),
transform.Crop([args.train_h, args.train_w], crop_type='rand', padding=mean, ignore_label=args.ignore_label),
transform.ToTensor(),
transform.Normalize(mean=mean, std=std)])
train_data = dataset.SemData(split='train', data_root=args.data_root, data_list=args.train_list, transform=train_transform)
if args.distributed:
train_sampler = torch.utils.data.distributed.DistributedSampler(train_data)
else:
train_sampler = None
train_loader = torch.utils.data.DataLoader(train_data, batch_size=args.batch_size, shuffle=(train_sampler is None), num_workers=args.workers, pin_memory=True, sampler=train_sampler, drop_last=True)
if args.evaluate:
val_transform = transform.Compose([
transform.Crop([args.train_h, args.train_w], crop_type='center', padding=mean, ignore_label=args.ignore_label),
transform.ToTensor(),
transform.Normalize(mean=mean, std=std)])
val_data = dataset.SemData(split='val', data_root=args.data_root, data_list=args.val_list, transform=val_transform)
if args.distributed:
val_sampler = torch.utils.data.distributed.DistributedSampler(val_data)
else:
val_sampler = None
val_loader = torch.utils.data.DataLoader(val_data, batch_size=args.batch_size_val, shuffle=False, num_workers=args.workers, pin_memory=True, sampler=val_sampler)
for epoch in range(args.start_epoch, args.epochs):
epoch_log = epoch + 1
if args.distributed:
train_sampler.set_epoch(epoch)
loss_train, mIoU_train, mAcc_train, allAcc_train = train(train_loader, model, optimizer, epoch)
if main_process():
writer.add_scalar('loss_train', loss_train, epoch_log)
writer.add_scalar('mIoU_train', mIoU_train, epoch_log)
writer.add_scalar('mAcc_train', mAcc_train, epoch_log)
writer.add_scalar('allAcc_train', allAcc_train, epoch_log)
if (epoch_log % args.save_freq == 0) and main_process():
filename = args.save_path + '/train_epoch_' + str(epoch_log) + '.pth'
logger.info('Saving checkpoint to: ' + filename)
torch.save({'epoch': epoch_log, 'state_dict': model.state_dict(), 'optimizer': optimizer.state_dict()}, filename)
if epoch_log / args.save_freq > 2:
deletename = args.save_path + '/train_epoch_' + str(epoch_log - args.save_freq * 2) + '.pth'
os.remove(deletename)
if args.evaluate:
loss_val, mIoU_val, mAcc_val, allAcc_val = validate(val_loader, model, criterion)
if main_process():
writer.add_scalar('loss_val', loss_val, epoch_log)
writer.add_scalar('mIoU_val', mIoU_val, epoch_log)
writer.add_scalar('mAcc_val', mAcc_val, epoch_log)
writer.add_scalar('allAcc_val', allAcc_val, epoch_log)
def test_init(self):
dataset = SegmentationDummyDataset()
PSPNet(dataset=dataset, frontend_name='resnet101')
def main():
global args
criterion = nn.CrossEntropyLoss(ignore_index=args.ignore_label)
model = PSPNet(layers=args.layers,
classes=args.classes,
zoom_factor=args.zoom_factor,
criterion=criterion,
pretrained=args.pretrained,
naive_ppm=args.naive_ppm)
# set diffrent learning rate on different part of models
modules_ori = [
model.layer0, model.layer1, model.layer2, model.layer3, model.layer4
]
modules_new = [model.ppm, model.cls_head, model.aux_head]
params_list = []
for module in modules_ori:
params_list.append(dict(params=module.parameters(), lr=args.base_lr))
for module in modules_new:
params_list.append(
dict(params=module.parameters(), lr=args.base_lr * 10))
args.index_split = 5
optimizer = torch.optim.SGD(params_list,
lr=args.base_lr,
momentum=args.momentum,
weight_decay=args.weight_decay)
global logger, writer
logger = get_logger()
writer = SummaryWriter(args.save_path)
logger.info(args)
logger.info("=> creating model ...")
logger.info("Classes: {}".format(args.classes))
logger.info(model)
model = model.cuda()
model = torch.nn.DataParallel(model).cuda()
if args.weight:
if os.path.isfile(args.weight):
logger.info("=> loading weight '{}'".format(args.weight))
checkpoint = torch.load(args.weight)
model.load_state_dict(checkpoint['state_dict'])
logger.info("=> loaded weight '{}'".format(args.weight))
else:
logger.info("=> no weight found at '{}'".format(args.weight))
if args.resume:
if os.path.isfile(args.resume):
logger.info("=> loading checkpoint '{}'".format(args.resume))
# checkpoint = torch.load(args.resume)
checkpoint = torch.load(
args.resume, map_location=lambda storage, loc: storage.cuda())
args.start_epoch = checkpoint['epoch']
model.load_state_dict(checkpoint['state_dict'])
optimizer.load_state_dict(checkpoint['optimizer'])
logger.info("=> loaded checkpoint '{}' (epoch {})".format(
args.resume, checkpoint['epoch']))
else:
logger.info("=> no checkpoint found at '{}'".format(args.resume))
# image pre-processing and augmentation
value_scale = 255
mean = [0.485, 0.456, 0.406]
mean = [item * value_scale for item in mean]
std = [0.229, 0.224, 0.225]
std = [item * value_scale for item in std]
train_transform = transform.Compose([
transform.Resize((args.train_h, args.train_w)),
# augmentation
transform.RandScale([args.scale_min, args.scale_max]),
transform.RandRotate([args.rotate_min, args.rotate_max],
padding=mean,
ignore_label=args.ignore_label),
transform.RandomGaussianBlur(),
transform.RandomHorizontalFlip(),
transform.Crop([args.train_h, args.train_w],
crop_type='rand',
padding=mean,
ignore_label=args.ignore_label),
transform.ToTensor(),
transform.Normalize(mean=mean, std=std)
])
# initialize dataloader
train_data = dataset.SemData(split='trainval', transform=train_transform)
train_sampler = None
train_loader = torch.utils.data.DataLoader(train_data,
batch_size=args.batch_size,
shuffle=(train_sampler is None),
num_workers=args.workers,
pin_memory=True,
sampler=train_sampler,
drop_last=True)
if args.evaluate:
val_transform = transform.Compose([
transform.Crop([args.train_h, args.train_w],
crop_type='center',
padding=mean,
ignore_label=args.ignore_label),
transform.ToTensor(),
transform.Normalize(mean=mean, std=std)
])
val_data = dataset.SemData(split='test', transform=val_transform)
val_sampler = None
val_loader = torch.utils.data.DataLoader(
val_data,
batch_size=args.batch_size_val,
shuffle=False,
num_workers=args.workers,
pin_memory=True,
sampler=val_sampler)
# start training
logger.info('Starting training.')
for epoch in range(args.start_epoch, args.epochs):
epoch_log = epoch + 1
loss_train, mIoU_train, mAcc_train, allAcc_train = train(
train_loader, model, optimizer, epoch)
writer.add_scalar('loss/train', loss_train, epoch_log)
writer.add_scalar('mIoU/train', mIoU_train, epoch_log)
writer.add_scalar('mAcc/train', mAcc_train, epoch_log)
writer.add_scalar('allAcc/train', allAcc_train, epoch_log)
if args.evaluate:
loss_val, mIoU_val, mAcc_val, allAcc_val = validate(
val_loader, model, criterion)
writer.add_scalar('loss/val', loss_val, epoch_log)
writer.add_scalar('mIoU/val', mIoU_val, epoch_log)
writer.add_scalar('mAcc/val', mAcc_val, epoch_log)
writer.add_scalar('allAcc/val', allAcc_val, epoch_log)
if (epoch_log % args.save_freq == 0):
filename = args.save_path + '/train_epoch_' + str(
epoch_log) + '.pth'
logger.info('Saving checkpoint to: ' + filename)
torch.save(
{
'epoch': epoch_log,
'state_dict': model.module.state_dict(),
'optimizer': optimizer.state_dict()
}, filename)
