def main():
global args, logger
args = get_parser().parse_args()
logger = get_logger()
# os.environ["CUDA_VISIBLE_DEVICES"] = ','.join(str(x) for x in args.gpu)
logger.info(args)
assert args.classes > 1
assert args.zoom_factor in [1, 2, 4, 8]
assert (args.crop_h - 1) % 8 == 0 and (args.crop_w - 1) % 8 == 0
assert args.split in ['train', 'val', 'test']
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]
train_transform = transforms.Compose([
transforms.RandScale([0.5, 2]),
transforms.RandRotate([-10, 10], padding=mean, ignore_label=args.ignore_label),
transforms.RandomGaussianBlur(),
transforms.RandomHorizontalFlip(),
transforms.Crop([args.crop_h, args.crop_w], crop_type='rand', padding=mean, ignore_label=args.ignore_label),
transforms.ToTensor()])
val_transform = transforms.Compose([transforms.Crop([args.crop_h, args.crop_w], crop_type='center', padding=mean, ignore_label=args.ignore_label),
transforms.ToTensor()])
val_data1 = datasets.SegData(split='train', data_root=args.data_root, data_list=args.val_list1, transform=val_transform)
val_loader1 = torch.utils.data.DataLoader(val_data1, batch_size=1, shuffle=False, num_workers=args.workers, pin_memory=True)
from pspnet import PSPNet
model = PSPNet(backbone = args.backbone, layers=args.layers, classes=args.classes, zoom_factor=args.zoom_factor, use_softmax=False, use_aux=False, pretrained=False, syncbn=False).cuda()
logger.info(model)
# model = torch.nn.DataParallel(model).cuda()
model = model.cuda()
cudnn.enabled = True
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))
pretrained_dict = {k.replace('module.',''): v for k, v in checkpoint['state_dict'].items()}
dict1 = model.state_dict()
model.load_state_dict(pretrained_dict, strict=False)
else:
raise RuntimeError("=> no checkpoint found at '{}'".format(args.model_path))
cv2.setNumThreads(0)
validate(val_loader1, val_data1.data_list, model, args.classes, mean, std, args.base_size1, args.crop_h, args.crop_w, args.scales)
def main():
global args, logger, writer
args = get_parser().parse_args()
import multiprocessing as mp
if mp.get_start_method(allow_none=True) != 'spawn':
mp.set_start_method('spawn', force=True)
rank, world_size = dist_init(args.port)
logger = get_logger()
writer = SummaryWriter(args.save_path)
#os.environ["CUDA_VISIBLE_DEVICES"] = ','.join(str(x) for x in args.gpu)
#if len(args.gpu) == 1:
# args.syncbn = False
if rank == 0:
logger.info(args)
assert args.classes > 1
assert args.zoom_factor in [1, 2, 4, 8]
assert (args.crop_h-1) % 8 == 0 and (args.crop_w-1) % 8 == 0
assert args.net_type in [0, 1, 2, 3]
if args.bn_group == 1:
args.bn_group_comm = None
else:
assert world_size % args.bn_group == 0
args.bn_group_comm = simple_group_split(world_size, rank, world_size // args.bn_group)
if rank == 0:
logger.info("=> creating model ...")
logger.info("Classes: {}".format(args.classes))
if args.net_type == 0:
from pspnet import PSPNet
model = PSPNet(backbone=args.backbone, layers=args.layers, classes=args.classes, zoom_factor=args.zoom_factor, syncbn=args.syncbn, group_size=args.bn_group, group=args.bn_group_comm).cuda()
elif args.net_type in [1, 2, 3]:
from pspnet_div4 import PSPNet
model = PSPNet(backbone=args.backbone, layers=args.layers, classes=args.classes, zoom_factor=args.zoom_factor, syncbn=args.syncbn, group_size=args.bn_group, group=args.bn_group_comm, net_type=args.net_type).cuda()
logger.info(model)
# optimizer = torch.optim.SGD(model.parameters(), args.base_lr, momentum=args.momentum, weight_decay=args.weight_decay)
# newly introduced layer with lr x10
if args.net_type == 0:
optimizer = torch.optim.SGD(
[{'params': model.layer0.parameters()},
{'params': model.layer1.parameters()},
{'params': model.layer2.parameters()},
{'params': model.layer3.parameters()},
{'params': model.layer4.parameters()},
{'params': model.ppm.parameters(), 'lr': args.base_lr * 10},
{'params': model.conv6.parameters(), 'lr': args.base_lr * 10},
{'params': model.conv1_1x1.parameters(), 'lr': args.base_lr * 10},
{'params': model.cls.parameters(), 'lr': args.base_lr * 10},
{'params': model.aux.parameters(), 'lr': args.base_lr * 10}],
lr=args.base_lr, momentum=args.momentum, weight_decay=args.weight_decay)
elif args.net_type == 1:
optimizer = torch.optim.SGD(
[{'params': model.layer0.parameters()},
{'params': model.layer1.parameters()},
{'params': model.layer2.parameters()},
{'params': model.layer3.parameters()},
{'params': model.layer4.parameters()},
{'params': model.layer4_p.parameters()},
{'params': model.ppm.parameters(), 'lr': args.base_lr * 10},
{'params': model.ppm_p.parameters(), 'lr': args.base_lr * 10},
{'params': model.cls.parameters(), 'lr': args.base_lr * 10},
{'params': model.cls_p.parameters(), 'lr': args.base_lr * 10},
{'params': model.aux.parameters(), 'lr': args.base_lr * 10}],
lr=args.base_lr, momentum=args.momentum, weight_decay=args.weight_decay)
elif args.net_type == 2:
optimizer = torch.optim.SGD(
[{'params': model.layer0.parameters()},
{'params': model.layer1.parameters()},
{'params': model.layer2.parameters()},
{'params': model.layer3.parameters()},
{'params': model.layer4.parameters()},
{'params': model.layer4_p.parameters()},
{'params': model.ppm.parameters(), 'lr': args.base_lr * 10},
{'params': model.ppm_p.parameters(), 'lr': args.base_lr * 10},
{'params': model.cls.parameters(), 'lr': args.base_lr * 10},
{'params': model.cls_p.parameters(), 'lr': args.base_lr * 10},
{'params': model.att.parameters(), 'lr': args.base_lr * 10},
{'params': model.att_p.parameters(), 'lr': args.base_lr * 10},
{'params': model.aux.parameters(), 'lr': args.base_lr * 10}],
lr=args.base_lr, momentum=args.momentum, weight_decay=args.weight_decay)
elif args.net_type == 3:
optimizer = torch.optim.SGD(
[{'params': model.layer0.parameters()},
{'params': model.layer1.parameters()},
{'params': model.layer2.parameters()},
{'params': model.layer3.parameters()},
{'params': model.layer4.parameters()},
{'params': model.layer4_p.parameters()},
{'params': model.ppm.parameters(), 'lr': args.base_lr * 10},
{'params': model.ppm_p.parameters(), 'lr': args.base_lr * 10},
{'params': model.cls.parameters(), 'lr': args.base_lr * 10},
{'params': model.cls_p.parameters(), 'lr': args.base_lr * 10},
{'params': model.att.parameters(), 'lr': args.base_lr * 10},
{'params': model.aux.parameters(), 'lr': args.base_lr * 10}],
lr=args.base_lr, momentum=args.momentum, weight_decay=args.weight_decay)
fcw = V11RFCN()
fcw_model = torch.load('checkpoint_e8.pth')['state_dict']
fcw_dict = fcw.state_dict()
pretrained_fcw = {k: v for k, v in fcw_model.items() if k in fcw_dict}
fcw_dict.update(pretrained_fcw)
fcw.load_state_dict(fcw_dict)
#fcw = DistModule(fcw)
#print(fcw)
fcw = fcw.cuda()
#model = torch.nn.DataParallel(model).cuda()
model = DistModule(model)
#if args.syncbn:
# from lib.syncbn import patch_replication_callback
# patch_replication_callback(model)
cudnn.enabled = True
cudnn.benchmark = True
criterion = nn.NLLLoss(ignore_index=args.ignore_label).cuda()
if args.weight:
def map_func(storage, location):
return storage.cuda()
if os.path.isfile(args.weight):
logger.info("=> loading weight '{}'".format(args.weight))
checkpoint = torch.load(args.weight, map_location=map_func)['state_dict']
checkpoint = {k: v for k, v in checkpoint.items() if 'ppm' not in k}
model_dict = model.state_dict()
model_dict.update(checkpoint)
model.load_state_dict(model_dict)
logger.info("=> loaded weight '{}'".format(args.weight))
else:
logger.info("=> no weight found at '{}'".format(args.weight))
if args.resume:
load_state(args.resume, model, optimizer)
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 = transforms.Compose([
transforms.RandScale([args.scale_min, args.scale_max]),
#transforms.RandRotate([args.rotate_min, args.rotate_max], padding=mean, ignore_label=args.ignore_label),
transforms.RandomGaussianBlur(),
transforms.RandomHorizontalFlip(),
transforms.Crop([args.crop_h, args.crop_w], crop_type='rand', padding=mean, ignore_label=args.ignore_label),
transforms.ToTensor(),
transforms.Normalize(mean=mean, std=std)])
train_data = datasets.SegData(split='train', data_root=args.data_root, data_list=args.train_list, transform=train_transform)
train_sampler = DistributedSampler(train_data)
train_loader = torch.utils.data.DataLoader(train_data, batch_size=args.batch_size, shuffle=False, num_workers=args.workers, pin_memory=False, sampler=train_sampler)
if args.evaluate:
val_transform = transforms.Compose([
transforms.Crop([args.crop_h, args.crop_w], crop_type='center', padding=mean, ignore_label=args.ignore_label),
transforms.ToTensor(),
transforms.Normalize(mean=mean, std=std)])
val_data = datasets.SegData(split='val', data_root=args.data_root, data_list=args.val_list, transform=val_transform)
val_loader = torch.utils.data.DataLoader(val_data, batch_size=args.batch_size_val, shuffle=False, num_workers=args.workers, pin_memory=True)
for epoch in range(args.start_epoch, args.epochs + 1):
loss_train, mIoU_train, mAcc_train, allAcc_train = train(train_loader, model, criterion, optimizer, epoch, args.zoom_factor, args.batch_size, args.aux_weight, fcw)
if rank == 0:
writer.add_scalar('loss_train', loss_train, epoch)
writer.add_scalar('mIoU_train', mIoU_train, epoch)
writer.add_scalar('mAcc_train', mAcc_train, epoch)
writer.add_scalar('allAcc_train', allAcc_train, epoch)
# write parameters histogram costs lots of time
# for name, param in model.named_parameters():
# writer.add_histogram(name, param, epoch)
if epoch % args.save_step == 0 and rank == 0:
filename = args.save_path + '/train_epoch_' + str(epoch) + '.pth'
logger.info('Saving checkpoint to: ' + filename)
torch.save({'epoch': epoch, 'state_dict': model.state_dict(), 'optimizer': optimizer.state_dict()}, filename)
#if epoch / args.save_step > 2:
# deletename = args.save_path + '/train_epoch_' + str(epoch - args.save_step*2) + '.pth'
# os.remove(deletename)
if args.evaluate:
loss_val, mIoU_val, mAcc_val, allAcc_val = validate(val_loader, model, criterion, args.classes, args.zoom_factor)
writer.add_scalar('loss_val', loss_val, epoch)
writer.add_scalar('mIoU_val', mIoU_val, epoch)
writer.add_scalar('mAcc_val', mAcc_val, epoch)
writer.add_scalar('allAcc_val', allAcc_val, epoch)
def main():
global args, logger, writer
args = get_parser().parse_args()
logger = get_logger()
writer = SummaryWriter(args.save_path)
# os.environ["CUDA_VISIBLE_DEVICES"] = ','.join(str(x) for x in args.gpu)
if args.dist:
dist_init(args.port, backend=args.backend)
if len(args.gpu) == 1:
args.syncbn = False
logger.info(args)
assert args.classes > 1
assert args.zoom_factor in [1, 2, 4, 8]
assert (args.crop_h - 1) % 8 == 0 and (args.crop_w - 1) % 8 == 0
world_size = 1
rank = 0
if args.dist:
rank = dist.get_rank()
world_size = dist.get_world_size()
if rank == 0:
logger.info('dist:{}'.format(args.dist))
logger.info("=> creating model ...")
logger.info("Classes: {}".format(args.classes))
# rank = dist.get_rank()
if args.bn_group > 1:
args.syncbn = True
bn_sync_stats = True
bn_group_comm = simple_group_split(world_size, rank,
world_size // args.bn_group)
else:
args.syncbn = False
bn_sync_stats = False
bn_group_comm = None
model = PSPNet(layers=args.layers,
classes=args.classes,
zoom_factor=args.zoom_factor,
syncbn=args.syncbn,
group_size=args.bn_group,
group=bn_group_comm,
sync_stats=bn_sync_stats)
if rank == 0:
logger.info(model)
# optimizer = torch.optim.SGD(model.parameters(), args.base_lr, momentum=args.momentum, weight_decay=args.weight_decay)
# newly introduced layer with lr x10
optimizer = torch.optim.SGD([{
'params': model.layer0.parameters()
}, {
'params': model.layer1.parameters()
}, {
'params': model.layer2.parameters()
}, {
'params': model.layer3.parameters()
}, {
'params': model.layer4.parameters()
}, {
'params': model.ppm.parameters(),
'lr': args.base_lr * 10
}, {
'params': model.cls.parameters(),
'lr': args.base_lr * 10
}, {
'params': model.aux.parameters(),
'lr': args.base_lr * 10
}],
lr=args.base_lr,
momentum=args.momentum,
weight_decay=args.weight_decay)
# model = torch.nn.DataParallel(model).cuda()
# if args.syncbn:
# from lib.syncbn import patch_replication_callback
# patch_replication_callback(model)
model = model.cuda()
cudnn.enabled = True
cudnn.benchmark = True
criterion = nn.NLLLoss(ignore_index=args.ignore_label)
if args.weight:
def map_func(storage, location):
return storage.cuda()
if os.path.isfile(args.weight):
logger.info("=> loading weight '{}'".format(args.weight))
checkpoint = torch.load(args.weight, map_location=map_func)
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)
# args.start_epoch = checkpoint['epoch']
# model.load_state_dict(checkpoint['state_dict'])
# optimizer.load_state_dict(checkpoint['optimizer'])
model, optimizer, args.start_epoch = restore_from(
model, optimizer, args.resume)
logger.info("=> loaded checkpoint '{}' (epoch {})".format(
args.resume, args.start_epoch))
else:
logger.info("=> no checkpoint found at '{}'".format(args.resume))
if args.dist:
broadcast_params(model)
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 = transforms.Compose([
transforms.RandScale([args.scale_min, args.scale_max]),
transforms.RandRotate([args.rotate_min, args.rotate_max],
padding=mean,
ignore_label=args.ignore_label),
transforms.RandomGaussianBlur(),
transforms.RandomHorizontalFlip(),
transforms.Crop([args.crop_h, args.crop_w],
crop_type='rand',
padding=mean,
ignore_label=args.ignore_label),
transforms.ToTensor(),
transforms.Normalize(mean=mean, std=std)
])
train_data = datasets.SegData(split='train',
data_root=args.data_root,
data_list=args.train_list,
transform=train_transform)
train_sampler = None
if args.dist:
train_sampler = DistributedSampler(train_data)
train_loader = torch.utils.data.DataLoader(
train_data,
batch_size=args.batch_size,
shuffle=False if train_sampler else True,
num_workers=args.workers,
pin_memory=False,
sampler=train_sampler)
if args.evaluate:
val_transform = transforms.Compose([
transforms.Crop([args.crop_h, args.crop_w],
crop_type='center',
padding=mean,
ignore_label=args.ignore_label),
transforms.ToTensor(),
transforms.Normalize(mean=mean, std=std)
])
val_data = datasets.SegData(split='val',
data_root=args.data_root,
data_list=args.val_list,
transform=val_transform)
val_sampler = None
if args.dist:
val_sampler = DistributedSampler(val_data)
val_loader = torch.utils.data.DataLoader(
val_data,
batch_size=args.batch_size_val,
shuffle=False,
num_workers=args.workers,
pin_memory=False,
sampler=val_sampler)
for epoch in range(args.start_epoch, args.epochs + 1):
loss_train, mIoU_train, mAcc_train, allAcc_train = train(
train_loader, model, criterion, optimizer, epoch, args.zoom_factor,
args.batch_size, args.aux_weight)
writer.add_scalar('loss_train', loss_train.cpu().numpy(), epoch)
writer.add_scalar('mIoU_train', mIoU_train, epoch)
writer.add_scalar('mAcc_train', mAcc_train, epoch)
writer.add_scalar('allAcc_train', allAcc_train, epoch)
# write parameters histogram costs lots of time
# for name, param in model.named_parameters():
# writer.add_histogram(name, param, epoch)
if args.evaluate and rank == 0:
loss_val, mIoU_val, mAcc_val, allAcc_val = validate(
val_loader, model, criterion, args.classes, args.zoom_factor)
writer.add_scalar('loss_val', loss_val.cpu().numpy(), epoch)
writer.add_scalar('mIoU_val', mIoU_val, epoch)
writer.add_scalar('mAcc_val', mAcc_val, epoch)
writer.add_scalar('allAcc_val', allAcc_val, epoch)
if epoch % args.save_step == 0 and (rank == 0):
filename = args.save_path + '/train_epoch_' + str(epoch) + '.pth'
logger.info('Saving checkpoint to: ' + filename)
torch.save(
{
'epoch': epoch,
'state_dict': model.cpu().state_dict(),
'optimizer': optimizer.state_dict()
}, filename)