def __init__(self, args):
if args.se_loss:
args.checkname = args.checkname + "_se"
self.args = args
# data transforms
input_transform = transform.Compose([
transform.ToTensor(),
transform.Normalize([.485, .456, .406], [.229, .224, .225])])
# dataset
data_kwargs = {'transform': input_transform, 'base_size': args.base_size,
'crop_size': args.crop_size}
trainset = get_segmentation_dataset(args.dataset, split='train', mode='train',
**data_kwargs)
testset = get_segmentation_dataset(args.dataset, split='val', mode ='val',
**data_kwargs)
# dataloader
kwargs = {'num_workers': args.workers, 'pin_memory': False} \
if args.cuda else {}
self.trainloader = data.DataLoader(trainset, batch_size=args.batch_size,
drop_last=True, shuffle=True, **kwargs)
self.valloader = data.DataLoader(testset, batch_size=args.batch_size,
drop_last=False, shuffle=False, **kwargs)
self.nclass = trainset.num_class
# model
model = get_segmentation_model(args.model, dataset=args.dataset,
backbone = args.backbone, aux = args.aux,
se_loss = args.se_loss, norm_layer = BatchNorm2d,
base_size=args.base_size, crop_size=args.crop_size)
print(model)
# count parameter number
pytorch_total_params = sum(p.numel() for p in model.parameters())
print("Total number of parameters: %d"%pytorch_total_params)
# optimizer using different LR
params_list = [{'params': model.pretrained.parameters(), 'lr': args.lr},]
if hasattr(model, 'head'):
if args.diflr:
params_list.append({'params': model.head.parameters(), 'lr': args.lr*10})
else:
params_list.append({'params': model.head.parameters(), 'lr': args.lr})
if hasattr(model, 'auxlayer'):
if args.diflr:
params_list.append({'params': model.auxlayer.parameters(), 'lr': args.lr*10})
else:
params_list.append({'params': model.auxlayer.parameters(), 'lr': args.lr})
optimizer = torch.optim.SGD(params_list,
lr=args.lr,
momentum=args.momentum,
weight_decay=args.weight_decay)
#optimizer = torch.optim.ASGD(params_list,
# lr=args.lr,
# weight_decay=args.weight_decay)
# criterions
self.criterion = SegmentationLosses(se_loss=args.se_loss, aux=args.aux,
nclass=self.nclass)
self.model, self.optimizer = model, optimizer
# using cuda
if args.cuda:
self.model = DataParallelModel(self.model).cuda()
self.criterion = DataParallelCriterion(self.criterion).cuda()
# resuming checkpoint
if args.resume is not None and len(args.resume)>0:
if not os.path.isfile(args.resume):
raise RuntimeError("=> no checkpoint found at '{}'" .format(args.resume))
checkpoint = torch.load(args.resume)
args.start_epoch = checkpoint['epoch']
if args.cuda:
# load weights for the same model
# self.model.module.load_state_dict(checkpoint['state_dict'])
# model and checkpoint have different strucutures
pretrained_dict = checkpoint['state_dict']
model_dict = self.model.module.state_dict()
for name, param in pretrained_dict.items():
if name not in model_dict:
continue
if isinstance(param, Parameter):
# backwards compatibility for serialized parameters
param = param.data
model_dict[name].copy_(param)
else:
self.model.load_state_dict(checkpoint['state_dict'])
if not args.ft:
self.optimizer.load_state_dict(checkpoint['optimizer'])
self.best_pred = checkpoint['best_pred']
print("=> loaded checkpoint '{}' (epoch {})"
.format(args.resume, checkpoint['epoch']))
# clear start epoch if fine-tuning
if args.ft:
args.start_epoch = 0
# lr scheduler
self.scheduler = utils.LR_Scheduler(args.lr_scheduler, args.lr,
args.epochs, len(self.trainloader),lr_step=args.lr_step)
self.best_pred = 0.0
def __init__(self, args):
self.args = args
# data transforms
input_transform = transform.Compose([
transform.ToTensor(),
transform.Normalize([.485, .456, .406], [.229, .224, .225])
])
# dataset
data_kwargs = {
'transform': input_transform,
'base_size': args.base_size,
'crop_size': args.crop_size
}
trainset = get_segmentation_dataset(args.dataset,
split=args.train_split,
mode='train',
**data_kwargs)
testset = get_segmentation_dataset(args.dataset,
split='val',
mode='val',
**data_kwargs)
# dataloader
kwargs = {'num_workers': args.workers, 'pin_memory': True} \
if args.cuda else {}
self.trainloader = data.DataLoader(trainset,
batch_size=args.batch_size,
drop_last=True,
shuffle=True,
**kwargs)
self.valloader = data.DataLoader(testset,
batch_size=args.batch_size,
drop_last=False,
shuffle=False,
**kwargs)
self.nclass = trainset.num_class
# model
model = get_segmentation_model(
args.model,
dataset=args.dataset,
backbone=args.backbone,
dilated=args.dilated,
lateral=args.lateral,
jpu=args.jpu,
aux=args.aux,
se_loss=args.se_loss,
norm_layer=torch.nn.BatchNorm2d, ## BatchNorm2d
base_size=args.base_size,
crop_size=args.crop_size)
print(model)
# optimizer using different LR
params_list = [
{
'params': model.pretrained.parameters(),
'lr': args.lr
},
]
if hasattr(model, 'jpu'):
params_list.append({
'params': model.jpu.parameters(),
'lr': args.lr * 10
})
if hasattr(model, 'head'):
params_list.append({
'params': model.head.parameters(),
'lr': args.lr * 10
})
if hasattr(model, 'auxlayer'):
params_list.append({
'params': model.auxlayer.parameters(),
'lr': args.lr * 10
})
optimizer = torch.optim.SGD(params_list,
lr=args.lr,
momentum=args.momentum,
weight_decay=args.weight_decay)
# criterions
self.criterion = SegmentationLosses(se_loss=args.se_loss,
aux=args.aux,
nclass=self.nclass,
se_weight=args.se_weight,
aux_weight=args.aux_weight)
self.model, self.optimizer = model, optimizer
# using cuda
if args.cuda:
self.model = DataParallelModel(self.model).cuda()
self.criterion = DataParallelCriterion(self.criterion).cuda()
# resuming checkpoint
if args.resume is not None:
if not os.path.isfile(args.resume):
raise RuntimeError("=> no checkpoint found at '{}'".format(
args.resume))
checkpoint = torch.load(args.resume)
args.start_epoch = checkpoint['epoch']
if args.cuda:
self.model.module.load_state_dict(checkpoint['state_dict'])
else:
self.model.load_state_dict(checkpoint['state_dict'])
if not args.ft:
self.optimizer.load_state_dict(checkpoint['optimizer'])
self.best_pred = checkpoint['best_pred']
print("=> loaded checkpoint '{}' (epoch {})".format(
args.resume, checkpoint['epoch']))
# clear start epoch if fine-tuning
if args.ft:
args.start_epoch = 0
# lr scheduler
self.scheduler = utils.LR_Scheduler(args.lr_scheduler, args.lr,
args.epochs, len(self.trainloader))
self.best_pred = 0.0
class Trainer():
def __init__(self, args):
self.args = args
# data transforms
input_transform = transform.Compose([
transform.ToTensor(),
transform.Normalize([.485, .456, .406], [.229, .224, .225])
])
# dataset
data_kwargs = {
'transform': input_transform,
'base_size': args.base_size,
'crop_size': args.crop_size
}
trainset = get_dataset(args.dataset,
split=args.train_split,
mode='train',
**data_kwargs)
valset = get_dataset(
args.dataset,
split='val',
mode='ms_val' if args.multi_scale_eval else 'fast_val',
**data_kwargs)
# dataloader
kwargs = {'num_workers': args.workers, 'pin_memory': True}
self.trainloader = data.DataLoader(trainset,
batch_size=args.batch_size,
drop_last=True,
shuffle=True,
**kwargs)
if self.args.multi_scale_eval:
kwargs['collate_fn'] = test_batchify_fn
self.valloader = data.DataLoader(valset,
batch_size=args.test_batch_size,
drop_last=False,
shuffle=False,
**kwargs)
self.nclass = trainset.num_class
# model
if args.norm_layer == 'bn':
norm_layer = BatchNorm2d
elif args.norm_layer == 'sync_bn':
assert args.multi_gpu, "SyncBatchNorm can only be used when multi GPUs are available!"
norm_layer = SyncBatchNorm
else:
raise ValueError('Invalid norm_layer {}'.format(args.norm_layer))
model = get_segmentation_model(
args.model,
dataset=args.dataset,
backbone=args.backbone,
aux=args.aux,
se_loss=args.se_loss,
norm_layer=norm_layer,
base_size=args.base_size,
crop_size=args.crop_size,
multi_grid=True,
multi_dilation=[2, 4, 8],
only_pam=True,
)
print(model)
# optimizer using different LR
params_list = [
{
'params': model.pretrained.parameters(),
'lr': args.lr
},
]
if hasattr(model, 'head'):
params_list.append({
'params': model.head.parameters(),
'lr': args.lr
})
if hasattr(model, 'auxlayer'):
params_list.append({
'params': model.auxlayer.parameters(),
'lr': args.lr
})
optimizer = torch.optim.SGD(params_list,
lr=args.lr,
momentum=args.momentum,
weight_decay=args.weight_decay)
# criterions
self.criterion = SegmentationMultiLosses()
self.model, self.optimizer = model, optimizer
# using cuda
if args.multi_gpu:
self.model = DataParallelModel(self.model).cuda()
self.criterion = DataParallelCriterion(self.criterion).cuda()
else:
self.model = self.model.cuda()
self.criterion = self.criterion.cuda()
self.single_device_model = self.model.module if self.args.multi_gpu else self.model
# resuming checkpoint
if args.resume is not None:
if not os.path.isfile(args.resume):
raise RuntimeError("=> no checkpoint found at '{}'".format(
args.resume))
checkpoint = torch.load(args.resume)
args.start_epoch = checkpoint['epoch']
self.single_device_model.load_state_dict(checkpoint['state_dict'])
if not args.ft and not (args.only_val or args.only_vis
or args.only_infer):
self.optimizer.load_state_dict(checkpoint['optimizer'])
self.best_pred = checkpoint['best_pred']
print("=> loaded checkpoint '{}' (epoch {}), best_pred {}".format(
args.resume, checkpoint['epoch'], checkpoint['best_pred']))
# clear start epoch if fine-tuning
if args.ft:
args.start_epoch = 0
# lr scheduler
self.lr_scheduler = torch.optim.lr_scheduler.ExponentialLR(
optimizer, 0.6)
self.best_pred = 0.0
def save_ckpt(self, epoch, score):
is_best = False
if score >= self.best_pred:
is_best = True
self.best_pred = score
utils.save_checkpoint(
{
'epoch': epoch + 1,
'state_dict': self.single_device_model.state_dict(),
'optimizer': self.optimizer.state_dict(),
'best_pred': self.best_pred,
}, self.args, is_best)
def training(self, epoch):
train_loss = 0.0
self.model.train()
self.lr_scheduler.step()
tbar = tqdm(self.trainloader, miniters=20)
for i, (image, target) in enumerate(tbar):
if not self.args.multi_gpu:
image = image.cuda()
target = target.cuda()
self.optimizer.zero_grad()
if torch_ver == "0.3":
image = Variable(image)
target = Variable(target)
outputs = self.model(image)
if self.args.multi_gpu:
loss = self.criterion(outputs, target)
else:
loss = self.criterion(*(outputs + (target, )))
loss.backward()
self.optimizer.step()
train_loss += loss.item()
ep_log = 'ep {}'.format(epoch + 1)
lr_log = 'lr ' + '{:.6f}'.format(
self.optimizer.param_groups[0]['lr']).rstrip('0')
loss_log = 'loss {:.3f}'.format(train_loss / (i + 1))
tbar.set_description(', '.join([ep_log, lr_log, loss_log]))
def validation(self, epoch):
def _get_pred(batch_im):
with torch.no_grad():
# metric.update also accepts list, so no need to gather results from multi gpus
if self.args.multi_scale_eval:
assert len(batch_im) <= torch.cuda.device_count(
), "Multi-scale testing only allows batch size <= number of GPUs"
scattered_pred = self.ms_evaluator.parallel_forward(
batch_im)
else:
outputs = self.model(batch_im)
scattered_pred = [
out[0] for out in outputs
] if self.args.multi_gpu else [outputs[0]]
return scattered_pred
# Lazy creation
if not hasattr(self, 'ms_evaluator'):
self.ms_evaluator = MultiEvalModule(self.single_device_model,
self.nclass,
scales=self.args.eval_scales,
crop=self.args.crop_eval)
self.metric = utils.SegmentationMetric(self.nclass)
self.model.eval()
tbar = tqdm(self.valloader, desc='\r')
for i, (batch_im, target) in enumerate(tbar):
# No need to put target to GPU, since the metrics are calculated by numpy.
# And no need to put data to GPU manually if we use data parallel.
if not self.args.multi_gpu and not isinstance(
batch_im, (list, tuple)):
batch_im = batch_im.cuda()
scattered_pred = _get_pred(batch_im)
scattered_target = []
ind = 0
for p in scattered_pred:
target_tmp = target[ind:ind + len(p)]
# Multi-scale testing. In fact, len(target_tmp) == 1
if isinstance(target_tmp, (list, tuple)):
assert len(target_tmp) == 1
target_tmp = torch.stack(target_tmp)
scattered_target.append(target_tmp)
ind += len(p)
self.metric.update(scattered_target, scattered_pred)
pixAcc, mIoU = self.metric.get()
tbar.set_description('ep {}, pixAcc: {:.4f}, mIoU: {:.4f}'.format(
epoch + 1, pixAcc, mIoU))
return self.metric.get()
def visualize(self, epoch):
if (self.args.dir_of_im_to_vis
== 'None') and (self.args.im_list_file_to_vis == 'None'):
return
if not hasattr(self, 'vis_im_paths'):
if self.args.dir_of_im_to_vis != 'None':
print('=> Visualize Dir {}'.format(self.args.dir_of_im_to_vis))
im_paths = list(
walkdir(self.args.dir_of_im_to_vis, exts=['.jpg', '.png']))
else:
print('=> Visualize Image List {}'.format(
self.args.im_list_file_to_vis))
im_paths = read_lines(self.args.im_list_file_to_vis)
print('=> Save Dir {}'.format(self.args.vis_save_dir))
im_paths = sorted(im_paths)
# np.random.RandomState(seed=1).shuffle(im_paths)
self.vis_im_paths = im_paths[:self.args.max_num_vis]
cfg = {
'save_path':
os.path.join(self.args.vis_save_dir,
'vis_epoch{}.png'.format(epoch)),
'multi_scale':
self.args.multi_scale_eval,
'crop':
self.args.crop_eval,
'num_class':
self.nclass,
'scales':
self.args.eval_scales,
'base_size':
self.args.base_size,
}
vis_im_list(self.single_device_model, self.vis_im_paths, cfg)
def infer_and_save(self, infer_dir, infer_save_dir):
print('=> Infer Dir {}'.format(infer_dir))
print('=> Save Dir {}'.format(infer_save_dir))
sub_im_paths = list(
walkdir(infer_dir, exts=['.jpg', '.png'], sub_path=True))
im_paths = [os.path.join(infer_dir, p) for p in sub_im_paths]
# NOTE: Don't save result as JPEG, since it causes aliasing.
save_paths = [
os.path.join(infer_save_dir, p.replace('.jpg', '.png'))
for p in sub_im_paths
]
cfg = {
'multi_scale': self.args.multi_scale_eval,
'crop': self.args.crop_eval,
'num_class': self.nclass,
'scales': self.args.eval_scales,
'base_size': self.args.base_size,
}
infer_and_save_im_list(self.single_device_model, im_paths, save_paths,
cfg)
def __init__(self, args):
self.args = args
# data transforms
input_transform = transform.Compose([
transform.ToTensor(),
transform.Normalize([.485, .456, .406], [.229, .224, .225])
])
# dataset
data_kwargs = {
'transform': input_transform,
'base_size': args.base_size,
'crop_size': args.crop_size
}
trainset = get_dataset(args.dataset,
split=args.train_split,
mode='train',
**data_kwargs)
valset = get_dataset(
args.dataset,
split='val',
mode='ms_val' if args.multi_scale_eval else 'fast_val',
**data_kwargs)
# dataloader
kwargs = {'num_workers': args.workers, 'pin_memory': True}
self.trainloader = data.DataLoader(trainset,
batch_size=args.batch_size,
drop_last=True,
shuffle=True,
**kwargs)
if self.args.multi_scale_eval:
kwargs['collate_fn'] = test_batchify_fn
self.valloader = data.DataLoader(valset,
batch_size=args.test_batch_size,
drop_last=False,
shuffle=False,
**kwargs)
self.nclass = trainset.num_class
# model
if args.norm_layer == 'bn':
norm_layer = BatchNorm2d
elif args.norm_layer == 'sync_bn':
assert args.multi_gpu, "SyncBatchNorm can only be used when multi GPUs are available!"
norm_layer = SyncBatchNorm
else:
raise ValueError('Invalid norm_layer {}'.format(args.norm_layer))
model = get_segmentation_model(
args.model,
dataset=args.dataset,
backbone=args.backbone,
aux=args.aux,
se_loss=args.se_loss,
norm_layer=norm_layer,
base_size=args.base_size,
crop_size=args.crop_size,
multi_grid=True,
multi_dilation=[2, 4, 8],
only_pam=True,
)
print(model)
# optimizer using different LR
params_list = [
{
'params': model.pretrained.parameters(),
'lr': args.lr
},
]
if hasattr(model, 'head'):
params_list.append({
'params': model.head.parameters(),
'lr': args.lr
})
if hasattr(model, 'auxlayer'):
params_list.append({
'params': model.auxlayer.parameters(),
'lr': args.lr
})
optimizer = torch.optim.SGD(params_list,
lr=args.lr,
momentum=args.momentum,
weight_decay=args.weight_decay)
# criterions
self.criterion = SegmentationMultiLosses()
self.model, self.optimizer = model, optimizer
# using cuda
if args.multi_gpu:
self.model = DataParallelModel(self.model).cuda()
self.criterion = DataParallelCriterion(self.criterion).cuda()
else:
self.model = self.model.cuda()
self.criterion = self.criterion.cuda()
self.single_device_model = self.model.module if self.args.multi_gpu else self.model
# resuming checkpoint
if args.resume is not None:
if not os.path.isfile(args.resume):
raise RuntimeError("=> no checkpoint found at '{}'".format(
args.resume))
checkpoint = torch.load(args.resume)
args.start_epoch = checkpoint['epoch']
self.single_device_model.load_state_dict(checkpoint['state_dict'])
if not args.ft and not (args.only_val or args.only_vis
or args.only_infer):
self.optimizer.load_state_dict(checkpoint['optimizer'])
self.best_pred = checkpoint['best_pred']
print("=> loaded checkpoint '{}' (epoch {}), best_pred {}".format(
args.resume, checkpoint['epoch'], checkpoint['best_pred']))
# clear start epoch if fine-tuning
if args.ft:
args.start_epoch = 0
# lr scheduler
self.lr_scheduler = torch.optim.lr_scheduler.ExponentialLR(
optimizer, 0.6)
self.best_pred = 0.0
def __init__(self, args):
self.args = args
args.log_name = str(args.checkname)
self.logger = utils.create_logger(args.log_root, args.log_name)
# data transforms
input_transform = transform.Compose([
transform.ToTensor(),
transform.Normalize([.485, .456, .406], [.229, .224, .225])
])
# dataset
data_kwargs = {
'transform': input_transform,
'base_size': args.base_size,
'crop_size': args.crop_size,
'logger': self.logger,
'scale': args.scale
}
trainset = get_segmentation_dataset(args.dataset,
split='train',
mode='train',
**data_kwargs)
testset = get_segmentation_dataset(args.dataset,
split='val',
mode='val',
**data_kwargs)
# dataloader
kwargs = {'num_workers': args.workers, 'pin_memory': True} \
if args.cuda else {}
self.trainloader = data.DataLoader(trainset,
batch_size=args.batch_size,
drop_last=True,
shuffle=True,
**kwargs)
self.valloader = data.DataLoader(testset,
batch_size=args.batch_size,
drop_last=False,
shuffle=False,
**kwargs)
self.nclass = trainset.num_class
# model
model = get_segmentation_model(args.model,
dataset=args.dataset,
backbone=args.backbone,
aux=args.aux,
se_loss=args.se_loss,
norm_layer=BatchNorm2d,
base_size=args.base_size,
crop_size=args.crop_size,
multi_grid=args.multi_grid,
multi_dilation=args.multi_dilation)
#print(model)
self.logger.info(model)
# optimizer using different LR
params_list = [
{
'params': model.pretrained.parameters(),
'lr': args.lr
},
]
if hasattr(model, 'head'):
params_list.append({
'params': model.head.parameters(),
'lr': args.lr * 10
})
if hasattr(model, 'auxlayer'):
params_list.append({
'params': model.auxlayer.parameters(),
'lr': args.lr * 10
})
cityscape_weight = torch.FloatTensor([
0.8373, 0.918, 0.866, 1.0345, 1.0166, 0.9969, 0.9754, 1.0489,
0.8786, 1.0023, 0.9539, 0.9843, 1.1116, 0.9037, 1.0865, 1.0955,
1.0865, 1.1529, 1.0507
])
optimizer = torch.optim.SGD(params_list,
lr=args.lr,
momentum=args.momentum,
weight_decay=args.weight_decay)
#weight for class imbalance
# self.criterion = SegmentationMultiLosses(nclass=self.nclass, weight=cityscape_weight)
self.criterion = SegmentationMultiLosses(nclass=self.nclass)
#self.criterion = SegmentationLosses(se_loss=args.se_loss, aux=args.aux,nclass=self.nclass)
self.model, self.optimizer = model, optimizer
# using cuda
if args.cuda:
self.model = DataParallelModel(self.model).cuda()
self.criterion = DataParallelCriterion(self.criterion).cuda()
# finetune from a trained model
if args.ft:
args.start_epoch = 0
checkpoint = torch.load(args.ft_resume)
if args.cuda:
self.model.module.load_state_dict(checkpoint['state_dict'],
strict=False)
else:
self.model.load_state_dict(checkpoint['state_dict'],
strict=False)
self.logger.info("=> loaded checkpoint '{}' (epoch {})".format(
args.ft_resume, checkpoint['epoch']))
# resuming checkpoint
if args.resume:
if not os.path.isfile(args.resume):
raise RuntimeError("=> no checkpoint found at '{}'".format(
args.resume))
checkpoint = torch.load(args.resume)
args.start_epoch = checkpoint['epoch']
if args.cuda:
self.model.module.load_state_dict(checkpoint['state_dict'])
else:
self.model.load_state_dict(checkpoint['state_dict'])
if not args.ft:
self.optimizer.load_state_dict(checkpoint['optimizer'])
self.best_pred = checkpoint['best_pred']
self.logger.info("=> loaded checkpoint '{}' (epoch {})".format(
args.resume, checkpoint['epoch']))
# lr scheduler
self.scheduler = utils.LR_Scheduler(args.lr_scheduler,
args.lr,
args.epochs,
len(self.trainloader),
logger=self.logger,
lr_step=args.lr_step)
self.best_pred = 0.0
def __init__(self, args):
self.args = args
args.log_name = str(args.checkname)
args.log_root = os.path.join(args.dataset, args.log_root) # dataset/log/
self.logger = utils.create_logger(args.log_root, args.log_name)
# data transforms
input_transform = transform.Compose([
transform.ToTensor(),
transform.Normalize([.485, .456, .406], [.229, .224, .225])])
# dataset
data_kwargs = {'transform': input_transform, 'base_size': args.base_size,
'crop_size': args.crop_size, 'logger': self.logger,
'scale': args.scale}
trainset = get_segmentation_dataset(args.dataset, split='trainval', mode='trainval',
**data_kwargs)
testset = get_segmentation_dataset(args.dataset, split='val', mode='val', # crop fixed size as model input
**data_kwargs)
# dataloader
kwargs = {'num_workers': args.workers, 'pin_memory': True} \
if args.cuda else {}
self.trainloader = data.DataLoader(trainset, batch_size=args.batch_size,
drop_last=True, shuffle=True, **kwargs)
self.valloader = data.DataLoader(testset, batch_size=args.batch_size,
drop_last=False, shuffle=False, **kwargs)
self.nclass = trainset.num_class
# model
model = get_segmentation_model(args.model, dataset=args.dataset,
backbone=args.backbone,
norm_layer=BatchNorm2d,
base_size=args.base_size, crop_size=args.crop_size,
)
#print(model)
self.logger.info(model)
# optimizer using different LR
params_list = [{'params': model.pretrained.parameters(), 'lr': args.lr},]
if hasattr(model, 'head'):
print("this model has object, head")
params_list.append({'params': model.head.parameters(), 'lr': args.lr*10})
optimizer = torch.optim.SGD(params_list,
lr=args.lr,
momentum=args.momentum,
weight_decay=args.weight_decay)
self.criterion = SegmentationLosses(nclass=self.nclass)
self.model, self.optimizer = model, optimizer
# using cuda
if args.cuda:
self.model = DataParallelModel(self.model).cuda()
self.criterion = DataParallelCriterion(self.criterion).cuda()
# resuming checkpoint
if args.resume:
if not os.path.isfile(args.resume):
raise RuntimeError("=> no checkpoint found at '{}'" .format(args.resume))
checkpoint = torch.load(args.resume)
args.start_epoch = checkpoint['epoch']
if args.cuda:
self.model.module.load_state_dict(checkpoint['state_dict'])
else:
self.model.load_state_dict(checkpoint['state_dict'])
if not args.ft:
self.optimizer.load_state_dict(checkpoint['optimizer'])
self.best_pred = checkpoint['best_pred']
self.logger.info("=> loaded checkpoint '{}' (epoch {})".format(args.resume, checkpoint['epoch']))
# lr scheduler
self.scheduler = utils.LR_Scheduler(args.lr_scheduler, args.lr,
args.epochs, len(self.trainloader), logger=self.logger,
lr_step=args.lr_step)
self.best_pred = 0.0
def main(args):
writer = SummaryWriter(log_dir=args.tensorboard_log_dir)
w, h = map(int, args.input_size.split(','))
w_target, h_target = map(int, args.input_size_target.split(','))
joint_transform = joint_transforms.Compose([
joint_transforms.FreeScale((h, w)),
joint_transforms.RandomHorizontallyFlip(),
])
normalize = ((0.485, 0.456, 0.406), (0.229, 0.224, 0.225))
input_transform = standard_transforms.Compose([
standard_transforms.ToTensor(),
standard_transforms.Normalize(*normalize),
])
target_transform = extended_transforms.MaskToTensor()
restore_transform = standard_transforms.ToPILImage()
if '5' in args.data_dir:
dataset = GTA5DataSetLMDB(
args.data_dir, args.data_list,
joint_transform=joint_transform,
transform=input_transform, target_transform=target_transform,
)
else:
dataset = CityscapesDataSetLMDB(
args.data_dir, args.data_list,
joint_transform=joint_transform,
transform=input_transform, target_transform=target_transform,
)
loader = data.DataLoader(
dataset, batch_size=args.batch_size,
shuffle=True, num_workers=args.num_workers, pin_memory=True
)
val_dataset = CityscapesDataSetLMDB(
args.data_dir_target, args.data_list_target,
# joint_transform=joint_transform,
transform=input_transform, target_transform=target_transform
)
val_loader = data.DataLoader(
val_dataset, batch_size=args.batch_size,
shuffle=False, num_workers=args.num_workers, pin_memory=True
)
upsample = nn.Upsample(size=(h_target, w_target),
mode='bilinear', align_corners=True)
net = PSP(
nclass = args.n_classes, backbone='resnet101',
root=args.model_path_prefix, norm_layer=BatchNorm2d,
)
params_list = [
{'params': net.pretrained.parameters(), 'lr': args.learning_rate},
{'params': net.head.parameters(), 'lr': args.learning_rate*10},
{'params': net.auxlayer.parameters(), 'lr': args.learning_rate*10},
]
optimizer = torch.optim.SGD(params_list,
lr=args.learning_rate,
momentum=args.momentum,
weight_decay=args.weight_decay)
criterion = SegmentationLosses(nclass=args.n_classes, aux=True, ignore_index=255)
# criterion = SegmentationMultiLosses(nclass=args.n_classes, ignore_index=255)
net = DataParallelModel(net).cuda()
criterion = DataParallelCriterion(criterion).cuda()
logger = utils.create_logger(args.tensorboard_log_dir, 'PSP_train')
scheduler = utils.LR_Scheduler(args.lr_scheduler, args.learning_rate,
args.num_epoch, len(loader), logger=logger,
lr_step=args.lr_step)
net_eval = Eval(net)
num_batches = len(loader)
best_pred = 0.0
for epoch in range(args.num_epoch):
loss_rec = AverageMeter()
data_time_rec = AverageMeter()
batch_time_rec = AverageMeter()
tem_time = time.time()
for batch_index, batch_data in enumerate(loader):
scheduler(optimizer, batch_index, epoch, best_pred)
show_fig = (batch_index+1) % args.show_img_freq == 0
iteration = batch_index+1+epoch*num_batches
net.train()
img, label, name = batch_data
img = img.cuda()
label_cuda = label.cuda()
data_time_rec.update(time.time()-tem_time)
output = net(img)
loss = criterion(output, label_cuda)
optimizer.zero_grad()
loss.backward()
optimizer.step()
loss_rec.update(loss.item())
writer.add_scalar('A_seg_loss', loss.item(), iteration)
batch_time_rec.update(time.time()-tem_time)
tem_time = time.time()
if (batch_index+1) % args.print_freq == 0:
print(
f'Epoch [{epoch+1:d}/{args.num_epoch:d}][{batch_index+1:d}/{num_batches:d}]\t'
f'Time: {batch_time_rec.avg:.2f} '
f'Data: {data_time_rec.avg:.2f} '
f'Loss: {loss_rec.avg:.2f}'
)
# if show_fig:
# # base_lr = optimizer.param_groups[0]["lr"]
# output = torch.argmax(output[0][0], dim=1).detach()[0, ...].cpu()
# # fig, axes = plt.subplots(2, 1, figsize=(12, 14))
# # axes = axes.flat
# # axes[0].imshow(colorize_mask(output.numpy()))
# # axes[0].set_title(name[0])
# # axes[1].imshow(colorize_mask(label[0, ...].numpy()))
# # axes[1].set_title(f'seg_true_{base_lr:.6f}')
# # writer.add_figure('A_seg', fig, iteration)
# output_mask = np.asarray(colorize_mask(output.numpy()))
# label = np.asarray(colorize_mask(label[0,...].numpy()))
# image_out = np.concatenate([output_mask, label])
# writer.add_image('A_seg', image_out, iteration)
mean_iu = test_miou(net_eval, val_loader, upsample,
'./style_seg/dataset/info.json')
torch.save(
net.module.state_dict(),
os.path.join(args.save_path_prefix, f'{epoch:d}_{mean_iu*100:.0f}.pth')
)
writer.close()
def __init__(self, args):
self.args = args
# data transforms
input_transform = transform.Compose([
transform.ToTensor(),
transform.Normalize([.485, .456, .406], [.229, .224, .225])
])
# dataset
data_kwargs = {
'transform': input_transform,
'base_size': args.base_size,
'crop_size': args.crop_size
}
trainset = get_segmentation_dataset(args.dataset,
split=args.train_split,
mode='train',
**data_kwargs)
testset = get_segmentation_dataset(args.dataset,
split='val',
mode='val',
**data_kwargs)
# dataloader
kwargs = {'num_workers': args.workers, 'pin_memory': True} \
if args.cuda else {}
self.trainloader = data.DataLoader(trainset,
batch_size=args.batch_size,
drop_last=False,
shuffle=True,
**kwargs)
self.valloader = data.DataLoader(testset,
batch_size=args.batch_size,
drop_last=False,
shuffle=False,
**kwargs)
self.nclass = trainset.num_class
# model
model = get_segmentation_model(args.model,
dataset=args.dataset,
backbone=args.backbone,
dilated=args.dilated,
multi_grid=args.multi_grid,
stride=args.stride,
lateral=args.lateral,
jpu=args.jpu,
aux=args.aux,
se_loss=args.se_loss,
norm_layer=SyncBatchNorm,
base_size=args.base_size,
crop_size=args.crop_size)
# print(model)
# optimizer using different LR
params_list = [
{
'params': model.pretrained.parameters(),
'lr': args.lr
},
]
if hasattr(model, 'jpu') and model.jpu:
params_list.append({
'params': model.jpu.parameters(),
'lr': args.lr * 10
})
if hasattr(model, 'head'):
params_list.append({
'params': model.head.parameters(),
'lr': args.lr * 10
})
if hasattr(model, 'auxlayer'):
params_list.append({
'params': model.auxlayer.parameters(),
'lr': args.lr * 10
})
optimizer = torch.optim.SGD(params_list,
lr=args.lr,
momentum=args.momentum,
weight_decay=args.weight_decay)
class_balance_weight = 'None'
if args.dataset == "pcontext60":
class_balance_weight = torch.tensor([
1.3225e-01, 2.0757e+00, 1.8146e+01, 5.5052e+00, 2.2060e+00,
2.8054e+01, 2.0566e+00, 1.8598e+00, 2.4027e+00, 9.3435e+00,
3.5990e+00, 2.7487e-01, 1.4216e+00, 2.4986e+00, 7.7258e-01,
4.9020e-01, 2.9067e+00, 1.2197e+00, 2.2744e+00, 2.0444e+01,
3.0057e+00, 1.8167e+01, 3.7405e+00, 5.6749e-01, 3.2631e+00,
1.5007e+00, 5.5519e-01, 1.0056e+01, 1.8952e+01, 2.6792e-01,
2.7479e-01, 1.8309e+00, 2.0428e+01, 1.4788e+01, 1.4908e+00,
1.9113e+00, 2.6166e+02, 2.3233e-01, 1.9096e+01, 6.7025e+00,
2.8756e+00, 6.8804e-01, 4.4140e+00, 2.5621e+00, 4.4409e+00,
4.3821e+00, 1.3774e+01, 1.9803e-01, 3.6944e+00, 1.0397e+00,
2.0601e+00, 5.5811e+00, 1.3242e+00, 3.0088e-01, 1.7344e+01,
2.1569e+00, 2.7216e-01, 5.8731e-01, 1.9956e+00, 4.4004e+00
])
elif args.dataset == "ade20k":
class_balance_weight = torch.tensor([
0.0772, 0.0431, 0.0631, 0.0766, 0.1095, 0.1399, 0.1502, 0.1702,
0.2958, 0.3400, 0.3738, 0.3749, 0.4059, 0.4266, 0.4524, 0.5725,
0.6145, 0.6240, 0.6709, 0.6517, 0.6591, 0.6818, 0.9203, 0.9965,
1.0272, 1.0967, 1.1202, 1.2354, 1.2900, 1.5038, 1.5160, 1.5172,
1.5036, 2.0746, 2.1426, 2.3159, 2.2792, 2.6468, 2.8038, 2.8777,
2.9525, 2.9051, 3.1050, 3.1785, 3.3533, 3.5300, 3.6120, 3.7006,
3.6790, 3.8057, 3.7604, 3.8043, 3.6610, 3.8268, 4.0644, 4.2698,
4.0163, 4.0272, 4.1626, 4.3702, 4.3144, 4.3612, 4.4389, 4.5612,
5.1537, 4.7653, 4.8421, 4.6813, 5.1037, 5.0729, 5.2657, 5.6153,
5.8240, 5.5360, 5.6373, 6.6972, 6.4561, 6.9555, 7.9239, 7.3265,
7.7501, 7.7900, 8.0528, 8.5415, 8.1316, 8.6557, 9.0550, 9.0081,
9.3262, 9.1391, 9.7237, 9.3775, 9.4592, 9.7883, 10.6705,
10.2113, 10.5845, 10.9667, 10.8754, 10.8274, 11.6427, 11.0687,
10.8417, 11.0287, 12.2030, 12.8830, 12.5082, 13.0703, 13.8410,
12.3264, 12.9048, 12.9664, 12.3523, 13.9830, 13.8105, 14.0345,
15.0054, 13.9801, 14.1048, 13.9025, 13.6179, 17.0577, 15.8351,
17.7102, 17.3153, 19.4640, 17.7629, 19.9093, 16.9529, 19.3016,
17.6671, 19.4525, 20.0794, 18.3574, 19.1219, 19.5089, 19.2417,
20.2534, 20.0332, 21.7496, 21.5427, 20.3008, 21.1942, 22.7051,
23.3359, 22.4300, 20.9934, 26.9073, 31.7362, 30.0784
])
elif args.dataset == "cocostuff":
class_balance_weight = torch.tensor([
4.8557e-02, 6.4709e-02, 3.9255e+00, 9.4797e-01, 1.2703e+00,
1.4151e+00, 7.9733e-01, 8.4903e-01, 1.0751e+00, 2.4001e+00,
8.9736e+00, 5.3036e+00, 6.0410e+00, 9.3285e+00, 1.5952e+00,
3.6090e+00, 9.8772e-01, 1.2319e+00, 1.9194e+00, 2.7624e+00,
2.0548e+00, 1.2058e+00, 3.6424e+00, 2.0789e+00, 1.7851e+00,
6.7138e+00, 2.1315e+00, 6.9813e+00, 1.2679e+02, 2.0357e+00,
2.2933e+01, 2.3198e+01, 1.7439e+01, 4.1294e+01, 7.8678e+00,
4.3444e+01, 6.7543e+01, 1.0066e+01, 6.7520e+00, 1.3174e+01,
3.3499e+00, 6.9737e+00, 2.1482e+00, 1.9428e+01, 1.3240e+01,
1.9218e+01, 7.6836e-01, 2.6041e+00, 6.1822e+00, 1.4070e+00,
4.4074e+00, 5.7792e+00, 1.0321e+01, 4.9922e+00, 6.7408e-01,
3.1554e+00, 1.5832e+00, 8.9685e-01, 1.1686e+00, 2.6487e+00,
6.5354e-01, 2.3801e-01, 1.9536e+00, 1.5862e+00, 1.7797e+00,
2.7385e+01, 1.2419e+01, 3.9287e+00, 7.8897e+00, 7.5737e+00,
1.9758e+00, 8.1962e+01, 3.6922e+00, 2.0039e+00, 2.7333e+00,
5.4717e+00, 3.9048e+00, 1.9184e+01, 2.2689e+00, 2.6091e+02,
4.7366e+01, 2.3844e+00, 8.3310e+00, 1.4857e+01, 6.5076e+00,
2.0854e-01, 1.0425e+00, 1.7386e+00, 1.1973e+01, 5.2862e+00,
1.7341e+00, 8.6124e-01, 9.3702e+00, 2.8545e+00, 6.0123e+00,
1.7560e-01, 1.8128e+00, 1.3784e+00, 1.3699e+00, 2.3728e+00,
6.2819e-01, 1.3097e+00, 4.7892e-01, 1.0268e+01, 1.2307e+00,
5.5662e+00, 1.2867e+00, 1.2745e+00, 4.7505e+00, 8.4029e+00,
1.8679e+00, 1.0519e+01, 1.1240e+00, 1.4975e-01, 2.3146e+00,
4.1265e-01, 2.5896e+00, 1.4537e+00, 4.5575e+00, 7.8143e+00,
1.4603e+01, 2.8812e+00, 1.8868e+00, 7.8131e+01, 1.9323e+00,
7.4980e+00, 1.2446e+01, 2.1856e+00, 3.0973e+00, 4.1270e-01,
4.9016e+01, 7.1001e-01, 7.4035e+00, 2.3395e+00, 2.9207e-01,
2.4156e+00, 3.3211e+00, 2.1300e+00, 2.4533e-01, 1.7081e+00,
4.6621e+00, 2.9199e+00, 1.0407e+01, 7.6207e-01, 2.7806e-01,
3.7711e+00, 1.1852e-01, 8.8280e+00, 3.1700e-01, 6.3765e+01,
6.6032e+00, 5.2177e+00, 4.3596e+00, 6.2965e-01, 1.0207e+00,
1.1731e+01, 2.3935e+00, 9.2767e+00, 1.1023e-01, 3.6947e+00,
1.3943e+00, 2.3407e+00, 1.2112e-01, 2.8518e+00, 2.8195e+00,
1.0078e+00, 1.6614e+00, 6.5307e-01, 1.9070e+01, 2.7231e+00,
6.0769e-01
])
# criterions
self.criterion = SegmentationLosses(se_loss=args.se_loss,
aux=args.aux,
nclass=self.nclass,
se_weight=args.se_weight,
aux_weight=args.aux_weight,
weight=class_balance_weight)
self.model, self.optimizer = model, optimizer
# using cuda
if args.cuda:
self.model = DataParallelModel(self.model).cuda()
self.criterion = DataParallelCriterion(self.criterion).cuda()
# resuming checkpoint
self.best_pred = 0.0
if args.resume is not None:
if not os.path.isfile(args.resume):
raise RuntimeError("=> no checkpoint found at '{}'".format(
args.resume))
checkpoint = torch.load(args.resume)
args.start_epoch = checkpoint['epoch']
if args.cuda:
self.model.module.load_state_dict(checkpoint['state_dict'])
else:
self.model.load_state_dict(checkpoint['state_dict'])
if not args.ft:
self.optimizer.load_state_dict(checkpoint['optimizer'])
self.best_pred = checkpoint['best_pred']
print("=> loaded checkpoint '{}' (epoch {})".format(
args.resume, checkpoint['epoch']))
# clear start epoch if fine-tuning
if args.ft:
args.start_epoch = 0
# lr scheduler
self.scheduler = utils.LR_Scheduler(args.lr_scheduler, args.lr,
args.epochs, len(self.trainloader))
def train(args):
weight_dir = args.log_root # os.path.join(args.log_root, 'weights')
log_dir = os.path.join(
args.log_root, 'logs',
'SS-Net-{}'.format(time.strftime("%Y-%m-%d-%H-%M-%S",
time.localtime())))
data_dir = os.path.join(args.data_root, args.dataset)
# +++++++++++++++++++++++++++++++++++++++++++++++++++ #
# 1. Setup DataLoader
# +++++++++++++++++++++++++++++++++++++++++++++++++++ #
print("> # +++++++++++++++++++++++++++++++++++++++++++++++++++++++ #")
print("> 0. Setting up DataLoader...")
net_h, net_w = int(args.img_row * args.crop_ratio), int(args.img_col *
args.crop_ratio)
augment_train = Compose([
RandomHorizontallyFlip(),
RandomSized((0.5, 0.75)),
RandomRotate(5),
RandomCrop((net_h, net_w))
])
augment_valid = Compose([
RandomHorizontallyFlip(),
Scale((args.img_row, args.img_col)),
CenterCrop((net_h, net_w))
])
train_loader = CityscapesLoader(data_dir,
gt='gtFine',
split='train',
img_size=(args.img_row, args.img_col),
is_transform=True,
augmentations=augment_train)
valid_loader = CityscapesLoader(data_dir,
gt='gtFine',
split='val',
img_size=(args.img_row, args.img_col),
is_transform=True,
augmentations=augment_valid)
num_classes = train_loader.n_classes
tra_loader = data.DataLoader(train_loader,
batch_size=args.batch_size,
num_workers=int(multiprocessing.cpu_count() /
2),
shuffle=True)
val_loader = data.DataLoader(valid_loader,
batch_size=args.batch_size,
num_workers=int(multiprocessing.cpu_count() /
2))
# +++++++++++++++++++++++++++++++++++++++++++++++++++ #
# 2. Setup Model
# +++++++++++++++++++++++++++++++++++++++++++++++++++ #
print("> # +++++++++++++++++++++++++++++++++++++++++++++++++++++++ #")
print("> 1. Setting up Model...")
model = RetinaNet(num_classes=num_classes, input_size=(net_h, net_w))
# model = torch.nn.DataParallel(model, device_ids=[0,1,2]).cuda()
model = DataParallelModel(model,
device_ids=args.device_ids).cuda() # multi-gpu
# 2.1 Setup Optimizer
# +++++++++++++++++++++++++++++++++++++++++++++++++++ #
# Check if model has custom optimizer
if hasattr(model.module, 'optimizer'):
print('> Using custom optimizer')
optimizer = model.module.optimizer
else:
optimizer = torch.optim.SGD(model.parameters(),
lr=args.learning_rate,
momentum=0.90,
weight_decay=5e-4,
nesterov=True)
# optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate, weight_decay=1e-5)
# scheduler = torch.optim.lr_scheduler.StepLR(optimizer, step_size=4, gamma=0.1)
# scheduler = torch.optim.lr_scheduler.ExponentialLR(optimizer, gamma=0.9)
# 2.2 Setup Loss
# +++++++++++++++++++++++++++++++++++++++++++++++++++ #
class_weight = np.array([
0.05570516, 0.32337477, 0.08998544, 1.03602707, 1.03413147, 1.68195437,
5.58540548, 3.56563995, 0.12704978, 1., 0.46783719, 1.34551528,
5.29974114, 0.28342531, 0.9396095, 0.81551811, 0.42679146, 3.6399074,
2.78376194
],
dtype=float)
class_weight = torch.from_numpy(class_weight).float().cuda()
sem_loss = bootstrapped_cross_entropy2d
sem_loss = DataParallelCriterion(sem_loss, device_ids=args.device_ids)
se_loss = SemanticEncodingLoss(num_classes=19,
ignore_label=250,
alpha=0.50).cuda()
se_loss_parallel = DataParallelCriterion(se_loss,
device_ids=args.device_ids)
"""
# multi-gpu
bootstrapped_cross_entropy2d = ContextBootstrappedCELoss2D(num_classes=num_classes,
ignore=250,
kernel_size=5,
padding=4,
dilate=2,
use_gpu=True)
loss_sem = DataParallelCriterion(bootstrapped_cross_entropy2d, device_ids=[0, 1])
"""
# 2.3 Setup Metrics
# +++++++++++++++++++++++++++++++++++++++++++++++++++ #
# !!!!! Here Metrics !!!!!
metrics = RunningScore(num_classes) # num_classes = 93
# +++++++++++++++++++++++++++++++++++++++++++++++++++ #
# 3. Resume Model
# +++++++++++++++++++++++++++++++++++++++++++++++++++ #
print("> # +++++++++++++++++++++++++++++++++++++++++++++++++++++++ #")
print("> 2. Model state init or resume...")
args.start_epoch = 1
args.start_iter = 0
beat_map = 0.
if args.resume is not None:
full_path = os.path.join(os.path.join(weight_dir, 'train_model'),
args.resume)
if os.path.isfile(full_path):
print("> Loading model and optimizer from checkpoint '{}'".format(
args.resume))
checkpoint = torch.load(full_path)
args.start_epoch = checkpoint['epoch']
args.start_iter = checkpoint['iter']
beat_map = checkpoint['beat_map']
model.load_state_dict(checkpoint['model_state']) # weights
optimizer.load_state_dict(
checkpoint['optimizer_state']) # gradient state
del checkpoint
print("> Loaded checkpoint '{}' (epoch {}, iter {})".format(
args.resume, args.start_epoch, args.start_iter))
else:
print("> No checkpoint found at '{}'".format(full_path))
raise Exception("> No checkpoint found at '{}'".format(full_path))
else:
# init_weights(model, pi=0.01,
# pre_trained=os.path.join(args.log_root, 'resnet50_imagenet.pth'))
if args.pre_trained is not None:
print("> Loading weights from pre-trained model '{}'".format(
args.pre_trained))
full_path = os.path.join(args.log_root, args.pre_trained)
pre_weight = torch.load(full_path)
prefix = "module.fpn.base_net."
model_dict = model.state_dict()
pretrained_dict = {(prefix + k): v
for k, v in pre_weight.items()
if (prefix + k) in model_dict}
model_dict.update(pretrained_dict)
model.load_state_dict(model_dict)
del pre_weight
del model_dict
del pretrained_dict
# +++++++++++++++++++++++++++++++++++++++++++++++++++ #
# 4. Train Model
# +++++++++++++++++++++++++++++++++++++++++++++++++++ #
# 4.0. Setup tensor-board for visualization
# +++++++++++++++++++++++++++++++++++++++++++++++++++ #
writer = None
if args.tensor_board:
writer = SummaryWriter(log_dir=log_dir, comment="SSnet_Cityscapes")
# dummy_input = Variable(torch.rand(1, 3, args.img_row, args.img_col).cuda(), requires_grad=True)
# writer.add_graph(model, dummy_input)
print("> # +++++++++++++++++++++++++++++++++++++++++++++++++++++++ #")
print("> 3. Model Training start...")
topk_init = 512
num_batches = int(
math.ceil(
len(tra_loader.dataset.files[tra_loader.dataset.split]) /
float(tra_loader.batch_size)))
# lr_period = 20 * num_batches
for epoch in np.arange(args.start_epoch - 1, args.num_epochs):
# +++++++++++++++++++++++++++++++++++++++++++++++++++ #
# 4.1 Mini-Batch Training
# +++++++++++++++++++++++++++++++++++++++++++++++++++ #
model.train()
topk_base = topk_init
if epoch == args.start_epoch - 1:
pbar = tqdm(np.arange(args.start_iter, num_batches))
start_iter = args.start_iter
else:
pbar = tqdm(np.arange(num_batches))
start_iter = 0
lr = args.learning_rate
# lr = adjust_learning_rate(optimizer, init_lr=args.learning_rate, decay_rate=0.1, curr_epoch=epoch,
# epoch_step=20, start_decay_at_epoch=args.start_decay_at_epoch,
# total_epoch=args.num_epochs, mode='exp')
# scheduler.step()
# for train_i, (images, gt_masks) in enumerate(tra_loader): # One mini-Batch datasets, One iteration
for train_i, (images, gt_masks) in zip(range(start_iter, num_batches),
tra_loader):
full_iter = (epoch * num_batches) + train_i + 1
lr = poly_lr_scheduler(optimizer,
init_lr=args.learning_rate,
iter=full_iter,
lr_decay_iter=1,
max_iter=args.num_epochs * num_batches,
power=0.9)
# lr = args.learning_rate * cosine_annealing_lr(lr_period, full_iter)
# optimizer = set_optimizer_lr(optimizer, lr)
images = images.cuda().requires_grad_()
se_labels = se_loss.unique_encode(gt_masks)
se_labels = se_labels.cuda()
gt_masks = gt_masks.cuda()
topk_base = poly_topk_scheduler(init_topk=topk_init,
iter=full_iter,
topk_decay_iter=1,
max_iter=args.num_epochs *
num_batches,
power=0.95)
optimizer.zero_grad()
se, sem_seg_pred = model(images)
# --------------------------------------------------- #
# Compute loss
# --------------------------------------------------- #
topk = topk_base * 512
train_loss = sem_loss(input=sem_seg_pred,
target=gt_masks,
K=topk,
weight=None)
train_se_loss = se_loss_parallel(predicts=se,
enc_cls_target=se_labels,
size_average=True,
reduction='elementwise_mean')
loss = train_loss + args.alpha * train_se_loss
loss.backward() # back-propagation
torch.nn.utils.clip_grad_norm_(model.parameters(), 1e3)
optimizer.step() # parameter update based on the current gradient
pbar.update(1)
pbar.set_description("> Epoch [%d/%d]" %
(epoch + 1, args.num_epochs))
pbar.set_postfix(Train_Loss=train_loss.item(),
Train_SE_Loss=train_se_loss.item(),
TopK=topk_base)
# pbar.set_postfix(Train_Loss=train_loss.item(), TopK=topk_base)
# +++++++++++++++++++++++++++++++++++++++++++++++++++ #
# 4.1.1 Verbose training process
# +++++++++++++++++++++++++++++++++++++++++++++++++++ #
if (train_i + 1) % args.verbose_interval == 0:
# ---------------------------------------- #
# 1. Training Losses
# ---------------------------------------- #
loss_log = "Epoch [%d/%d], Iter: %d Loss1: \t %.4f " % (
epoch + 1, args.num_epochs, train_i + 1, loss.item())
# ---------------------------------------- #
# 2. Training Metrics
# ---------------------------------------- #
sem_seg_pred = F.softmax(sem_seg_pred, dim=1)
pred = sem_seg_pred.data.max(1)[1].cpu().numpy()
gt = gt_masks.data.cpu().numpy()
metrics.update(
gt,
pred) # accumulate the metrics (confusion_matrix and ious)
score, _ = metrics.get_scores()
metric_log = ""
for k, v in score.items():
metric_log += " {}: \t %.4f, ".format(k) % v
metrics.reset() # reset the metrics for each train_i steps
logs = loss_log + metric_log
if args.tensor_board:
writer.add_scalar('Training/Train_Loss', train_loss.item(),
full_iter)
writer.add_scalar('Training/Train_SE_Loss',
train_se_loss.item(), full_iter)
writer.add_scalar('Training/Loss', loss.item(), full_iter)
writer.add_scalar('Training/Lr', lr, full_iter)
writer.add_scalars('Training/Metrics', score, full_iter)
writer.add_text('Training/Text', logs, full_iter)
for name, param in model.named_parameters():
writer.add_histogram(name,
param.clone().cpu().data.numpy(),
full_iter)
"""
# each 2000 iterations save model
if (train_i + 1) % args.iter_interval_save_model == 0:
pbar.set_postfix(Loss=train_loss.item(), lr=lr)
state = {"epoch": epoch + 1,
"iter": train_i + 1,
'beat_map': beat_map,
"model_state": model.state_dict(),
"optimizer_state": optimizer.state_dict()}
save_dir = os.path.join(os.path.join(weight_dir, 'train_model'),
"ssnet_model_sem_se_{}epoch_{}iter.pkl".format(epoch+1, train_i+1))
torch.save(state, save_dir)
"""
# end of this training phase
state = {
"epoch": epoch + 1,
"iter": num_batches,
'beat_map': beat_map,
"model_state": model.state_dict(),
"optimizer_state": optimizer.state_dict()
}
save_dir = os.path.join(
os.path.join(args.log_root, 'train_model'),
"ssnet_model_sem_se_{}_{}epoch_{}iter.pkl".format(
args.model_details, epoch + 1, num_batches))
torch.save(state, save_dir)
# +++++++++++++++++++++++++++++++++++++++++++++++++++ #
# 4.2 Mini-Batch Validation
# +++++++++++++++++++++++++++++++++++++++++++++++++++ #
model.eval()
val_loss = 0.0
vali_count = 0
with torch.no_grad():
for i_val, (images_val, gt_masks_val) in enumerate(val_loader):
vali_count += 1
images_val = images_val.cuda()
se_labels_val = se_loss.unique_encode(gt_masks_val)
se_labels_val = se_labels_val.cuda()
gt_masks_val = gt_masks_val.cuda()
se_val, sem_seg_pred_val = model(images_val)
# !!!!!! Loss !!!!!!
topk_val = topk_base * 512
loss = sem_loss(sem_seg_pred_val, gt_masks_val, topk_val, weight=None) + \
args.alpha * se_loss_parallel(predicts=se_val, enc_cls_target=se_labels_val,
size_average=True, reduction='elementwise_mean')
val_loss += loss.item()
# accumulating the confusion matrix and ious
sem_seg_pred_val = F.softmax(sem_seg_pred_val, dim=1)
pred = sem_seg_pred_val.data.max(1)[1].cpu().numpy()
gt = gt_masks_val.data.cpu().numpy()
metrics.update(gt, pred)
# ---------------------------------------- #
# 1. Validation Losses
# ---------------------------------------- #
val_loss /= vali_count
loss_log = "Epoch [%d/%d], Loss: \t %.4f" % (
epoch + 1, args.num_epochs, val_loss)
# ---------------------------------------- #
# 2. Validation Metrics
# ---------------------------------------- #
metric_log = ""
score, _ = metrics.get_scores()
for k, v in score.items():
metric_log += " {}: \t %.4f, ".format(k) % v
metrics.reset() # reset the metrics
logs = loss_log + metric_log
pbar.set_postfix(
Vali_Loss=val_loss, Lr=lr,
Vali_mIoU=score['Mean_IoU']) # Train_Loss=train_loss.item()
if args.tensor_board:
writer.add_scalar('Validation/Loss', val_loss, epoch)
writer.add_scalars('Validation/Metrics', score, epoch)
writer.add_text('Validation/Text', logs, epoch)
for name, param in model.named_parameters():
writer.add_histogram(name,
param.clone().cpu().data.numpy(),
epoch)
# +++++++++++++++++++++++++++++++++++++++++++++++++++ #
# 4.3 End of one Epoch
# +++++++++++++++++++++++++++++++++++++++++++++++++++ #
# !!!!! Here choose suitable Metric for the best model selection !!!!!
if score['Mean_IoU'] >= beat_map:
beat_map = score['Mean_IoU']
state = {
"epoch": epoch + 1,
"beat_map": beat_map,
"model_state": model.state_dict(),
"optimizer_state": optimizer.state_dict()
}
save_dir = os.path.join(
weight_dir,
"SSnet_best_sem_se_{}_model.pkl".format(args.model_details))
torch.save(state, save_dir)
# Note that step should be called after validate()
pbar.close()
# +++++++++++++++++++++++++++++++++++++++++++++++++++ #
# 4.4 End of Training process
# +++++++++++++++++++++++++++++++++++++++++++++++++++ #
if args.tensor_board:
# export scalar datasets to JSON for external processing
# writer.export_scalars_to_json("{}/all_scalars.json".format(log_dir))
writer.close()
print("> # +++++++++++++++++++++++++++++++++++++++++++++++++++++++ #")
print("> Training Done!!!")
print("> # +++++++++++++++++++++++++++++++++++++++++++++++++++++++ #")
def __init__(self, args):
self.args = args
args.log_name = str(args.checkname)
root_dir = getattr(args, "data_root", '../datasets')
wo_head = getattr(args, "resume_wo_head", False)
self.logger = utils.create_logger(args.log_root, args.log_name)
# data transforms
input_transform = transform.Compose([
transform.ToTensor(),
transform.Normalize([.485, .456, .406], [.229, .224, .225])
])
# dataset
data_kwargs = {
'transform': input_transform,
'base_size': args.base_size,
'crop_size': args.crop_size,
'logger': self.logger,
'scale': args.scale
}
trainset = get_segmentation_dataset(args.dataset,
split='train',
mode='train',
root=root_dir,
**data_kwargs)
testset = get_segmentation_dataset(args.dataset,
split='val',
mode='val',
root=root_dir,
**data_kwargs)
# dataloader
kwargs = {'num_workers': args.workers, 'pin_memory': True} \
if args.cuda else {}
self.trainloader = data.DataLoader(trainset,
batch_size=args.batch_size,
drop_last=True,
shuffle=True,
**kwargs)
self.valloader = data.DataLoader(testset,
batch_size=args.batch_size,
drop_last=False,
shuffle=False,
**kwargs)
self.nclass = trainset.num_class
# model
model = get_segmentation_model(args.model,
dataset=args.dataset,
backbone=args.backbone,
aux=args.aux,
se_loss=args.se_loss,
norm_layer=BatchNorm2d,
base_size=args.base_size,
crop_size=args.crop_size,
multi_grid=args.multi_grid,
multi_dilation=args.multi_dilation)
#print(model)
self.logger.info(model)
# optimizer using different LR
if not args.wo_backbone:
params_list = [
{
'params': model.pretrained.parameters(),
'lr': args.lr
},
]
else:
params_list = []
if hasattr(model, 'head'):
params_list.append({
'params': model.head.parameters(),
'lr': args.lr * 10
})
if hasattr(model, 'auxlayer'):
params_list.append({
'params': model.auxlayer.parameters(),
'lr': args.lr * 10
})
optimizer = torch.optim.SGD(params_list,
lr=args.lr,
momentum=args.momentum,
weight_decay=args.weight_decay)
self.criterion = SegmentationMultiLosses(nclass=self.nclass)
#self.criterion = SegmentationLosses(se_loss=args.se_loss, aux=args.aux,nclass=self.nclass)
self.model, self.optimizer = model, optimizer
# using cuda
if args.cuda:
self.model = DataParallelModel(self.model).cuda()
self.criterion = DataParallelCriterion(self.criterion).cuda()
# finetune from a trained model
if args.ft:
args.start_epoch = 0
checkpoint = torch.load(args.ft_resume)
if wo_head:
print("WITHout HEAD !!!!!!!!!!")
from collections import OrderedDict
new = OrderedDict()
for k, v in checkpoint['state_dict'].items():
if not k.startswith("head"):
new[k] = v
checkpoint['state_dict'] = new
else:
print("With HEAD !!!!!!!!!!")
if args.cuda:
self.model.module.load_state_dict(checkpoint['state_dict'],
strict=False)
else:
self.model.load_state_dict(checkpoint['state_dict'],
strict=False)
# self.logger.info("=> loaded checkpoint '{}' (epoch {})".format(args.ft_resume, checkpoint['epoch']))
# resuming checkpoint
if args.resume:
if not os.path.isfile(args.resume):
raise RuntimeError("=> no checkpoint found at '{}'".format(
args.resume))
checkpoint = torch.load(args.resume)
args.start_epoch = checkpoint['epoch']
if args.cuda:
self.model.module.load_state_dict(checkpoint['state_dict'])
else:
self.model.load_state_dict(checkpoint['state_dict'])
if not args.ft:
self.optimizer.load_state_dict(checkpoint['optimizer'])
self.best_pred = checkpoint['best_pred']
self.logger.info("=> loaded checkpoint '{}' (epoch {})".format(
args.resume, checkpoint['epoch']))
# lr scheduler
self.scheduler = utils.LR_Scheduler(args.lr_scheduler,
args.lr,
args.epochs,
len(self.trainloader),
logger=self.logger,
lr_step=args.lr_step)
self.best_pred = 0.0
def __init__(self, args):
self.args = args
self.args.start_epoch = 0
self.args.cuda = True
# data transforms
input_transform = transform.Compose([
transform.ToTensor(),
transform.Normalize([.490, .490, .490], [.247, .247, .247])
]) # TODO: change mean and std
# dataset
train_chain = Compose([
HorizontalFlip(p=0.5),
OneOf([
ElasticTransform(
alpha=300, sigma=300 * 0.05, alpha_affine=300 * 0.03),
GridDistortion(),
OpticalDistortion(distort_limit=2, shift_limit=0.5),
],
p=0.3),
RandomSizedCrop(
min_max_height=(900, 1024), height=1024, width=1024, p=0.5),
ShiftScaleRotate(rotate_limit=20, p=0.5),
Resize(self.args.size, self.args.size)
],
p=1)
val_chain = Compose([Resize(self.args.size, self.args.size)], p=1)
num_fold = self.args.num_fold
df_train = pd.read_csv(os.path.join(args.imagelist_path, 'train.csv'))
df_val = pd.read_csv(os.path.join(args.imagelist_path, 'val.csv'))
df_full = pd.concat((df_train, df_val), ignore_index=True, axis=0)
df_full['lbl'] = (df_full['mask_name'].astype(str) == '-1').astype(int)
skf = StratifiedKFold(8, shuffle=True, random_state=777)
train_ids, val_ids = list(
skf.split(df_full['mask_name'], df_full['lbl']))[num_fold]
df_test = pd.read_csv(
os.path.join(args.imagelist_path, 'test_true.csv'))
df_new_train = pd.concat((df_full.iloc[train_ids], df_test),
ignore_index=True,
axis=0,
sort=False)
df_new_val = df_full.iloc[val_ids]
df_new_train.to_csv(f'/tmp/train_new_pneumo_{num_fold}.csv')
df_new_val.to_csv(f'/tmp/val_new_pneumo_{num_fold}.csv')
trainset = SegmentationDataset(f'/tmp/train_new_pneumo_{num_fold}.csv',
args.image_path,
args.masks_path,
input_transform=input_transform,
transform_chain=train_chain,
base_size=1024)
testset = SegmentationDataset(f'/tmp/val_new_pneumo_{num_fold}.csv',
args.image_path,
args.masks_path,
input_transform=input_transform,
transform_chain=val_chain,
base_size=1024)
imgs = trainset.mask_img_map[:, [0, 3]]
weights = make_weights_for_balanced_classes(imgs, 2)
weights = torch.DoubleTensor(weights)
train_sampler = (torch.utils.data.sampler.WeightedRandomSampler(
weights, len(weights)))
# dataloader
kwargs = {'num_workers': args.workers, 'pin_memory': True}
self.trainloader = data.DataLoader(
trainset,
batch_size=args.batch_size,
drop_last=True,
sampler=train_sampler, #shuffle=True,
**kwargs)
self.valloader = data.DataLoader(testset,
batch_size=args.batch_size,
drop_last=False,
shuffle=False,
**kwargs)
self.nclass = 1
if self.args.model == 'unet':
model = UNet(n_classes=self.nclass, norm_layer=SyncBatchNorm)
params_list = [
{
'params': model.parameters(),
'lr': args.lr
},
]
elif self.args.model == 'encnet':
model = EncNet(
nclass=self.nclass,
backbone=args.backbone,
aux=args.aux,
se_loss=args.se_loss,
norm_layer=SyncBatchNorm #nn.BatchNorm2d
)
# optimizer using different LR
params_list = [
{
'params': model.pretrained.parameters(),
'lr': args.lr
},
]
if hasattr(model, 'head'):
params_list.append({
'params': model.head.parameters(),
'lr': args.lr * 10
})
if hasattr(model, 'auxlayer'):
params_list.append({
'params': model.auxlayer.parameters(),
'lr': args.lr * 10
})
print(model)
optimizer = torch.optim.SGD(params_list,
lr=args.lr,
momentum=0.9,
weight_decay=args.wd)
# criterions
if self.nclass == 1:
self.criterion = SegmentationLossesBCE(se_loss=args.se_loss,
aux=args.aux,
nclass=self.nclass,
se_weight=args.se_weight,
aux_weight=args.aux_weight,
use_dice=args.use_dice)
else:
self.criterion = SegmentationLosses(
se_loss=args.se_loss,
aux=args.aux,
nclass=self.nclass,
se_weight=args.se_weight,
aux_weight=args.aux_weight,
)
self.model, self.optimizer = model, optimizer
self.best_pred = 0.0
self.model = DataParallelModel(self.model).cuda()
self.criterion = DataParallelCriterion(self.criterion).cuda()
# resuming checkpoint
if args.resume is not None:
if not os.path.isfile(args.resume):
raise RuntimeError("=> no checkpoint found at '{}'".format(
args.resume))
checkpoint = torch.load(args.resume) #, map_location='cpu')
self.args.start_epoch = checkpoint['epoch']
state_dict = {k: v for k, v in checkpoint['state_dict'].items()}
self.model.load_state_dict(state_dict)
self.optimizer.load_state_dict(checkpoint['optimizer'])
for g in self.optimizer.param_groups:
g['lr'] = args.lr
self.best_pred = checkpoint['best_pred']
print("=> loaded checkpoint '{}' (epoch {})".format(
args.resume, checkpoint['epoch']))
print(f'Best dice: {checkpoint["best_pred"]}')
print(f'LR: {get_lr(self.optimizer):.5f}')
self.scheduler = ReduceLROnPlateau(self.optimizer,
mode='min',
factor=0.8,
patience=4,
threshold=0.001,
threshold_mode='abs',
min_lr=0.00001)
self.logger = Logger(args.logger_dir)
self.step_train = 0
self.best_loss = 20
self.step_val = 0