def main():
global args, best_loss
args = parser.parse_args()
if args.tensorboard:
from tensorboard_logger import configure
print("Using tensorboard")
configure("%s" % (args.dir))
# loading core configuration
c_config = CoreConfig()
if args.core_config == '':
print('No core config file given, using default core configuration')
if not os.path.exists(args.core_config):
sys.exit('Cannot find the config file: {}'.format(args.core_config))
else:
c_config.read(args.core_config)
print('Using core configuration from {}'.format(args.core_config))
# loading Unet configuration
u_config = UnetConfig()
if args.unet_config == '':
print('No unet config file given, using default unet configuration')
if not os.path.exists(args.unet_config):
sys.exit('Cannot find the unet configuration file: {}'.format(
args.unet_config))
else:
# need train_image_size for validation
u_config.read(args.unet_config, args.train_image_size)
print('Using unet configuration from {}'.format(args.unet_config))
offset_list = c_config.offsets
print("offsets are: {}".format(offset_list))
# model configurations from core config
num_classes = c_config.num_classes
num_colors = c_config.num_colors
num_offsets = len(c_config.offsets)
# model configurations from unet config
start_filters = u_config.start_filters
up_mode = u_config.up_mode
merge_mode = u_config.merge_mode
depth = u_config.depth
train_data = args.train_dir + '/train'
val_data = args.train_dir + '/val'
trainset = WaldoDataset(train_data, c_config, args.train_image_size)
trainloader = torch.utils.data.DataLoader(trainset,
num_workers=4,
batch_size=args.batch_size,
shuffle=True)
valset = WaldoDataset(val_data, c_config, args.train_image_size)
valloader = torch.utils.data.DataLoader(valset,
num_workers=4,
batch_size=args.batch_size)
NUM_TRAIN = len(trainset)
NUM_VAL = len(valset)
NUM_ALL = NUM_TRAIN + NUM_VAL
print('Total samples: {0} \n'
'Using {1} samples for training, '
'{2} samples for validation'.format(NUM_ALL, NUM_TRAIN, NUM_VAL))
# create model
model = UNet(num_classes,
num_offsets,
in_channels=num_colors,
depth=depth,
start_filts=start_filters,
up_mode=up_mode,
merge_mode=merge_mode).cuda()
# get the number of model parameters
print('Number of model parameters: {}'.format(
sum([p.data.nelement() for p in model.parameters()])))
# optionally resume from a checkpoint
if args.resume:
if os.path.isfile(args.resume):
print("=> loading checkpoint '{}'".format(args.resume))
checkpoint = torch.load(args.resume)
args.start_epoch = checkpoint['epoch']
best_loss = checkpoint['best_loss']
model.load_state_dict(checkpoint['state_dict'])
print("=> loaded checkpoint '{}' (epoch {})".format(
args.resume, checkpoint['epoch']))
else:
print("=> no checkpoint found at '{}'".format(args.resume))
# define optimizer
# optimizer = t.optim.Adam(model.parameters(), lr=1e-3)
optimizer = torch.optim.SGD(model.parameters(),
args.lr,
momentum=args.momentum,
nesterov=args.nesterov,
weight_decay=args.weight_decay)
# Train
for epoch in range(args.start_epoch, args.epochs):
Train(trainloader, model, optimizer, epoch)
val_loss = Validate(valloader, model, epoch)
is_best = val_loss < best_loss
best_loss = min(val_loss, best_loss)
save_checkpoint(
{
'epoch': epoch + 1,
'state_dict': model.state_dict(),
'best_prec1': best_loss,
}, is_best)
print('Best validation loss: ', best_loss)
# visualize some example outputs
outdir = '{}/imgs'.format(args.dir)
if not os.path.exists(outdir):
os.makedirs(outdir)
sample(model, valloader, outdir, c_config)
def main():
global args, best_loss
args = parser.parse_args()
if args.tensorboard:
print("Using tensorboard")
configure("exp/%s" % (args.name))
if not (os.path.exists(args.train_data) and os.path.exists(args.train_data)
and os.path.exists(args.test_data)):
train, val, test = DataProcess(args.train_path, args.test_path, 0.9,
args.img_channels)
torch.save(train, args.train_data)
torch.save(val, args.val_data)
torch.save(test, args.test_data)
s_trans = tsf.Compose([
tsf.ToPILImage(),
tsf.Resize((args.img_height, args.img_width)),
tsf.ToTensor(),
])
offset_list = [(1, 1), (0, -2)]
# split the training set into training set and validation set
trainset = Dataset(args.train_data, s_trans, offset_list, args.num_classes,
args.img_height, args.img_width)
trainloader = torch.utils.data.DataLoader(trainset,
num_workers=1,
batch_size=args.batch_size)
valset = Dataset(args.val_data, s_trans, offset_list, args.num_classes,
args.img_height, args.img_width)
valloader = torch.utils.data.DataLoader(valset,
num_workers=1,
batch_size=args.batch_size)
# datailer = iter(trainloader)
# img, bound, class_id = datailer.next()
# # print img.shape, bound.shape, class_id.shape
# torch.set_printoptions(threshold=5000)
# print bound.shape
# torchvision.utils.save_image(img, 'raw.png')
# torchvision.utils.save_image(bound[:, 0:1, :, :], 'bound1.png')
# torchvision.utils.save_image(bound[:, 1:2, :, :], 'bound2.png')
# torchvision.utils.save_image(class_id[:, 0:1, :, :], 'class1.png')
# torchvision.utils.save_image(class_id[:, 1:2, :, :], 'class2.png')
# sys.exit('stop')
NUM_TRAIN = len(trainset)
NUM_VAL = len(valset)
NUM_ALL = NUM_TRAIN + NUM_VAL
print(
'Total samples: {0} \n'
'Using {1} samples for training, '
'{2} samples for validation'.format(NUM_ALL, NUM_TRAIN, NUM_VAL))
# create model
model = UNet(args.num_classes,
len(offset_list),
in_channels=3,
depth=args.depth).cuda()
# model = UNet(3, 1, len(offset_list))
# get the number of model parameters
print('Number of model parameters: {}'.format(
sum([p.data.nelement() for p in model.parameters()])))
# optionally resume from a checkpoint
if args.resume:
if os.path.isfile(args.resume):
print("=> loading checkpoint '{}'".format(args.resume))
checkpoint = torch.load(args.resume)
args.start_epoch = checkpoint['epoch']
best_loss = checkpoint['best_loss']
model.load_state_dict(checkpoint['state_dict'])
print("=> loaded checkpoint '{}' (epoch {})".format(
args.resume, checkpoint['epoch']))
else:
print("=> no checkpoint found at '{}'".format(args.resume))
# define optimizer
# optimizer = t.optim.Adam(model.parameters(), lr=1e-3)
optimizer = torch.optim.SGD(model.parameters(),
args.lr,
momentum=args.momentum,
nesterov=args.nesterov,
weight_decay=args.weight_decay)
# Train
for epoch in range(args.start_epoch, args.epochs):
Train(trainloader, model, optimizer, epoch)
val_loss = Validate(valloader, model, epoch)
is_best = val_loss < best_loss
best_loss = min(val_loss, best_loss)
save_checkpoint(
{
'epoch': epoch + 1,
'state_dict': model.state_dict(),
'best_prec1': best_loss,
}, is_best)
print 'Best validation loss: ', best_loss
# Visualize some predicted masks on training data to get a better intuition
# about the performance. Comment it if not necessary.
datailer = iter(trainloader)
img, classification, bound = datailer.next()
torchvision.utils.save_image(img, 'imgs/raw.png')
for i in range(len(offset_list)):
torchvision.utils.save_image(bound[:, i:i + 1, :, :],
'imgs/bound_{}.png'.format(i))
for i in range(args.num_classes):
torchvision.utils.save_image(classification[:, i:i + 1, :, :],
'imgs/class_{}.png'.format(i))
img = torch.autograd.Variable(img).cuda()
predictions = model(img)
predictions = predictions.data
class_pred = predictions[:, :args.num_classes, :, :]
bound_pred = predictions[:, args.num_classes:, :, :]
for i in range(len(offset_list)):
torchvision.utils.save_image(bound_pred[:, i:i + 1, :, :],
'imgs/bound_pred{}.png'.format(i))
for i in range(args.num_classes):
torchvision.utils.save_image(class_pred[:, i:i + 1, :, :],
'imgs/class_pred{}.png'.format(i))
def main():
global args, best_loss
args = parser.parse_args()
if args.tensorboard:
print("Using tensorboard")
configure("exp/%s" % (args.name))
if not (os.path.exists(args.train_data) and os.path.exists(args.train_data)
and os.path.exists(args.test_data)):
train, val, test = DataProcess(args.train_path, args.test_path, 0.9,
args.img_channels)
t.save(train, args.train_data)
t.save(val, args.val_data)
t.save(test, args.test_data)
s_trans = tsf.Compose([
tsf.ToPILImage(),
tsf.Resize((args.img_height, args.img_width)),
tsf.ToTensor(),
])
t_trans = tsf.Compose([
tsf.ToPILImage(),
tsf.Resize((args.img_height, args.img_width),
interpolation=PIL.Image.NEAREST),
tsf.ToTensor(),
])
# split the training set into training set and validation set
trainset = TrainDataset(args.train_data, s_trans, t_trans)
trainloader = t.utils.data.DataLoader(trainset,
num_workers=1,
batch_size=args.batch_size,
shuffle=True)
valset = TrainDataset(args.val_data, s_trans, t_trans)
valloader = t.utils.data.DataLoader(valset,
num_workers=1,
batch_size=args.batch_size)
NUM_TRAIN = len(trainset)
NUM_VAL = len(valset)
NUM_ALL = NUM_TRAIN + NUM_VAL
print(
'Total samples: {0} \n'
'Using {1} samples for training, '
'{2} samples for validation'.format(NUM_ALL, NUM_TRAIN, NUM_VAL))
testset = TestDataset(args.test_data, s_trans)
testloader = t.utils.data.DataLoader(testset, num_workers=1, batch_size=1)
# create model
model = UNet(1, in_channels=3, depth=args.depth).cuda()
# get the number of model parameters
print('Number of model parameters: {}'.format(
sum([p.data.nelement() for p in model.parameters()])))
# optionally resume from a checkpoint
if args.resume:
if os.path.isfile(args.resume):
print("=> loading checkpoint '{}'".format(args.resume))
checkpoint = t.load(args.resume)
args.start_epoch = checkpoint['epoch']
best_prec1 = checkpoint['best_loss']
model.load_state_dict(checkpoint['state_dict'])
print("=> loaded checkpoint '{}' (epoch {})".format(
args.resume, checkpoint['epoch']))
else:
print("=> no checkpoint found at '{}'".format(args.resume))
# define optimizer
optimizer = t.optim.Adam(model.parameters(), lr=1e-3)
# Train
for epoch in range(args.start_epoch, args.epochs):
Train(trainloader, model, optimizer, epoch)
val_loss = Validate(valloader, model, epoch)
is_best = val_loss < best_loss
best_loss = min(val_loss, best_loss)
save_checkpoint(
{
'epoch': epoch + 1,
'state_dict': model.state_dict(),
'best_prec1': best_loss,
}, is_best)
print 'Best validation loss: ', best_loss
# Visualize some predicted masks on training data to get a better intuition
# about the performance. Comment it if not necessary.
datailer = iter(trainloader)
img, mask = datailer.next()
torchvision.utils.save_image(img, 'raw.png')
torchvision.utils.save_image(mask, 'mask.png')
img = t.autograd.Variable(img).cuda()
img_pred = model(img)
img_pred = img_pred.data
torchvision.utils.save_image(img_pred > 0.5, 'predicted.png')
# Load the best model and evaluate on test set
checkpoint = t.load('exp/%s/' % (args.name) + 'model_best.pth.tar')
model.load_state_dict(checkpoint['state_dict'])
Test(testloader, model)
init_util.print_network(model)
# model = nn.DataParallel(model, device_ids=[0]) # multi-GPU
log = logger.Logger('./output/{}'.format(args.save))
best = [0, np.inf] # 初始化最优模型的epoch和performance
trigger = 0 # early stop 计数器
for epoch in range(1, args.epochs + 1):
common.adjust_learning_rate(optimizer, epoch, args)
train_log = train(model, train_loader)
val_log = val(model, val_loader)
log.update(epoch, train_log, val_log)
# Save checkpoint.
state = {
'net': model.state_dict(),
'optimizer': optimizer.state_dict(),
'epoch': epoch
}
torch.save(
state,
os.path.join('./output/{}'.format(args.save), 'latest_model.pth'))
trigger += 1
if val_log['Val Loss'] < best[1]:
print('Saving best model')
torch.save(
state,
os.path.join('./output/{}'.format(args.save),
'best_model.pth'))
best[0] = epoch
best[1] = val_log['Val Loss']
transform = transforms.Compose([
utils.transforms.RandomMirror(),
utils.transforms.ToTensor(),
utils.transforms.Downsize(2)
])
dataset = utils.datasets.SteelDefectDataset(csv_file='train.csv',
root_dir='data/severstal-steel-defect-detection',transform=transform)
train_loader = DataLoader(dataset, batch_size=1,shuffle=True)
criterion = utils.loss.SegmentMSELoss()
for e in range(1,epoch+1):
print('Epoch {}:'.format(e))
total_loss = 0
for batch, data in tqdm(enumerate(train_loader),total=len(train_loader),leave=False):
optimizer.zero_grad()
imgs, cs, targets = data['img'], data['c'], data['target']
imgs = imgs.to(device)
targets = targets.to(device)
out = model(imgs)
loss = criterion(out,cs,targets)
loss.backward()
total_loss += loss.data
optimizer.step()
if batch == 500:
print(total_loss/batch)
print('Loss: {:.3f}'.format(total_loss/(batch)))
torch.save(model.state_dict(), os.path.join('weights','Unet_e{}.pth'.format(e)))