def main():
TTA2_preprocess = [preprocess, preprocess_hflip]
TTA10_preprocess = [preprocess_tencrop]
TTA12_preprocess = [preprocess, preprocess_hflip, preprocess_tencrop]
id = 0
print("testing {}.....".format(ckp_path))
for trans in TTA10_preprocess:
print("id is: {}".format(id))
test_dataset = McDataset(test_root, test_source, transform=trans)
test_loader = DataLoader(test_dataset,
batch_size,
shuffle=False,
pin_memory=False)
print("test loading....")
model = get_model('senet154', pretrained=False)
# model.cuda()
model = torch.nn.DataParallel(model).cuda()
checkpoint = load_checkpoint(ckp_path)
model.load_state_dict(checkpoint['state_dict'])
tester = TenCropTester(model)
# if id == 2:
# tester = TenCropTester(model)
# else:
# tester = BaseTester(model)
pred = tester.extract(test_loader)
np.save("./rst/prob_dense{}.npy".format(id), pred)
id += 1
def main():
# tta_preprocess = [preprocess, preprocess_hflip]
tta_preprocess = [preprocess_tencrop]
id = 0
print("testing {}.....".format(ckp_path))
for trans in tta_preprocess:
print("id is: {}".format(id))
test_dataset = McDataset(test_root, test_source, transform=trans)
test_loader = DataLoader(test_dataset,
batch_size,
shuffle=False,
pin_memory=False)
print("test loading....")
model = models.__dict__[arch]()
model = FineTuneModel(model, arch, 128)
# model.cuda()
model = torch.nn.DataParallel(model).cuda()
checkpoint = load_checkpoint(ckp_path)
model.load_state_dict(checkpoint['state_dict'])
# tester = BaseTester(model)
tester = TenCropTester(model)
pred = tester.extract(test_loader)
np.save("./rst/prob_dense{}.npy".format(id), pred)
id += 1
def main():
normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
test_dataset = McDataset(
test_root,
test_source,
transforms.Compose([
transforms.Resize(image_size),
transforms.CenterCrop(input_size),
transforms.ToTensor(),
normalize,
]))
test_loader = DataLoader(test_dataset, batch_size, shuffle=False, pin_memory=False)
model = models.__dict__[arch]()
model = FineTuneModel(model, arch, 128)
# model.cuda()
model = torch.nn.DataParallel(model).cuda()
checkpoint = load_checkpoint(ckp_path)
model.load_state_dict(checkpoint['state_dict'])
tester = BaseTester(model)
pred = tester.extract(test_loader)
np.save("./rst/prob_dense.npy", pred)
def main():
global args, best_prec1, min_loss
args = parser.parse_args()
rank, world_size = dist_init(args.port)
print("world_size is: {}".format(world_size))
assert (args.batch_size % world_size == 0)
assert (args.workers % world_size == 0)
args.batch_size = args.batch_size // world_size
args.workers = args.workers // world_size
# create model
print("=> creating model '{}'".format("inceptionv4"))
print("save_path is: {}".format(args.save_path))
image_size = 341
input_size = 299
model = get_model('inceptionv4', pretrained=True)
# print("model is: {}".format(model))
model.cuda()
model = DistModule(model)
# optionally resume from a checkpoint
if args.load_path:
if args.resume_opt:
best_prec1, start_epoch = load_state(args.load_path,
model,
optimizer=optimizer)
else:
# print('load weights from', args.load_path)
load_state(args.load_path, model)
cudnn.benchmark = True
# Data loading code
normalize = transforms.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])
train_dataset = McDataset(
args.train_root, args.train_source,
transforms.Compose([
transforms.Resize(image_size),
transforms.RandomCrop(input_size),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
ColorAugmentation(),
normalize,
]))
val_dataset = McDataset(
args.val_root, args.val_source,
transforms.Compose([
transforms.Resize(image_size),
transforms.CenterCrop(input_size),
transforms.ToTensor(),
normalize,
]))
train_sampler = DistributedSampler(train_dataset)
val_sampler = DistributedSampler(val_dataset)
train_loader = DataLoader(train_dataset,
batch_size=args.batch_size,
shuffle=False,
num_workers=args.workers,
pin_memory=False,
sampler=train_sampler)
val_loader = DataLoader(val_dataset,
batch_size=args.batch_size,
shuffle=False,
num_workers=args.workers,
pin_memory=False,
sampler=val_sampler)
# define loss function (criterion) and optimizer
criterion = nn.CrossEntropyLoss()
lr = 0
patience = 0
for epoch in range(args.start_epoch, args.epochs):
# adjust_learning_rate(optimizer, epoch)
train_sampler.set_epoch(epoch)
if epoch == 1:
lr = 0.00003
if patience == 2:
patience = 0
checkpoint = load_checkpoint(args.save_path + '_best.pth.tar')
model.load_state_dict(checkpoint['state_dict'])
print("Loading checkpoint_best.............")
# model.load_state_dict(torch.load('checkpoint_best.pth.tar'))
lr = lr / 10.0
if epoch == 0:
lr = 0.001
for name, param in model.named_parameters():
# print("name is: {}".format(name))
if (name not in last_layer_names):
param.requires_grad = False
optimizer = torch.optim.RMSprop(filter(lambda p: p.requires_grad,
model.parameters()),
lr=lr)
# optimizer = torch.optim.Adam(
# filter(lambda p: p.requires_grad, model.parameters()), lr=lr)
else:
for param in model.parameters():
param.requires_grad = True
optimizer = torch.optim.RMSprop(model.parameters(),
lr=lr,
weight_decay=0.0001)
# optimizer = torch.optim.Adam(
# model.parameters(), lr=lr, weight_decay=0.0001)
print("lr is: {}".format(lr))
# train for one epoch
train(train_loader, model, criterion, optimizer, epoch)
# evaluate on validation set
val_prec1, val_losses = validate(val_loader, model, criterion)
print("val_losses is: {}".format(val_losses))
# remember best prec@1 and save checkpoint
if rank == 0:
# remember best prec@1 and save checkpoint
if val_losses < min_loss:
is_best = True
save_checkpoint(
{
'epoch': epoch + 1,
'arch': 'inceptionv4',
'state_dict': model.state_dict(),
'best_prec1': best_prec1,
'optimizer': optimizer.state_dict(),
}, is_best, args.save_path)
# torch.save(model.state_dict(), 'best_val_weight.pth')
print(
'val score improved from {:.5f} to {:.5f}. Saved!'.format(
min_loss, val_losses))
min_loss = val_losses
patience = 0
else:
patience += 1
if rank == 1 or rank == 2 or rank == 3 or rank == 4 or rank == 5 or rank == 6 or rank == 7:
if val_losses < min_loss:
min_loss = val_losses
patience = 0
else:
patience += 1
print("patience is: {}".format(patience))
print("min_loss is: {}".format(min_loss))
print("min_loss is: {}".format(min_loss))
def main():
global args, config, best_prec1
args = parser.parse_args()
with open(args.config) as f:
config = yaml.load(f)
config = EasyDict(config['common'])
config.save_path = os.path.dirname(args.config)
rank, world_size = dist_init()
# create model
bn_group_size = config.model.kwargs.bn_group_size
bn_var_mode = config.model.kwargs.get('bn_var_mode', 'L2')
if bn_group_size == 1:
bn_group = None
else:
assert world_size % bn_group_size == 0
bn_group = simple_group_split(world_size, rank,
world_size // bn_group_size)
config.model.kwargs.bn_group = bn_group
config.model.kwargs.bn_var_mode = (link.syncbnVarMode_t.L1 if bn_var_mode
== 'L1' else link.syncbnVarMode_t.L2)
model = model_entry(config.model)
if rank == 0:
print(model)
model.cuda()
if config.optimizer.type == 'FP16SGD' or config.optimizer.type == 'FusedFP16SGD':
args.fp16 = True
else:
args.fp16 = False
if args.fp16:
# if you have modules that must use fp32 parameters, and need fp32 input
# try use link.fp16.register_float_module(your_module)
# if you only need fp32 parameters set cast_args=False when call this
# function, then call link.fp16.init() before call model.half()
if config.optimizer.get('fp16_normal_bn', False):
print('using normal bn for fp16')
link.fp16.register_float_module(link.nn.SyncBatchNorm2d,
cast_args=False)
link.fp16.register_float_module(torch.nn.BatchNorm2d,
cast_args=False)
link.fp16.init()
model.half()
model = DistModule(model, args.sync)
# create optimizer
opt_config = config.optimizer
opt_config.kwargs.lr = config.lr_scheduler.base_lr
if config.get('no_wd', False):
param_group, type2num = param_group_no_wd(model)
opt_config.kwargs.params = param_group
else:
opt_config.kwargs.params = model.parameters()
optimizer = optim_entry(opt_config)
# optionally resume from a checkpoint
last_iter = -1
best_prec1 = 0
if args.load_path:
if args.recover:
best_prec1, last_iter = load_state(args.load_path,
model,
optimizer=optimizer)
else:
load_state(args.load_path, model)
cudnn.benchmark = True
# Data loading code
normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
# augmentation
aug = [
transforms.RandomResizedCrop(config.augmentation.input_size),
transforms.RandomHorizontalFlip()
]
for k in config.augmentation.keys():
assert k in [
'input_size', 'test_resize', 'rotation', 'colorjitter', 'colorold'
]
rotation = config.augmentation.get('rotation', 0)
colorjitter = config.augmentation.get('colorjitter', None)
colorold = config.augmentation.get('colorold', False)
if rotation > 0:
aug.append(transforms.RandomRotation(rotation))
if colorjitter is not None:
aug.append(transforms.ColorJitter(*colorjitter))
aug.append(transforms.ToTensor())
if colorold:
aug.append(ColorAugmentation())
aug.append(normalize)
# train
train_dataset = McDataset(config.train_root,
config.train_source,
transforms.Compose(aug),
fake=args.fake)
# val
val_dataset = McDataset(
config.val_root, config.val_source,
transforms.Compose([
transforms.Resize(config.augmentation.test_resize),
transforms.CenterCrop(config.augmentation.input_size),
transforms.ToTensor(),
normalize,
]), args.fake)
train_sampler = DistributedGivenIterationSampler(
train_dataset,
config.lr_scheduler.max_iter,
config.batch_size,
last_iter=last_iter)
val_sampler = DistributedSampler(val_dataset, round_up=False)
train_loader = DataLoader(train_dataset,
batch_size=config.batch_size,
shuffle=False,
num_workers=config.workers,
pin_memory=True,
sampler=train_sampler)
val_loader = DataLoader(val_dataset,
batch_size=config.batch_size,
shuffle=False,
num_workers=config.workers,
pin_memory=True,
sampler=val_sampler)
config.lr_scheduler['optimizer'] = optimizer.optimizer if isinstance(
optimizer, FP16SGD) else optimizer
config.lr_scheduler['last_iter'] = last_iter
lr_scheduler = get_scheduler(config.lr_scheduler)
if rank == 0:
tb_logger = SummaryWriter(config.save_path + '/events')
logger = create_logger('global_logger', config.save_path + '/log.txt')
logger.info('args: {}'.format(pprint.pformat(args)))
logger.info('config: {}'.format(pprint.pformat(config)))
else:
tb_logger = None
if args.evaluate:
if args.fusion_list is not None:
validate(val_loader,
model,
fusion_list=args.fusion_list,
fuse_prob=args.fuse_prob)
else:
validate(val_loader, model)
link.finalize()
return
train(train_loader, val_loader, model, optimizer, lr_scheduler,
last_iter + 1, tb_logger)
link.finalize()
def main():
global args, best_prec1, timer
args = parser.parse_args()
rank, world_size = dist_init(args.port)
assert (args.batch_size % world_size == 0)
assert (args.workers % world_size == 0)
args.batch_size = args.batch_size // world_size
args.workers = args.workers // world_size
# step1: create model
print("=> creating model '{}'".format(args.arch))
if args.arch.startswith('inception_v3'):
print('inception_v3 without aux_logits!')
image_size = 341
input_size = 299
model = models.__dict__[args.arch](aux_logits=False)
elif args.arch.startswith('ir18'):
image_size = 640
input_size = 448
model = IR18()
else:
image_size = 256
input_size = 224
model = models.__dict__[args.arch]()
if args.pretrained:
print("=> using pre-trained model '{}'".format(args.arch))
if os.path.isfile(args.pretrained):
print("=> loading pretrained_model '{}'".format(args.pretrained))
pretrained_model = torch.load(args.pretrained)
model.load_state_dict(pretrained_model['state_dict'], strict=False)
print("=> loaded pretrained_model '{}'".format(args.pretrained))
else:
print("=> no checkpoint found at '{}'".format(args.pretrained))
model.cuda()
model = DistModule(model)
# step2: define loss function (criterion) and optimizer
criterion = nn.CrossEntropyLoss()
optimizer = torch.optim.SGD(model.parameters(),
args.lr,
momentum=args.momentum,
weight_decay=args.weight_decay)
# 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_prec1 = checkpoint['best_prec1']
model.load_state_dict(checkpoint['state_dict'])
optimizer.load_state_dict(checkpoint['optimizer'])
print("=> loaded checkpoint '{}' (epoch {})".format(
args.resume, checkpoint['epoch']))
else:
print("=> no checkpoint found at '{}'".format(args.resume))
cudnn.benchmark = True
# step3: Data loading code
normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
train_dataset = McDataset(
args.train_root,
args.train_source,
transforms.Compose([
transforms.RandomResizedCrop(input_size),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
# ColorAugmentation(),
# normalize,
]))
val_dataset = McDataset(
args.val_root,
args.val_source,
transforms.Compose([
transforms.Resize(image_size),
transforms.CenterCrop(input_size),
transforms.ToTensor(),
# normalize,
]))
train_sampler = DistributedSampler(train_dataset)
val_sampler = DistributedSampler(val_dataset)
train_loader = DataLoader(train_dataset,
batch_size=args.batch_size,
shuffle=False,
num_workers=args.workers,
pin_memory=False,
sampler=train_sampler)
val_loader = DataLoader(val_dataset,
batch_size=args.batch_size,
shuffle=False,
num_workers=args.workers,
pin_memory=False,
sampler=val_sampler)
if args.evaluate:
validate(val_loader, model, criterion)
return
timer = Timer(
len(train_loader) + len(val_loader), args.epochs - args.start_epoch)
for epoch in range(args.start_epoch, args.epochs):
adjust_learning_rate(optimizer, epoch)
train_sampler.set_epoch(epoch)
# train for one epoch
train(train_loader, model, criterion, optimizer, epoch)
# evaluate on validation set
prec1 = validate(val_loader, model, criterion)
if rank == 0:
# remember best prec@1 and save checkpoint
is_best = prec1 > best_prec1
best_prec1 = max(prec1, best_prec1)
save_checkpoint(
{
'epoch': epoch + 1,
'arch': args.arch,
'state_dict': model.state_dict(),
'best_prec1': best_prec1,
'optimizer': optimizer.state_dict(),
}, is_best, args.save_path)
print('* Best Prec 1: {best:.3f}'.format(best=best_prec1))
def main():
global writer, best_prec1
if not os.path.exists('{}/checkpoints'.format(args.ckpt)):
os.makedirs('{}/checkpoints'.format(args.ckpt))
if not os.path.exists('{}/images'.format(args.ckpt)):
os.makedirs('{}/images'.format(args.ckpt))
if not os.path.exists('{}/logs'.format(args.ckpt)):
os.makedirs('{}/logs'.format(args.ckpt))
if not os.path.exists('{}/plots'.format(args.ckpt)):
os.makedirs('{}/plots'.format(args.ckpt))
if os.path.exists('{}/runs'.format(args.ckpt)):
shutil.rmtree('{}/runs'.format(args.ckpt))
os.makedirs('{}/runs'.format(args.ckpt))
logger = create_logger('global_logger', '{}/logs/{}.txt'.format(args.ckpt,time.time()))
logger.info('{}'.format(args))
writer = SummaryWriter('{}/runs'.format(args.ckpt))
#net = nn.parallel.distributed.DistributedDataParallel(net)
# build dataset
data_dir = args.data_dir
data_list = args.data_list
val_data_dir = args.val_data_dir
val_data_list = args.val_data_list
train_dataset = McDataset(data_dir, data_list)
train_loader = torch.utils.data.DataLoader(
train_dataset, batch_size=args.bs, shuffle=True,
num_workers=1, pin_memory=True, collate_fn=fast_collate)
val_dataset = McDataset(val_data_dir, val_data_list)
val_loader = torch.utils.data.DataLoader(
val_dataset, batch_size=args.bs, shuffle=True,
num_workers=4, pin_memory=True, collate_fn=fast_collate)
# create model
print("=> creating model '{}'".format(args.net))
net = model.Net(classnum=args.classnum, feature_dim=2, head=args.loss_type, radius=args.radius, sample_feat=args.sample_feat)
net = net.cuda()
print(net)
# build optimizer
optimizer = torch.optim.SGD(net.parameters(), lr=args.base_lr, momentum=0.9, weight_decay=5e-4)
if args.loss_type == 'a-softmax':
criterion = model.AngleLoss(LambdaMax=args.LambdaMax, gamma=args.gamma, power=args.power).cuda()
if args.loss_type == 'softmax' or args.loss_type == 'gaussian':
criterion = torch.nn.CrossEntropyLoss()
start_epoch = 0
best_prec1 = 0
# optionally resume from a pretrained model
if args.evaluate:
model_path = os.path.join(args.ckpt, 'checkpoints', args.evaluate)
checkpoint = torch.load(model_path)
epoch = int(checkpoint['epoch'])
net.load_state_dict(checkpoint['state_dict'])
val_feat, val_target, prec1 = validate(net, val_loader, criterion, epoch)
print('Prec1: {}'.format(prec1))
return
if args.resume:
model_path = os.path.join(args.ckpt, 'checkpoints', args.resume)
checkpoint = torch.load(model_path)
start_epoch = int(checkpoint['epoch'])
best_prec1 = float(checkpoint['best_prec1'])
net.load_state_dict(checkpoint['state_dict'])
optimizer.load_state_dict(checkpoint['optimizer'])
if args.pretrained:
start_epoch = 0
## start training
net.train()
freq = args.print_freq
end = time.time()
for epoch in range(start_epoch,args.epochs):
train_feat, train_target, train_prec1 = train(net, epoch, train_loader, args, criterion, optimizer)
val_feat, val_target, prec1 = validate(net, val_loader, criterion, epoch)
#pdb.set_trace()
if (epoch+1) % args.vis_freq == 0:
visualize(train_feat, train_target, val_feat, val_target, epoch, args, train_prec1, prec1)
is_best = prec1>best_prec1
best_prec1 = max(prec1, best_prec1)
save_checkpoint({
'epoch': epoch+1,
'state_dict': net.state_dict(),
'optimizer': optimizer.state_dict(),
'best_prec1': best_prec1
},args, is_best)
print('Best Prec1: {}'.format(best_prec1))
writer.close()