def load_data():
train_data = dataset.Market_DataLoader(imgs_path=cfg.TRAIN.imgs_path,
pose_path=cfg.TRAIN.pose_path,
idx_path=cfg.TRAIN.idx_path,
transform=dataset.train_transform(),
loader=dataset.val_loader)
train_loader = Data.DataLoader(train_data,
batch_size=cfg.TRAIN.BATCH_SIZE,
shuffle=True,
num_workers=cfg.TRAIN.NUM_WORKERS,
drop_last=True)
val_data = dataset.Market_DataLoader(imgs_path=cfg.TRAIN.imgs_path,
pose_path=cfg.TRAIN.pose_path,
idx_path=cfg.TEST.idx_path,
transform=dataset.val_transform(),
loader=dataset.val_loader)
val_loader = Data.DataLoader(val_data,
batch_size=cfg.TEST.BATCH_SIZE,
shuffle=False,
num_workers=cfg.TRAIN.NUM_WORKERS)
train = [train_data, train_loader]
val = [val_data, val_loader]
return train, val
def create_data_loaders(fold, batch_size, workers):
train_ids, val_ids = dataset.get_split(fold)
labeled_size = len(train_ids)
unlabeled_size = 18000
sampler = dataset.TwoStreamBatchSampler(
range(labeled_size), # labeled ids
list(range(labeled_size,
labeled_size + unlabeled_size)), # unlabeled ids
batch_size, # total batch size (labeled + unlabeled)
LABELED_BATCH_SIZE # labeled batch size # TODO: was .5
)
train_loader = dataset.DataLoader(
dataset=dataset.MeanTeacherTGSDataset(
train_ids, transform=dataset.train_transform(), mode='train'),
# shuffle=True,
num_workers=workers,
batch_sampler=sampler,
pin_memory=torch.cuda.is_available())
valid_loader = dataset.make_loader(
val_ids,
transform=dataset.val_transform(),
shuffle=False,
# batch_size=len(device_ids),
batch_size=batch_size, # len(device_ids),
workers=workers)
return train_loader, valid_loader
def initialize_PN_GAN():
model_path = './G_12.pkl'
pn_gan = load_network(model_path)
pose_data = dataset.Pose_Loader(pose_path='./poses/',
transform=dataset.val_transform(),
loader=dataset.val_loader)
use_gpu = torch.cuda.is_available()
pin_memory = True if use_gpu else False
return pn_gan[0], DataLoader(pose_data,
batch_size=8,
shuffle=False,
num_workers=1,
pin_memory=pin_memory)
def load_data(samples):
val_data = dataset.Market_test(imgs_name=samples,
imgs_path=cfg.TEST.imgs_path,
pose_path=cfg.TEST.pose_path,
transform=dataset.val_transform(),
loader=dataset.val_loader)
val_loader = Data.DataLoader(val_data,
batch_size=cfg.TEST.BATCH_SIZE,
shuffle=False,
num_workers=cfg.TEST.BATCH_SIZE)
val = [val_data, val_loader]
return val
def predict_tta(model, ids, output, kind='test', batch_size=32, n_tta=5):
size = dataset.SIZE
base_transform = dataset.val_transform()
preds = np.zeros((1+n_tta, len(ids), dataset.NUM_CLASSES), dtype=np.float32)
preds[0] = predict(model, ids, transform=base_transform, kind=kind, batch_size=batch_size)
tta_transform = dataset.train_transform()
for tta_idx in range(1, n_tta):
preds[tta_idx] = predict(model, ids, transform=tta_transform, kind=kind, batch_size=batch_size)
mean_preds = np.mean(preds, axis=0)
np.save(output, mean_preds)
def main():
parser = argparse.ArgumentParser()
arg = parser.add_argument
arg('--name', type=str)
arg('--jaccard-weight', default=0.25, type=float)
arg('--device-ids',
type=str,
default='0',
help='For example 0,1 to run on two GPUs')
arg('--fold', type=int, help='fold', default=0)
arg('--output-dir', default='../data/runs', help='checkpoint root')
arg('--batch-size', type=int, default=32)
arg('--iter-size', type=int, default=1)
arg('--n-epochs', type=int, default=100)
arg('--lr', type=float, default=0.0001)
arg('--workers', type=int, default=4)
arg('--seed', type=int, default=0)
arg('--model', type=str, default=models.archs[0], choices=models.archs)
arg('--loss',
type=str,
default='focal',
choices=[
'focal', 'lovasz', 'bjd', 'bce_jaccard', 'bce_dice', 'cos_dice',
'hinge'
])
arg('--focal-gamma', type=float, default=.5)
arg('--num-channels', type=int, default=3)
arg('--weighted-sampler', action="store_true")
arg('--ignore-empty-masks', action='store_true')
arg('--remove-suspicious', action='store_true')
arg('--resume', action="store_true")
args = parser.parse_args()
random.seed(args.seed)
torch.manual_seed(args.seed)
if not args.name:
experiment = uuid.uuid4().hex
else:
experiment = args.name
output_dir = Path(args.output_dir) / experiment
output_dir.mkdir(exist_ok=True, parents=True)
output_dir.joinpath('params.json').write_text(
json.dumps(vars(args), indent=True, sort_keys=True))
# in case --resume is provided it will be loaded later
model = models.get_model(None, args.model)
# model = models.get_model(f"../data/runs/exp81/model_{args.fold}.pth", args.model)
if torch.cuda.is_available():
if args.device_ids:
device_ids = list(map(int, args.device_ids.split(',')))
else:
device_ids = None
model = nn.DataParallel(model, device_ids=device_ids).cuda()
train_ids, val_ids = dataset.get_split(args.fold)
cudnn.benchmark = True
train_loader = dataset.make_loader(
train_ids,
num_channels=args.num_channels,
transform=dataset.train_transform(),
shuffle=True,
weighted_sampling=args.weighted_sampler,
ignore_empty_masks=args.ignore_empty_masks,
remove_suspicious=args.remove_suspicious,
batch_size=args.batch_size,
workers=args.workers)
valid_loader = dataset.make_loader(
val_ids,
num_channels=args.num_channels,
transform=dataset.val_transform(),
shuffle=False,
#batch_size=len(device_ids),
batch_size=args.batch_size, # len(device_ids),
workers=args.workers)
# optimizer = Adam([p for p in model.parameters() if p.requires_grad], lr=args.lr)
optimizer = Adam(model.parameters(), lr=args.lr)
# loss = LossBinary(jaccard_weight=args.jaccard_weight)
# loss = LossBinaryMixedDiceBCE(dice_weight=0.5, bce_weight=0.5)
if args.loss == 'focal':
loss = FocalLoss(args.focal_gamma)
elif args.loss == 'lovasz':
loss = LossLovasz()
elif args.loss == 'bjd':
loss = BCEDiceJaccardLoss({'bce': 0.25, 'jaccard': None, 'dice': 0.75})
elif args.loss == 'bce_jaccard':
loss = LossBinary(args.jaccard_weight)
elif args.loss == 'bce_dice':
import loss2
bce_weight = 1
dice_weight = 2
loss = loss2.make_loss(bce_weight, dice_weight)
elif args.loss == 'cos_dice':
import loss2
loss = loss2.make_cos_dice_loss()
elif args.loss == 'hinge':
loss = LossHinge()
else:
raise NotImplementedError
validation = validation_binary
scheduler = ReduceLROnPlateau(optimizer,
verbose=True,
min_lr=1e-7,
factor=0.5)
snapshot = utils.fold_snapshot(output_dir,
args.fold) if args.resume else None
utils.train(experiment=experiment,
output_dir=output_dir,
optimizer=optimizer,
args=args,
model=model,
criterion=loss,
scheduler=scheduler,
train_loader=train_loader,
valid_loader=valid_loader,
validation=validation,
fold=args.fold,
batch_size=args.batch_size,
n_epochs=args.n_epochs,
snapshot=snapshot,
iter_size=args.iter_size)
def main():
args = build_train_args()
random.seed(args.seed)
torch.manual_seed(args.seed)
experiment = args.name if args.name else uuid.uuid4().hex
output_dir = Path(args.output_dir) / experiment
output_dir.mkdir(exist_ok=True, parents=True)
output_dir.joinpath('params.json').write_text(
json.dumps(vars(args), indent=True, sort_keys=True))
# this is different than --resume, which loads optimizer state as well
initial_model_path = os.path.join(
output_dir, f"model_{args.fold}.pth") if args.load_weights else None
model = models.get_model(initial_model_path, args.model)
train_ids, val_ids = dataset.get_split(args.fold)
cudnn.benchmark = True
train_loader = dataset.make_loader(train_ids,
num_channels=args.num_channels,
transform=dataset.train_transform(),
shuffle=True,
weighted_sampling=args.weighted_sampler,
batch_size=args.batch_size,
workers=args.workers)
valid_loader = dataset.make_loader(
val_ids,
num_channels=args.num_channels,
transform=dataset.val_transform(),
shuffle=False,
weighted_sampling=False,
batch_size=args.batch_size, # len(device_ids),
workers=args.workers)
# train last layer only
# for m in model.children():
# if m in [model.fc]:
# continue
# for param in m.parameters():
# param.requires_grad = False
# set all layers except fc to lr=0
#fc_params = list(map(id, model.fc.parameters()))
#base_params = filter(lambda p: id(p) not in fc_params, model.parameters())
#optimizer = torch.optim.Adam([
# {'params': base_params, 'lr': args.lr * 0.001},
# {'params': model.fc.parameters(), 'lr': args.lr}
#], lr=args.lr * 0.001)
optimizer = torch.optim.Adam(model.parameters(), lr=args.lr)
# optimizer = torch.optim.SGD(model.parameters(), lr=0.03, momentum=0.9, weight_decay=1e-4)
# criterion = FocalLoss(gamma=args.focal_gamma)
criterion = nn.BCEWithLogitsLoss().cuda()
# criterion = losses.f1_loss
validation_fn = validation_multi
#scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(
# optimizer, mode='max', verbose=True, min_lr=1e-7, factor=0.5, patience=5)
scheduler = torch.optim.lr_scheduler.StepLR(optimizer,
step_size=10,
gamma=0.1)
snapshot = utils.fold_snapshot(output_dir,
args.fold) if args.resume else None
device_ids = list(map(
int, args.device_ids.split(','))) if args.device_ids else None
wrapped_model = nn.DataParallel(model, device_ids=device_ids).cuda()
# unfreeze
# for m in model.children():
# for param in m.parameters():
# param.requires_grad = True
# fc1_params = list(map(id, model.fc1.parameters()))
# fc2_params = list(map(id, model.fc2.parameters()))
# base_params = filter(lambda p: id(p) not in fc1_params and id(p) not in fc2_params, model.parameters())
#
# # fc at args.lr and rest with args.lr * 0.1
# optimizer = torch.optim.Adam([
# {'params': base_params},
# {'params': model.fc1.parameters(), 'lr': args.lr},
# {'params': model.fc2.parameters(), 'lr': args.lr},
# ], lr=args.lr * 0.1)
# aternatively, add the unfrozen fc2 weight to the current optimizer
# optimizer.add_param_group({'params': net.fc2.parameters()})
utils.train(experiment=experiment,
output_dir=output_dir,
optimizer=optimizer,
args=args,
model=wrapped_model,
criterion=criterion,
scheduler=scheduler,
train_loader=train_loader,
valid_loader=valid_loader,
validation_fn=validation_fn,
fold=args.fold,
batch_size=args.batch_size,
n_epochs=args.n_epochs,
snapshot=snapshot,
iter_size=args.iter_size)
return criterion
#%%
if __name__ == '__main__':
sys.stdout = logger.Logger('./log_GAN_ep2.txt')
train_data = dataset.Market_DataLoader(imgs_path=cfg.TRAIN.imgs_path, pose_path=cfg.TRAIN.pose_path, idx_path=cfg.TRAIN.idx_path,
transform=dataset.train_transform(), img_loader=dataset.val_loader, pose_loader=dataset.pose_loader)
train_loader = Data.DataLoader(train_data, batch_size=cfg.TRAIN.BATCH_SIZE, shuffle=True, num_workers=cfg.TRAIN.NUM_WORKERS, drop_last=True)
val_data = dataset.Market_DataLoader(imgs_path=cfg.TRAIN.imgs_path, pose_path=cfg.TRAIN.pose_path, idx_path=cfg.TEST.idx_path,
transform=dataset.val_transform(), img_loader=dataset.val_loader, pose_loader=dataset.pose_loader)
val_loader = Data.DataLoader(val_data, batch_size=cfg.TEST.BATCH_SIZE, shuffle=False,
num_workers=cfg.TRAIN.NUM_WORKERS)
train_file = [train_data, train_loader]
val_file = [val_data, val_loader]
# train_file, val_file = load_data()
nets = load_network()
optimizers, schedulers, summary = Optimizer(nets)
criterion = loss_func()
# train(train_file, val_file, nets, optimizers, schedulers, summary, criterion)
loss = loss = LossLovasz()
validation = validation_binary
scheduler = ReduceLROnPlateau(optimizer,
verbose=True,
min_lr=1e-7,
factor=0.5,
patience=5)
output_dir = Path('tmp')
output_dir.mkdir(exist_ok=True, parents=True)
train_ids, val_ids = dataset.get_split(fold)
train_loader = DataLoader(dataset=TGSDataset(
train_ids,
num_channels=3,
transform=dataset.val_transform(),
mode='train'),
shuffle=True,
num_workers=4,
batch_size=24,
sampler=None,
pin_memory=torch.cuda.is_available())
val_loader = DataLoader(dataset=TGSDataset(
val_ids,
num_channels=3,
transform=dataset.val_transform(),
mode='train'),
shuffle=False,
num_workers=4,
batch_size=1,
sampler=None,