with open(os.path.join(args.dir, 'command.sh'), 'w') as f:
f.write(' '.join(sys.argv))
f.write('\n')
torch.backends.cudnn.benchmark = True
torch.manual_seed(args.seed)
torch.cuda.manual_seed(args.seed)
print('Using model %s' % args.model)
model_cfg = getattr(models, args.model)
print('Loading dataset %s from %s' % (args.dataset, args.data_path))
loaders, num_classes = data.loaders(args.dataset,
args.data_path,
args.batch_size,
args.num_workers,
model_cfg.transform_train,
model_cfg.transform_test,
use_validation=not args.use_test,
split_classes=args.split_classes)
print('Preparing model')
print(*model_cfg.args)
model = model_cfg.base(*model_cfg.args,
num_classes=num_classes,
**model_cfg.kwargs)
model.cuda()
swag_model = SWAG(model_cfg.base,
num_classes=num_classes,
subspace_type='pca',
from swag import data, models, utils, losses
import numpy as np
loaders, num_classes = data.loaders("CIFAR10",
"~/datasets/",
1000,
4,
None,
None,
use_validation=False,
split_classes=None,
shuffle_train=False)
train_x = loaders['train'].dataset.data
train_y = np.array(loaders['train'].dataset.targets)
np.save("train_x", train_x)
np.save("train_y", train_y)
np.random.shuffle(train_y)
np.save("shuffled_train_y", train_y)
test_x = loaders['test'].dataset.data
test_y = np.array(loaders['test'].dataset.targets)
np.save("test_x", test_x)
np.save("test_y", test_y)
choices=["cross_entropy", "aleatoric"],
default="cross_entropy")
args = parser.parse_args()
torch.backends.cudnn.benchmark = True
torch.manual_seed(args.seed)
torch.cuda.manual_seed(args.seed)
model_cfg = getattr(models, "FCDenseNet67")
loaders, num_classes = data.loaders(
"CamVid",
args.data_path,
args.batch_size,
4,
ft_batch_size=1,
transform_train=model_cfg.transform_train,
transform_test=model_cfg.transform_test,
joint_transform=model_cfg.joint_transform,
ft_joint_transform=model_cfg.ft_joint_transform,
target_transform=model_cfg.target_transform,
)
# criterion = nn.NLLLoss(weight=camvid.class_weight[:-1].cuda(), reduction='none').cuda()
if args.loss == "cross_entropy":
criterion = losses.seg_cross_entropy
else:
criterion = losses.seg_ale_cross_entropy
# construct and load model
if args.swa_resume is not None:
checkpoint = torch.load(args.swa_resume)
print('Preparing directory %s' % args.dir)
os.makedirs(args.dir, exist_ok=True)
with open(os.path.join(args.dir, 'command.sh'), 'w') as f:
f.write(' '.join(sys.argv))
f.write('\n')
print('Using model %s' % args.model)
model_cfg = getattr(models, args.model)
loaders, num_classes = data.loaders(
args.dataset,
args.data_path,
args.batch_size,
args.num_workers,
ft_batch_size=args.ft_batch_size,
transform_train=model_cfg.transform_train,
transform_test=model_cfg.transform_test,
joint_transform=model_cfg.joint_transform,
ft_joint_transform=model_cfg.ft_joint_transform,
target_transform=model_cfg.target_transform)
print('Beginning with cropped images')
train_loader = loaders['train']
print('Preparing model')
model = model_cfg.base(*model_cfg.args,
num_classes=num_classes,
**model_cfg.kwargs,
use_aleatoric=args.loss == 'aleatoric')
model.cuda()
model.apply(train_utils.weights_init)
f.write(' '.join(sys.argv))
f.write('\n')
torch.backends.cudnn.benchmark = True
torch.manual_seed(args.seed)
torch.cuda.manual_seed(args.seed)
print('Using model %s' % args.model)
model_cfg = getattr(models, args.model)
print('Loading dataset %s from %s' % (args.dataset, args.data_path))
loaders, num_classes = data.loaders(args.dataset,
args.data_path,
args.batch_size,
args.num_workers,
transform_train=model_cfg.transform_test,
transform_test=model_cfg.transform_test,
shuffle_train=False,
use_validation=not args.use_test,
split_classes=args.split_classes)
loaders_bn, _ = data.loaders(args.dataset,
args.data_path,
args.batch_size,
args.num_workers,
transform_train=model_cfg.transform_train,
transform_test=model_cfg.transform_test,
shuffle_train=True,
use_validation=not args.use_test,
split_classes=args.split_classes)
