def main():
parser = argparse.ArgumentParser(
description='GAN-based unsupervised segmentation train')
parser.add_argument('--out', type=str, required=True)
parser.add_argument('--gan_weights', type=str, default=WEIGHTS['BigGAN'])
parser.add_argument('--deformator_weights', type=str, required=True)
parser.add_argument('--deformator_type',
type=str,
choices=DEFORMATOR_TYPE_DICT.keys(),
required=True)
parser.add_argument('--background_dim', type=int, required=True)
parser.add_argument('--classes',
type=int,
nargs='*',
default=list(range(1000)))
parser.add_argument('--device', type=int, default=0)
parser.add_argument('--seed', type=int, default=2)
parser.add_argument('--val_images_dir', type=str, default=None)
parser.add_argument('--val_masks_dir', type=str, default=None)
for key, val in SegmentationTrainParams().__dict__.items():
val_type = type(val) if key is not 'synthezing' else str
parser.add_argument('--{}'.format(key), type=val_type, default=None)
args = parser.parse_args()
torch.random.manual_seed(args.seed)
torch.cuda.set_device(args.device)
# save run p
save_command_run_params(args)
if len(args.classes) == 0:
print('using all ImageNet')
args.classes = list(range(1000))
G = make_big_gan(args.gan_weights, args.classes).eval().cuda()
deformator = LatentDeformator(
G.dim_z, type=DEFORMATOR_TYPE_DICT[args.deformator_type])
deformator.load_state_dict(
torch.load(args.deformator_weights, map_location=torch.device('cpu')))
deformator.eval().cuda()
model = UNet().train().cuda()
train_params = SegmentationTrainParams(**args.__dict__)
print(f'run train with p: {train_params.__dict__}')
train_segmentation(G,
deformator,
model,
train_params,
args.background_dim,
args.out,
val_dirs=[args.val_images_dir, args.val_masks_dir])
def load_from_dir(root_dir, model_index=None, G_weights=None, shift_in_w=True):
args = json.load(open(os.path.join(root_dir, 'args.json')))
args['w_shift'] = shift_in_w
models_dir = os.path.join(root_dir, 'models')
if model_index is None:
models = os.listdir(models_dir)
model_index = max([
int(name.split('.')[0].split('_')[-1]) for name in models
if name.startswith('deformator')
])
if G_weights is None:
G_weights = args['gan_weights']
if G_weights is None or not os.path.isfile(G_weights):
print('Using default local G weights')
G_weights = WEIGHTS[args['gan_type']]
if isinstance(G_weights, dict):
G_weights = G_weights[str(args['resolution'])]
if 'resolution' not in args.keys():
args['resolution'] = 128
G = load_generator(args, G_weights)
deformator = LatentDeformator(
shift_dim=G.dim_shift,
input_dim=args['directions_count']
if 'directions_count' in args.keys() else None,
out_dim=args['max_latent_dim']
if 'max_latent_dim' in args.keys() else None,
type=DEFORMATOR_TYPE_DICT[args['deformator']])
if 'shift_predictor' not in args.keys(
) or args['shift_predictor'] == 'ResNet':
shift_predictor = LatentShiftPredictor(G.dim_shift)
elif args['shift_predictor'] == 'LeNet':
shift_predictor = LeNetShiftPredictor(
G.dim_shift, 1 if args['gan_type'] == 'SN_MNIST' else 3)
deformator_model_path = os.path.join(
models_dir, 'deformator_{}.pt'.format(model_index))
shift_model_path = os.path.join(
models_dir, 'shift_predictor_{}.pt'.format(model_index))
if os.path.isfile(deformator_model_path):
deformator.load_state_dict(
torch.load(deformator_model_path,
map_location=torch.device('cpu')))
if os.path.isfile(shift_model_path):
shift_predictor.load_state_dict(
torch.load(shift_model_path, map_location=torch.device('cpu')))
setattr(
deformator, 'annotation',
load_human_annotation(os.path.join(root_dir, HUMAN_ANNOTATION_FILE)))
return deformator.eval().cuda(), G.eval().cuda(), shift_predictor.eval(
).cuda()
def load_from_dir(root_dir, model_index=None, G_weights=None, verbose=False):
args = json.load(open(os.path.join(root_dir, 'args.json')))
models_dir = os.path.join(root_dir, 'models')
if model_index is None:
models = os.listdir(models_dir)
model_index = max([
int(name.split('.')[0].split('_')[-1]) for name in models
if name.startswith('deformator')
])
if verbose:
print('using max index {}'.format(model_index))
if G_weights is None:
G_weights = args['gan_weights']
if G_weights is None or not os.path.isfile(G_weights):
if verbose:
print('Using default local G weights')
G_weights = WEIGHTS[args['gan_type']]
if args['gan_type'] == 'BigGAN':
G = make_big_gan(G_weights, args['target_class']).eval()
elif args['gan_type'] in ['ProgGAN', 'PGGAN']:
G = make_proggan(G_weights)
else:
G = make_external(G_weights)
deformator = LatentDeformator(
G.dim_z, type=DEFORMATOR_TYPE_DICT[args['deformator']])
if 'shift_predictor' not in args.keys(
) or args['shift_predictor'] == 'ResNet':
shift_predictor = ResNetShiftPredictor(G.dim_z)
elif args['shift_predictor'] == 'LeNet':
shift_predictor = LeNetShiftPredictor(
G.dim_z, 1 if args['gan_type'] == 'SN_MNIST' else 3)
deformator_model_path = os.path.join(
models_dir, 'deformator_{}.pt'.format(model_index))
shift_model_path = os.path.join(
models_dir, 'shift_predictor_{}.pt'.format(model_index))
if os.path.isfile(deformator_model_path):
deformator.load_state_dict(torch.load(deformator_model_path))
if os.path.isfile(shift_model_path):
shift_predictor.load_state_dict(torch.load(shift_model_path))
# try to load dims annotation
directions_json = os.path.join(root_dir, 'directions.json')
if os.path.isfile(directions_json):
with open(directions_json, 'r') as f:
directions_dict = json.load(f, object_pairs_hook=OrderedDict)
setattr(deformator, 'directions_dict', directions_dict)
return deformator.eval().cuda(), G.eval().cuda(), shift_predictor.eval(
).cuda()
def main():
parser = argparse.ArgumentParser(
description='GAN-based unsupervised segmentation train')
parser.add_argument('--args',
type=str,
default=None,
help='json with all arguments')
parser.add_argument('--out', type=str, required=True)
parser.add_argument('--gan_weights', type=str, default=WEIGHTS['BigGAN'])
parser.add_argument('--deformator_weights', type=str, required=True)
parser.add_argument('--deformator_type',
type=str,
choices=DEFORMATOR_TYPE_DICT.keys(),
required=True)
parser.add_argument('--background_dim', type=int, required=True)
parser.add_argument('--classes',
type=int,
nargs='*',
default=list(range(1000)))
parser.add_argument('--device', type=int, default=0)
parser.add_argument('--seed', type=int, default=2)
parser.add_argument('--val_images_dir', type=str)
parser.add_argument('--val_masks_dir', type=str)
for key, val in SegmentationTrainParams().__dict__.items():
parser.add_argument('--{}'.format(key), type=type(val), default=None)
args = parser.parse_args()
torch.random.manual_seed(args.seed)
torch.cuda.set_device(args.device)
if args.args is not None:
with open(args.args) as args_json:
args_dict = json.load(args_json)
args.__dict__.update(**args_dict)
# save run params
if not os.path.isdir(args.out):
os.makedirs(args.out)
with open(os.path.join(args.out, 'args.json'), 'w') as args_file:
json.dump(args.__dict__, args_file)
with open(os.path.join(args.out, 'command.sh'), 'w') as command_file:
command_file.write(' '.join(sys.argv))
command_file.write('\n')
if len(args.classes) == 0:
print('using all ImageNet')
args.classes = list(range(1000))
G = make_big_gan(args.gan_weights, args.classes).eval().cuda()
deformator = LatentDeformator(
G.dim_z, type=DEFORMATOR_TYPE_DICT[args.deformator_type])
deformator.load_state_dict(
torch.load(args.deformator_weights, map_location=torch.device('cpu')))
deformator.eval().cuda()
model = UNet().train().cuda()
train_params = SegmentationTrainParams(**args.__dict__)
print('run train with params: {}'.format(train_params.__dict__))
train_segmentation(G, deformator, model, train_params, args.background_dim,
args.out)
if args.val_images_dir is not None:
evaluate(model, args.val_images_dir, args.val_masks_dir,
os.path.join(args.out, 'score.json'), 128)