parser.add_argument('--max_minibatch', type=int, default=9)
parser.add_argument('--num_samples', type=int, default=9)
args = parser.parse_args()
### ---- initialize necessary --- ###
# (1) pretrained generative model
model = BigGAN().cuda().eval()
# (2) variable creator
var_manager = VariableManager()
# (3) default l1 + lpips loss function
loss_fn = LF.ProjectionLoss()
target = image.read(args.fp, as_transformed_tensor=True, im_size=256)
weight = image.read(args.mask_fp, as_transformed_tensor=True, im_size=256)
weight = ((weight + 1.) / 2.).clamp_(0.3, 1.0)
class_lbl = 153
fn = args.fp.split('/')[-1].split('.')[0]
save_dir = f'./results/biggan_256/hybridng_{fn}'
var_manager = VariableManager()
loss_fn = LF.ProjectionLoss()
# (4) define input output variable structure. the variable name must match
# the argument name of the model and loss function call
var_manager.register(
variable_name='z',
parser.add_argument('--latent_noise', type=float, default=0.05)
parser.add_argument('--truncate', type=float, default=2.0)
parser.add_argument('--make_video', action='store_true')
parser.add_argument('--num_samples', type=int, default=4)
parser.add_argument('--max_minibatch', type=int, default=9)
args = parser.parse_args()
### ---- initialize --- ###
model = StyleGAN2(model='cars', search='z')
filename = './images/car-example.png'
target = image.read(filename, as_transformed_tensor=True, im_size=512,
transform_style='stylegan')
# we apply a mask since the generated resolution is 384 x 512
loss_mask = torch.zeros((3, 512, 512))
loss_mask[:, 64:-64, :].data += 1.0
weight = loss_mask
fn = filename.split('/')[-1].split('.')[0]
save_dir = f'./results/stylegan2_cars/hybridng_{args.ng_method}_{fn}'
model = StyleGAN2(search='z')
model = nn.DataParallel(model)
