def create_model(opt):
arch = opt.arch
if opt.data == 'cifar100':
model, _ = inversefed.construct_model(arch,
num_classes=100,
num_channels=3)
elif opt.data == 'FashionMinist':
model, _ = inversefed.construct_model(arch,
num_classes=10,
num_channels=1)
return model
**setup)[:, None, None]
ds = torch.as_tensor(
getattr(inversefed.consts, f'{args.dataset.lower()}_std'),
**setup)[:, None, None]
if args.dataset == 'ImageNet':
if args.model == 'ResNet152':
model = torchvision.models.resnet152(pretrained=args.trained_model)
elif args.model == 'ResNet50':
model = torchvision.models.resnet50(pretrained=args.trained_model)
else:
model = torchvision.models.resnet18(pretrained=args.trained_model)
model_seed = None
else:
model, model_seed = inversefed.construct_model(args.model,
num_classes=10,
num_channels=3)
model.to(**setup)
model.eval()
# Sanity check: Validate model accuracy
training_stats = defaultdict(list)
# inversefed.training.training_routine.validate(model, loss_fn, validloader, defs, setup, training_stats)
# name, format = loss_fn.metric()
# print(f'Val loss is {training_stats["valid_losses"][-1]:6.4f}, Val {name}: {training_stats["valid_" + name][-1]:{format}}.')
# Choose example images from the validation set or from third-party sources
if args.num_images == 1:
if args.target_id == -1: # demo image
# Specify PIL filter for lower pillow versions
ground_truth = torch.as_tensor(