def _build_model(self):
"""Initializes the generator and discriminator networks.
"""
self.generator, self.discriminator = networks.build_networks(
gen=self.gen_net,
disc=self.disc_net,
ngf=64,
ndf=self.disc_size,
imsize=self.imsize,
nc=self.nc,
k=self.noise_dim,
z=self.dout_dim,
bn=self.batch_norm)
print('Generator', self.generator)
print('Discriminator', self.discriminator)
self.generator.apply(networks.weights_init)
self.discriminator.apply(networks.weights_init)
self.generator.cuda()
self.discriminator.cuda()
cudnn.benchmark = True
# Other options can change with every run
parser.add_argument('--batch_size', type=int, default=64, help='Batch size [default: 64]')
parser.add_argument('--fold', type=str, default='train', help='Fold [default: train]')
parser.add_argument('--start_epoch', type=int, help='Epoch to start from (defaults to most recent epoch)')
parser.add_argument('--count', type=int, default=1, help='Number of counterfactuals to generate')
options = vars(parser.parse_args())
sys.path.append(os.path.dirname(os.path.dirname(__file__)))
from dataloader import CustomDataloader
import counterfactual
from networks import build_networks
from options import load_options
# TODO: Right now, to edit cf_speed et al, you need to edit params.json
start_epoch = options['start_epoch']
options = load_options(options)
options['epoch'] = start_epoch
dataloader = CustomDataloader(**options)
# Batch size must be large enough to make a square grid visual
options['batch_size'] = dataloader.num_classes + 1
networks = build_networks(dataloader.num_classes, **options)
for i in range(options['count']):
counterfactual.generate_counterfactual(networks, dataloader, **options)
dataloader_train = torch.utils.data.DataLoader(
train_dataset,
batch_size=options['batch_size'],
shuffle=False,
num_workers=8,
pin_memory=True,
drop_last=False)
dataloader_off = torch.utils.data.DataLoader(dataset_off,
batch_size=options['batch_size'],
shuffle=False,
num_workers=8,
pin_memory=True,
drop_last=False)
networks = build_networks(**options)
new_results = evaluate_with_comparison(networks,
dataloader,
comparison_dataloader=dataloader_off,
dataloader_train=dataloader_train,
**options)
pprint(new_results)
if not os.path.exists('evaluation'):
os.mkdir('evaluation')
result = {options['mode']: new_results}
result_dir = os.path.join('evaluation',
'result_{}.json'.format(options['image_size']))
if os.path.exists(result_dir):
old_result = json.load(open(result_dir))
help='Size of the image.')
parser.add_argument(
'--png',
action='store_true',
help='Whether to generate a png '
'(overrides num_samples and generates a 8x8 grid of images).')
parser.add_argument('--o',
'--out_dir',
type=str,
default='./',
help='Output directory.')
args = parser.parse_args()
generator, _ = networks.build_networks(gen=args.gen,
imsize=args.imsize,
k=args.k)
generator.load_state_dict(torch.load(args.ckpt))
print('[*] Generator loaded')
if not os.path.exists(args.o):
os.makedirs(args.o)
generator.cuda()
batch_size = 64 if args.png else 256
noise = torch.cuda.FloatTensor(batch_size, args.k, 1, 1)
n_loops = 1 if args.png else args.n // batch_size + 1
samples = []
for i in tqdm(range(n_loops)):
noise.normal_(0, 1)
with torch.no_grad():
images = generator(noise).detach().cpu().numpy()
if args.png: # Generate 8x8 if png
