def download_dataset(dataset):
cfg = get_cfg()
cfg.cls.dataset.name = dataset
dsname = dataset.lower()
cfg.cls.dataset.root = f"{settings.ROOT_PATH}/data/{dsname}/"
cfg.cls.dataset.download = True
get_dataset(cfg, 'cls')
def test_disent_over_epochs(cfg):
epoch_num_list = [0, 10, 50, 100, 120, 150]
for epoch_num in epoch_num_list:
cfg.disent.epoch_num = epoch_num
test_disent(cfg)
def plot_th_tensor(ax, i, j, dec_ij):
dec_ij = dec_ij.to('cpu').detach().numpy()[0, 0]
ax[i, j].imshow(dec_ij, cmap='Greys_r', interpolation=None)
if __name__ == "__main__":
cfg = get_cfg()
cfg.exp_name = "static_noise"
cfg.disent = edict()
cfg.disent.epochs = 5000
cfg.disent.load = False
cfg.disent.epoch_num = 0
model_path = Path(f"{settings.ROOT_PATH}/output/disent/{cfg.exp_name}/")
optim_path = Path(
f"{settings.ROOT_PATH}/output/disent/{cfg.exp_name}/optim/")
if not model_path.exists(): model_path.mkdir(parents=True)
cfg.disent.model_path = model_path
cfg.disent.optim_path = optim_path
cfg.disent.workers = 1
def get_denoising_cfg(args):
cfg = get_cfg()
if args.new is True: setup_new_exp(args)
cfg.cl.device = torch.device("cuda:{}".format(args.gpuid))
cfg.cls.device = cfg.cl.device
# set the name
cfg.exp_name = args.name
if cfg.exp_name is None:
cfg.exp_name = str(uuid.uuid4())
cfg.disent = edict()
cfg.disent.epochs = 1200
cfg.disent.device = cfg.cl.device
cfg.disent.load = args.epoch_num > 0
cfg.disent.epoch_num = args.epoch_num
cfg.disent.dataset = edict()
cfg.disent.dataset.root = f"{settings.ROOT_PATH}/data/"
cfg.disent.dataset.n_classes = 10
cfg.disent.dataset.name = args.dataset
cfg.disent.noise_level = args.noise_level
cfg.disent.N = args.N
cfg.disent.img_loss_type = args.img_loss_type
cfg.disent.share_enc = args.share_enc
cfg.disent.hyper_h = args.hyper_h
cfg.disent.lr = edict()
cfg.disent.lr.start = args.lr_start
cfg.disent.lr.policy = args.lr_policy
cfg.disent.lr.params = args.lr_params
dsname = cfg.disent.dataset.name.lower()
model_path = Path(
f"{settings.ROOT_PATH}/output/disent/{dsname}/{cfg.exp_name}")
optim_path = Path(
f"{settings.ROOT_PATH}/output/disent/{dsname}/{cfg.exp_name}/optim/")
if not model_path.exists(): model_path.mkdir(parents=True)
cfg.disent.model_path = model_path
cfg.disent.optim_path = optim_path
cfg.disent.workers = 1
cfg.disent.batch_size = args.batch_size
cfg.disent.global_step = 0
cfg.disent.device = cfg.cl.device
cfg.disent.current_epoch = 0
cfg.disent.checkpoint_interval = 1
cfg.disent.test_interval = 5
cfg.disent.log_interval = 50
cfg.disent.random_crop = True
if cfg.disent.dataset.name.lower() == "mnist":
cfg.disent.n_channels = 1
else:
cfg.disent.n_channels = 3
cfg.disent.agg_enc_fxn = 'mean'
if cfg.disent.share_enc is False:
cfg.disent.agg_enc_fxn = 'id'
# info = {'noise':args.noise_level,'N':args.N,
# 'dataset':args.dataset,'batch_size':args.batch_size}
# write_settings(cfg.exp_name,info)
# print(info)
return cfg