def add_additional_params_to_args(args):
dataset = args.src_dataset if "src_dataset" in args.__dict__.keys(
) else args.tgt_dataset
args.n_class = get_n_class(dataset)
args.machine = os.uname()[1]
return args
def add_additional_params_to_args(args):
dataset = args.src_dataset if "src_dataset" in args.__dict__.keys() else args.tgt_dataset
args.n_class = get_n_class(dataset)
args.machine = os.uname()[1]
def add_img_shape_to_args(args):
if "src_dataset" in args.__dict__.keys() and "train_img_shape" in args.__dict__.keys():
if args.train_img_shape is None:
args.train_img_shape = get_img_shape(args.src_dataset, is_train=True)
print("args.train_img_shape is set to %s" % args.train_img_shape)
if "tgt_dataset" in args.__dict__.keys() and "test_img_shape" in args.__dict__.keys():
if args.test_img_shape is None:
args.test_img_shape = get_img_shape(args.tgt_dataset, is_train=False)
print("args.test_img_shape is set to %s" % args.test_img_shape)
return args
args = add_img_shape_to_args(args)
return args
parser.add_argument("--add_bg_loss",
action="store_true",
help='whether you add background loss or not')
parser.add_argument("--adjust_lr",
action="store_true",
help='whether you change lr')
parser.add_argument("--max_iter", type=int, default=5000)
parser.add_argument(
"--fix_bn",
action="store_true",
help='whether you fix the paramters of batch normalization layer')
args = parser.parse_args()
check_src_tgt_ok(args.src_dataset, args.tgt_dataset)
args.n_class = get_n_class(args.src_dataset, args.tgt_dataset)
weight = torch.ones(args.n_class)
if args.loss_weights_file:
import pandas as pd
loss_df = pd.read_csv(args.loss_weights_file)
loss_df.sort_values("class_id", inplace=True)
weight *= torch.FloatTensor(loss_df.weight.values)
if not args.add_bg_loss:
weight[args.n_class - 1] = 0 # Ignore background loss
# print ("loss weight %s" % weight)
if args.net == "fcn":