return factor * base_lr
def learning_rate_schedule_tr(base_lr, epoch, total_epochs):
alpha = epoch / total_epochs
if alpha <= 0.3:
factor = 1.0
elif alpha <= 0.6:
factor = 1.0 - (alpha - 0.6) / 0.4 * 0.99
else:
factor = 0.01
return factor * base_lr
criterion = F.cross_entropy
regularizer = None if args.curve is None else curves.l2_regularizer(args.wd)
optimizer = torch.optim.SGD(
filter(lambda param: param.requires_grad, model.parameters()),
lr=args.lr,
momentum=args.momentum,
weight_decay=args.wd if args.curve is None else 0.0)
checkpoint = torch.load(args.ckpt[0])
record_train = []
record_test = []
# for i, ckp in enumerate(args.ckpt):
for i in range(args.num_scale):
# checkpoint = torch.load(ckp)
print('next_scale')
key_weight_name = []
key_bias_name = []
architecture = getattr(models, args.model)
curve = getattr(curves, args.curve)
model = curves.CurveNet(
num_classes,
curve,
architecture.curve,
args.num_bends,
architecture_kwargs=architecture.kwargs,
)
model.cuda()
checkpoint = torch.load(args.ckpt)
model.load_state_dict(checkpoint['model_state'])
criterion = F.cross_entropy
regularizer = curves.l2_regularizer(args.wd)
T = args.num_points
ts = np.linspace(0.0, 1.0, T)
tr_loss = np.zeros(T)
tr_nll = np.zeros(T)
tr_acc = np.zeros(T)
te_loss = np.zeros(T)
te_nll = np.zeros(T)
te_acc = np.zeros(T)
tr_err = np.zeros(T)
te_err = np.zeros(T)
dl = np.zeros(T)
previous_weights = None