def main():
parser = argparse.ArgumentParser(description='sgd-ma')
parser.add_argument('dname', help='dataset name')
args = parser.parse_args()
# load the dataset
fpath = os.path.join(args.data_dir, f"{args.dname}.dat")
X, y = load_dat(fpath)
batch_size = int(np.sqrt(X.shape[0]) + 10)
eps = compute_epsilon(X, y, batch_size)
print(eps)
def main(args):
# load the dataset
dname = f"{args.dname}.dat"
fpath = os.path.join(args.data_dir, dname)
X, y = load_dat(fpath, normalize=args.norm, shuffle=args.shuffle)
N, dim = X.shape
# order of Renyi divergence
alpha = np.linspace(1.5, 1024, 1000)
delta = args.delta
sigma = np.array([0.005, 0.01, 0.025, 0.05, 0.1, 0.2, 0.3, 0.4, 0.6, 0.8])
sigma = np.flip(sigma, 0)
batch_size = args.batch_size
n_sigma = len(sigma)
cv_rep = 10
K = 5
n_rep = K * cv_rep
eps = np.zeros((n_sigma, n_rep))
acc = np.zeros_like(eps)
obj = np.zeros_like(eps)
j = 0
# task
if args.svm:
loss_func = hsvm_loss
grad_func = hsvm_grad
test_func = svm_test
y[y < 0.5] = -1.0
task = 'svm'
else:
loss_func = logit_loss
grad_func = logistic_grad
test_func = logistic_test
task = 'logres'
rkf = RepeatedKFold(n_splits=K, n_repeats=cv_rep)
for train_idx, test_idx in rkf.split(X):
train_X, train_y = X[train_idx, :], y[train_idx]
test_X, test_y = X[test_idx, :], y[test_idx]
noise = np.random.randn(dim)
# new recurrence relation
w, sens = sgd_recur(train_X,
train_y,
grad_func,
batch_size,
args.T,
args.L,
reg_coeff=args.mu,
R=args.norm,
init_step=args.init_step,
verbose=False)
sigma_sq = 2.0 * np.square(sigma)
eps[:, j] = epsilon_worst_case(sens[-1, :], alpha, sigma_sq, delta)
noisy_w = w[-1, :] + sigma[:, np.newaxis] * noise
acc[:, j] = test_func(noisy_w, test_X, test_y) * 100
obj[:, j] = loss_func(noisy_w, train_X, train_y, reg_coeff=args.mu)
j += 1
avg_acc = np.mean(acc, axis=1)
avg_eps = np.mean(eps, axis=1)
avg_obj = np.mean(obj, axis=1)
str_mu = "{0}".format(args.mu)[2:]
str_is = "{0}".format(args.init_step).replace('.', '').rstrip('0')
filename = "rsgdd_{5}_T{0}B{1}mu{2}IS{3}_{4}".format(
args.T, args.batch_size, str_mu, str_is, args.dname, task)
rs_dir = "./plot/results"
np.savetxt("{0}/{1}_eps.out".format(rs_dir, filename), avg_eps, fmt='%.5f')
np.savetxt("{0}/{1}_acc.out".format(rs_dir, filename), avg_acc, fmt='%.5f')
np.savetxt("{0}/{1}_obj.out".format(rs_dir, filename), avg_obj, fmt='%.5f')
parser.add_argument('dname', help='dataset name')
parser.add_argument('T', type=int, help='epoch')
parser.add_argument('rst', type=int, help='reset interval')
parser.add_argument('--data_dir', type=str, default=None)
parser.add_argument('--batch_size', type=int, default=4000)
parser.add_argument('--init_step', type=float, default=0.5)
parser.add_argument('--mu', type=float, default=0.001)
parser.add_argument('--L', type=float, default=1.81)
parser.add_argument('--norm', type=float, default=1.0)
parser.add_argument('--delta', type=float, default=1e-12)
args = parser.parse_args()
# load the dataset
fpath = os.path.join(args.data_dir, f"{args.dname}.dat")
X, y = load_dat(fpath, normalize=args.norm)
# y[y < 0.5] = -1.0
batch_size = args.batch_size
n_epoch = args.T
w, sens = sgd_restart(X,
y,
logistic_grad,
batch_size,
n_epoch,
args.L,
reg_coeff=args.mu,
reset_intvl=args.rst,
R=args.norm,
init_step=0.5,
verbose=True,
log_eta = logsumexp(expo, axis=0) - np.log(m)
return w, log_eta
if __name__ == "__main__":
parser = argparse.ArgumentParser(description='recursive mechanism')
parser.add_argument('dname', help='dataset name')
parser.add_argument('T', type=int, help='epoch')
parser.add_argument('--data_dir', type=str, default=None)
args = parser.parse_args()
# load the dataset
fpath = os.path.join(args.data_dir, f"{args.dname}.dat")
X, y = load_dat(fpath, minmax=(0, 1), bias_term=True)
# y[y < 0.5] = -1.0
w, sen = sgd_recur(X,
y,
logistic_grad,
4000,
args.T,
0.25,
reg_coeff=0.001,
init_step=0.5,
verbose=True,
loss_func=logistic_loss,
test_func=logistic_test)
acc = logistic_test(w[-1], X, y)
print("accuracy={}".format(acc))