for hetero_lv in hetero_lv_list:
LSGD_train_result_sub = []
LSGD_est_result_sub = []
LSGD_time_sub = []
lambda_param = 1e-2 / hetero_lv # Penalty parameter
np.random.seed(111)
beta_mu = np.random.normal(loc=0.0, scale=hetero_lv, size=M) # means of shift for each devices
beta_true_list = [] # list of true local parameters
for m in range(M):
shift = np.random.uniform(beta_mu[m]-0.01, beta_mu[m]+0.01, size=d_gl)
beta_true_list.append(w_true + shift)
data_all = data_generate_fn(n, M, d_gl, beta_true_list)
for i_repeat in range(n_repeat):
# Parameters of LSGD_PFL
avg_period = 5
w0 = np.zeros(d_gl)
beta0_list = []
for m in range(M):
beta0_list.append(np.zeros(d_gl))
eta_LSGD = 0.01
# train by LSGD_PFL
start_time = time.time()
print("hetero_lv: {} | Repeat: {} | Training by LSGD_PFL".format(hetero_lv, i_repeat+1))
loss_train_LSGD, loss_est_LSGD, w_LSGD, beta_list_LSGD = LSGD_PFL(data_all, loss_fun_MX2, fmj_grads_MX2, eta_LSGD, n_commun,
avg_period, est_err_MX2, w_true, beta_true_list, w0, beta0_list,
start_time_total = time.time()
for hetero_lv in hetero_lv_list:
LSGD_train_result_sub = []
LSGD_est_result_sub = []
LSGD_time_sub = []
np.random.seed(111)
beta_mu = np.random.normal(loc=0.0, scale=hetero_lv,
size=M) # means of shift for each devices
beta_true_list = [] # list of true local parameters
for m in range(M):
beta_true_list.append(
np.random.uniform(beta_mu[m] - 0.01, beta_mu[m] + 0.01, size=d_l))
data_all = data_generate_fn(n, w_true, beta_true_list)
for i_repeat in range(n_repeat):
# Parameters of LSGD_PFL
avg_period = 5
w0 = np.zeros(d_g)
beta0_list = []
for m in range(M):
beta0_list.append(np.zeros(d_l))
eta_LSGD = 0.01
# train by LSGD_PFL
start_time = time.time()
print("hetero_lv: {} | Repeat: {} | Training by LSGD_PFL".format(
hetero_lv, i_repeat + 1))
loss_train_LSGD, loss_est_LSGD, w_LSGD, beta_list_LSGD = LSGD_PFL(