def run_master(config):
w_i, num_worker = config.site_id, config.num_worker
beta, S, tau, steps = config.beta, config.S, config.tau, config.steps
device = config.device
X = load_data(w_i, num_worker)
x_dim = tuple(X.shape[1:])
master = AvgMaster(X, num_worker, x_dim, beta, S, tau, device)
addrport = ('localhost', config.port)
server = Server(addrport, master.send_queue, master.recv_queue, num_worker)
server.accept()
recv_thread = threading.Thread(target=server.recv_loop, daemon=True)
recv_thread.start()
# For logging
import pandas as pd
log_dir = f'./logs/mnist_avg/S={S} tau={tau}'
os.makedirs(log_dir, exist_ok=True)
time_vals = []
z_vals = torch.zeros((steps, ) + x_dim, device=device)
print('Started computing!')
time_0 = time.time()
for _ in server.send_iter():
if master.stop:
print('Optimization Ended, calculating objective values.')
data = []
for step in range(steps):
obj = master.objective(z_vals[step])
data.append([step + 1, time_vals[step], obj.item()])
dataframe = pd.DataFrame(
data, columns=['global_step', 'wall_time', 'objective'])
dataframe.to_pickle(log_dir + '/logs.pkl')
dataframe.to_csv(log_dir + '/logs.csv')
np.save(log_dir + '/results.npy', master.z.data.cpu().numpy())
print('Done!')
break
master.receive()
if master.update():
if master.k == 1: time_0 = time.time()
z_vals[master.k - 1] = master.z
time_vals.append(time.time() - time_0)
if master.k == steps:
master.stop_algorithm()
def run_master(config):
w_i, num_worker = config.site_id, config.num_worker
beta, S, tau, steps = config.beta, config.S, config.tau, config.steps
device = config.device
X, Y, X_test, Y_test, xm, xd = load_data(w_i, num_worker)
x_dim = (MNIST_SHAPE + 1, 10)
master = MCMaster(X, Y, X_test, Y_test, num_worker, x_dim, beta, S, tau,
device)
addrport = ('localhost', config.port)
server = Server(addrport, master.send_queue, master.recv_queue, num_worker)
server.accept()
recv_thread = threading.Thread(target=server.recv_loop, daemon=True)
recv_thread.start()
# For logging
import pandas as pd
log_dir = f'./logs/mnist_logistic/S={S} tau={tau}'
os.makedirs(log_dir, exist_ok=True)
time_vals = []
z_vals = torch.zeros((steps, ) + x_dim, device=device)
print('Started computing!')
time_0 = time.time()
for _ in server.send_iter():
if master.stop:
print('Optimization Ended, calculating objective values.')
data = []
for step in range(steps):
train_loss, test_loss = master.objective(z_vals[step])
data.append([
step + 1, time_vals[step],
train_loss.item(),
test_loss.item()
])
dataframe = pd.DataFrame(data,
columns=[
'global_step', 'wall_time',
'train_loss', 'test_loss'
])
dataframe.to_pickle(log_dir + '/logs.pkl')
dataframe.to_csv(log_dir + '/logs.csv')
w, b = W2wb(master.z)
save_results(log_dir + '/results.npz', w=w, b=b, xm=xm, xd=xd)
print('Done!')
break
master.receive()
if master.update():
if master.k == 1: time_0 = time.time()
z_vals[master.k - 1] = master.z
time_vals.append(time.time() - time_0)
if master.k == steps:
master.stop_algorithm()
