idxs=dict_users_test[user])
if best_acc is None or acc_test > best_acc:
best_acc = acc_test
net_best = copy.deepcopy(net_local)
# torch.save(net_local_list[user].state_dict(), model_save_path)
print('User {}, Epoch {}, Acc {:.2f}'.format(user, iter, acc_test))
if iter > 50 and acc_test >= 99:
break
net_local_list[user] = net_best
acc_test_local, loss_test_local = test_img_local_all(
net_local_list, args, dataset_test, dict_users_test)
acc_test_avg, loss_test_avg = test_img_avg_all(net_glob, net_local_list,
args, dataset_test)
acc_test_ens_avg, loss_test, acc_test_ens_maj = test_img_ensemble_all(
net_local_list, args, dataset_test)
print(
'Final: acc: {:.2f}, acc (avg): {:.2f}, acc (ens,avg): {:.2f}, acc (ens,maj): {:.2f}'
.format(acc_test_local, acc_test_avg, acc_test_ens_avg,
acc_test_ens_maj))
final_results = np.array(
[[acc_test_local, acc_test_avg, acc_test_ens_avg, acc_test_ens_maj]])
final_results = pd.DataFrame(final_results,
columns=[
'acc_test_local', 'acc_test_avg',
'acc_test_ens_avg', 'acc_test_ens_maj'
])
num_param_glob = 0
num_param_local = 0
for key in net_glob.state_dict().keys():
num_param_local += net_glob.state_dict()[key].numel()
if key in w_glob_keys:
num_param_glob += net_glob.state_dict()[key].numel()
percentage_param = 100 * float(num_param_glob) / num_param_local
print('# Params: {} (local), {} (global); Percentage {:.2f} ({}/{})'.format(
num_param_local, num_param_glob, percentage_param, num_param_glob, num_param_local))
# generate list of local models for each user
net_local_list = []
for user in range(args.num_users):
net_local_list.append(copy.deepcopy(net_glob))
acc_test_avg, loss_test_avg = test_img_avg_all(net_glob, net_local_list, args, dataset_test)
acc_test_local_list, _ = test_img_local_all(net_local_list, args, dataset_test, dict_users_test, return_all=True)
acc_test_local = acc_test_local_list.mean()
# training
results_save_path = os.path.join(base_save_dir, 'results.csv')
results_columns = ['epoch', 'acc_test_local', 'acc_test_avg', 'best_acc_local', 'acc_test_ens_avg', 'acc_test_ens_maj']
loss_train = []
best_iter = -1
best_acc_local = -1
best_acc_list = acc_test_local_list
best_net_list = copy.deepcopy(net_local_list)
fina_net_list = copy.deepcopy(net_local_list)
results = []