PFLlib/system/flcore/servers/serverpac.py

# PFLlib: Personalized Federated Learning Algorithm Library
# Copyright (C) 2021  Jianqing Zhang

# This program is free software; you can redistribute it and/or modify
# it under the terms of the GNU General Public License as published by
# the Free Software Foundation; either version 2 of the License, or
# (at your option) any later version.

# This program is distributed in the hope that it will be useful,
# but WITHOUT ANY WARRANTY; without even the implied warranty of
# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
# GNU General Public License for more details.

# You should have received a copy of the GNU General Public License along
# with this program; if not, write to the Free Software Foundation, Inc.,
# 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.

import time
import numpy as np
import random
import torch
import cvxpy as cvx
import copy
from flcore.clients.clientpac import clientPAC
from flcore.servers.serverbase import Server
from threading import Thread
from collections import defaultdict


class FedPAC(Server):
    def __init__(self, args, times):
        super().__init__(args, times)

        # select slow clients
        self.set_slow_clients()
        self.set_clients(clientPAC)

        print(f"\nJoin ratio / total clients: {self.join_ratio} / {self.num_clients}")
        print("Finished creating server and clients.")

        # self.load_model()
        self.Budget = []
        self.num_classes = args.num_classes
        self.global_protos = [None for _ in range(args.num_classes)]

        self.Vars = []
        self.Hs = []
        self.uploaded_heads = []


    def train(self):
        for i in range(self.global_rounds+1):
            s_t = time.time()
            self.selected_clients = self.select_clients()
            self.send_models()

            self.Vars = []
            self.Hs = []
            for client in self.selected_clients:
                self.Vars.append(client.V)
                self.Hs.append(client.h)

            if i%self.eval_gap == 0:
                print(f"\n-------------Round number: {i}-------------")
                print("\nEvaluate personalized models")
                self.evaluate()

            for client in self.selected_clients:
                client.train()

            # threads = [Thread(target=client.train)
            #            for client in self.selected_clients]
            # [t.start() for t in threads]
            # [t.join() for t in threads]

            self.receive_protos()
            self.global_protos = proto_aggregation(self.uploaded_protos)
            self.send_protos()

            self.receive_models()
            self.aggregate_parameters()

            self.aggregate_and_send_heads()

            self.Budget.append(time.time() - s_t)
            print('-'*50, self.Budget[-1])

            if self.auto_break and self.check_done(acc_lss=[self.rs_test_acc], top_cnt=self.top_cnt):
                break

        print("\nBest accuracy.")
        # self.print_(max(self.rs_test_acc), max(
        #     self.rs_train_acc), min(self.rs_train_loss))
        print(max(self.rs_test_acc))
        print(sum(self.Budget[1:])/len(self.Budget[1:]))

        self.save_results()
        

    def send_protos(self):
        assert (len(self.clients) > 0)

        for client in self.clients:
            start_time = time.time()

            client.set_protos(self.global_protos)

            client.send_time_cost['num_rounds'] += 1
            client.send_time_cost['total_cost'] += 2 * (time.time() - start_time)

    def receive_protos(self):
        assert (len(self.selected_clients) > 0)

        self.uploaded_ids = []
        self.uploaded_protos = []
        for client in self.selected_clients:
            self.uploaded_ids.append(client.id)
            self.uploaded_protos.append(client.protos)

    def evaluate(self, acc=None, loss=None):
        stats = self.test_metrics()
        # stats_train = self.train_metrics()

        test_acc = sum(stats[2])*1.0 / sum(stats[1])
        # train_loss = sum(stats_train[2])*1.0 / sum(stats_train[1])
        accs = [a / n for a, n in zip(stats[2], stats[1])]
        
        if acc == None:
            self.rs_test_acc.append(test_acc)
        else:
            acc.append(test_acc)
        
        # if loss == None:
        #     self.rs_train_loss.append(train_loss)
        # else:
        #     loss.append(train_loss)

        # print("Averaged Train Loss: {:.4f}".format(train_loss))
        print("Averaged Test Accurancy: {:.4f}".format(test_acc))
        # self.print_(test_acc, train_acc, train_loss)
        print("Std Test Accurancy: {:.4f}".format(np.std(accs)))

    def receive_models(self):
        assert (len(self.selected_clients) > 0)

        active_clients = random.sample(
            self.selected_clients, int((1-self.client_drop_rate) * self.current_num_join_clients))

        self.uploaded_ids = []
        self.uploaded_weights = []
        self.uploaded_models = []
        self.uploaded_heads = []
        tot_samples = 0
        for client in active_clients:
            try:
                client_time_cost = client.train_time_cost['total_cost'] / client.train_time_cost['num_rounds'] + \
                        client.send_time_cost['total_cost'] / client.send_time_cost['num_rounds']
            except ZeroDivisionError:
                client_time_cost = 0
            if client_time_cost <= self.time_threthold:
                tot_samples += client.train_samples
                self.uploaded_ids.append(client.id)
                self.uploaded_weights.append(client.train_samples)
                self.uploaded_models.append(client.model.base)
                self.uploaded_heads.append(client.model.head)
        for i, w in enumerate(self.uploaded_weights):
            self.uploaded_weights[i] = w / tot_samples

    def aggregate_and_send_heads(self):
        head_weights = solve_quadratic(len(self.uploaded_ids), self.Vars, self.Hs)

        for idx, cid in enumerate(self.uploaded_ids):
            print('(Client {}) Weights of Classifier Head'.format(cid))
            print(head_weights[idx],'\n')

            if head_weights[idx] is not None:
                new_head = self.add_heads(head_weights[idx])
            else:
                new_head = self.uploaded_heads[cid]

            self.clients[cid].set_head(new_head)

    def add_heads(self, weights):
        new_head = copy.deepcopy(self.uploaded_heads[0])
        for param in new_head.parameters():
            param.data.zero_()
                    
        for w, head in zip(weights, self.uploaded_heads):
            for server_param, client_param in zip(new_head.parameters(), head.parameters()):
                server_param.data += client_param.data.clone() * w
        return new_head


# https://github.com/yuetan031/fedproto/blob/main/lib/utils.py#L221
def proto_aggregation(local_protos_list):
    agg_protos_label = defaultdict(list)
    for local_protos in local_protos_list:
        for label in local_protos.keys():
            agg_protos_label[label].append(local_protos[label])

    for [label, proto_list] in agg_protos_label.items():
        if len(proto_list) > 1:
            proto = 0 * proto_list[0].data
            for i in proto_list:
                proto += i.data
            agg_protos_label[label] = proto / len(proto_list)
        else:
            agg_protos_label[label] = proto_list[0].data

    return agg_protos_label


# https://github.com/JianXu95/FedPAC/blob/main/tools.py#L94
def solve_quadratic(num_users, Vars, Hs):
    device = Hs[0][0].device
    num_cls = Hs[0].shape[0] # number of classes
    d = Hs[0].shape[1] # dimension of feature representation
    avg_weight = []
    for i in range(num_users):
        # ---------------------------------------------------------------------------
        # variance ter
        v = torch.tensor(Vars, device=device)
        # ---------------------------------------------------------------------------
        # bias term
        h_ref = Hs[i]
        dist = torch.zeros((num_users, num_users), device=device)
        for j1, j2 in pairwise(tuple(range(num_users))):
            h_j1 = Hs[j1]
            h_j2 = Hs[j2]
            h = torch.zeros((d, d), device=device)
            for k in range(num_cls):
                h += torch.mm((h_ref[k]-h_j1[k]).reshape(d,1), (h_ref[k]-h_j2[k]).reshape(1,d))
            dj12 = torch.trace(h)
            dist[j1][j2] = dj12
            dist[j2][j1] = dj12

        # QP solver
        p_matrix = torch.diag(v) + dist
        p_matrix = p_matrix.cpu().numpy()  # coefficient for QP problem
        evals, evecs = torch.linalg.eig(torch.tensor(p_matrix))
        
        # for numerical stablity
        p_matrix_new = 0
        p_matrix_new = 0
        for ii in range(num_users):
            if evals[ii].real >= 0.01:
                p_matrix_new += evals[ii].real*torch.mm(evecs[:,ii].reshape(num_users,1), evecs[:,ii].reshape(1, num_users))
        p_matrix = p_matrix_new.numpy() if not np.all(np.linalg.eigvals(p_matrix)>=0.0) else p_matrix
        
        # solve QP
        alpha = 0
        eps = 1e-3
        if np.all(np.linalg.eigvals(p_matrix)>=0):
            alphav = cvx.Variable(num_users)
            obj = cvx.Minimize(cvx.quad_form(alphav, p_matrix))
            prob = cvx.Problem(obj, [cvx.sum(alphav) == 1.0, alphav >= 0])
            prob.solve()
            alpha = alphav.value
            alpha = [(i)*(i>eps) for i in alpha] # zero-out small weights (<eps)
        else:
            alpha = None # if no solution for the optimization problem, use local classifier only
        
        avg_weight.append(alpha)

    return avg_weight

# https://github.com/JianXu95/FedPAC/blob/main/tools.py#L10
def pairwise(data):
    n = len(data)
    for i in range(n):
        for j in range(i, n):
            yield (data[i], data[j])