PyTorch实现FedProx算法
Cyril_KI 人气:0I. 前言
FedProx的原理请见:FedAvg联邦学习FedProx异质网络优化实验总结
联邦学习中存在多个客户端,每个客户端都有自己的数据集,这个数据集他们是不愿意共享的。
数据集为某城市十个地区的风电功率,我们假设这10个地区的电力部门不愿意共享自己的数据,但是他们又想得到一个由所有数据统一训练得到的全局模型。
III. FedProx
算法伪代码:
1. 模型定义
客户端的模型为一个简单的四层神经网络模型:
# -*- coding:utf-8 -*- """ @Time: 2022/03/03 12:23 @Author: KI @File: model.py @Motto: Hungry And Humble """ from torch import nn class ANN(nn.Module): def __init__(self, args, name): super(ANN, self).__init__() self.name = name self.len = 0 self.loss = 0 self.fc1 = nn.Linear(args.input_dim, 20) self.relu = nn.ReLU() self.sigmoid = nn.Sigmoid() self.dropout = nn.Dropout() self.fc2 = nn.Linear(20, 20) self.fc3 = nn.Linear(20, 20) self.fc4 = nn.Linear(20, 1) def forward(self, data): x = self.fc1(data) x = self.sigmoid(x) x = self.fc2(x) x = self.sigmoid(x) x = self.fc3(x) x = self.sigmoid(x) x = self.fc4(x) x = self.sigmoid(x) return x
2. 服务器端
服务器端和FedAvg一致,即重复进行客户端采样、参数传达、参数聚合三个步骤:
# -*- coding:utf-8 -*- """ @Time: 2022/03/03 12:50 @Author: KI @File: server.py @Motto: Hungry And Humble """ import copy import random import numpy as np import torch from model import ANN from client import train, test class FedProx: def __init__(self, args): self.args = args self.nn = ANN(args=self.args, name='server').to(args.device) self.nns = [] for i in range(self.args.K): temp = copy.deepcopy(self.nn) temp.name = self.args.clients[i] self.nns.append(temp) def server(self): for t in range(self.args.r): print('round', t + 1, ':') # sampling m = np.max([int(self.args.C * self.args.K), 1]) index = random.sample(range(0, self.args.K), m) # st # dispatch self.dispatch(index) # local updating self.client_update(index, t) # aggregation self.aggregation(index) return self.nn def aggregation(self, index): s = 0 for j in index: # normal s += self.nns[j].len params = {} for k, v in self.nns[0].named_parameters(): params[k] = torch.zeros_like(v.data) for j in index: for k, v in self.nns[j].named_parameters(): params[k] += v.data * (self.nns[j].len / s) for k, v in self.nn.named_parameters(): v.data = params[k].data.clone() def dispatch(self, index): for j in index: for old_params, new_params in zip(self.nns[j].parameters(), self.nn.parameters()): old_params.data = new_params.data.clone() def client_update(self, index, global_round): # update nn for k in index: self.nns[k] = train(self.args, self.nns[k], self.nn, global_round) def global_test(self): model = self.nn model.eval() for client in self.args.clients: model.name = client test(self.args, model)
3. 客户端更新
FedProx中客户端需要优化的函数为:
作者在FedAvg损失函数的基础上,引入了一个proximal term,我们可以称之为近端项。引入近端项后,客户端在本地训练后得到的模型参数 w将不会与初始时的服务器参数wt偏离太多。
对应的代码为:
def train(args, model, server, global_round): model.train() Dtr, Dte = nn_seq_wind(model.name, args.B) model.len = len(Dtr) global_model = copy.deepcopy(server) if args.weight_decay != 0: lr = args.lr * pow(args.weight_decay, global_round) else: lr = args.lr if args.optimizer == 'adam': optimizer = torch.optim.Adam(model.parameters(), lr=lr, weight_decay=args.weight_decay) else: optimizer = torch.optim.SGD(model.parameters(), lr=lr, momentum=0.9, weight_decay=args.weight_decay) print('training...') loss_function = nn.MSELoss().to(args.device) loss = 0 for epoch in range(args.E): for (seq, label) in Dtr: seq = seq.to(args.device) label = label.to(args.device) y_pred = model(seq) optimizer.zero_grad() # compute proximal_term proximal_term = 0.0 for w, w_t in zip(model.parameters(), global_model.parameters()): proximal_term += (w - w_t).norm(2) loss = loss_function(y_pred, label) + (args.mu / 2) * proximal_term loss.backward() optimizer.step() print('epoch', epoch, ':', loss.item()) return model
我们在原有MSE损失函数的基础上加上了一个近端项:
for w, w_t in zip(model.parameters(), global_model.parameters()): proximal_term += (w - w_t).norm(2)
然后再反向传播求梯度,然后优化器step更新参数。
原始论文中还提出了一个不精确解的概念:
不过值得注意的是,我并没有在原始论文的实验部分找到如何选择 γ \gamma γ的说明。查了一下资料后发现是涉及到了近端梯度下降的知识,本文代码并没有考虑不精确解,后期可能会补上。
IV. 完整代码
链接:http://pan.baidu.com/s/1hj2EOcqIUmM-C6R1cyjE5Q
提取码:fghp
项目结构:
其中:
- server.py为服务器端操作。
- client.py为客户端操作。
- data_process.py为数据处理部分。
- model.py为模型定义文件。
- args.py为参数定义文件。
- main.py为主文件,如想要运行此项目可直接运行:
python main.py
加载全部内容