唐山建网站公司,家电网站建设总体目标,企业策划书模板范文大学生,无锡网站建设工作室目录
1. 联邦学习介绍
2. 实验流程
3. 数据加载
4. 模型构建
5. 数据采样函数
6. 模型训练 1. 联邦学习介绍
联邦学习是一种分布式机器学习方法#xff0c;中心节点为server#xff08;服务器#xff09;#xff0c;各分支节点为本地的client#xff08;设备#…目录
1. 联邦学习介绍
2. 实验流程
3. 数据加载
4. 模型构建
5. 数据采样函数
6. 模型训练 1. 联邦学习介绍
联邦学习是一种分布式机器学习方法中心节点为server服务器各分支节点为本地的client设备。联邦学习的模式是在各分支节点分别利用本地数据训练模型再将训练好的模型汇合到中心节点获得一个更好的全局模型。
联邦学习的提出是为了充分利用用户的数据特征训练效果更佳的模型同时为了保证隐私联邦学习在训练过程中server和clients之间通信的是模型的参数或梯度、参数更新量本地的数据不会上传到服务器。
本项目主要是升级1.8版本的联邦学习fedavg算法至2.3版本内容取材于基于PaddlePaddle实现联邦学习算法FedAvg - 飞桨AI Studio星河社区
2. 实验流程
联邦学习的基本流程是
1. server初始化模型参数所有的clients将这个初始模型下载到本地
2. clients利用本地产生的数据进行SGD训练
3. 选取K个clients将训练得到的模型参数上传到server
4. server对得到的模型参数整合所有的clients下载新的模型。
5. 重复执行2-5直至收敛或达到预期要求
import os
import numpy as np
import matplotlib
import matplotlib.pyplot as plt
import random
import time
import paddle
import paddle.nn as nn
import numpy as np
from paddle.io import Dataset,DataLoader
import paddle.nn.functional as F3. 数据加载
mnist_data_trainnp.load(data/data2489/train_mnist.npy)
mnist_data_testnp.load(data/data2489/test_mnist.npy)
print(There are {} images for training.format(len(mnist_data_train)))
print(There are {} images for testing.format(len(mnist_data_test)))
# 数据和标签分离便于后续处理
Label[int(i[0]) for i in mnist_data_train]
Data[i[1:] for i in mnist_data_train]There are 60000 images for training
There are 10000 images for testing4. 模型构建
class CNN(nn.Layer):def __init__(self):super(CNN,self).__init__()self.conv1nn.Conv2D(1,32,5)self.relu nn.ReLU()self.pool1nn.MaxPool2D(kernel_size2,stride2)self.conv2nn.Conv2D(32,64,5)self.pool2nn.MaxPool2D(kernel_size2,stride2)self.fc1nn.Linear(1024,512)self.fc2nn.Linear(512,10)# self.softmax nn.Softmax()def forward(self,inputs):x self.conv1(inputs)x self.relu(x)x self.pool1(x)x self.conv2(x)x self.relu(x)x self.pool2(x)xpaddle.reshape(x,[-1,1024])x self.relu(self.fc1(x))y self.fc2(x)return y5. 数据采样函数
# 均匀采样分配到各个client的数据集都是IID且数量相等的
def IID(dataset, clients):num_items_per_client int(len(dataset)/clients)client_dict {}image_idxs [i for i in range(len(dataset))]for i in range(clients):client_dict[i] set(np.random.choice(image_idxs, num_items_per_client, replaceFalse)) # 为每个client随机选取数据image_idxs list(set(image_idxs) - client_dict[i]) # 将已经选取过的数据去除client_dict[i] list(client_dict[i])return client_dict# 非均匀采样同时各个client上的数据分布和数量都不同
def NonIID(dataset, clients, total_shards, shards_size, num_shards_per_client):shard_idxs [i for i in range(total_shards)]client_dict {i: np.array([], dtypeint64) for i in range(clients)}idxs np.arange(len(dataset))data_labels Labellabel_idxs np.vstack((idxs, data_labels)) # 将标签和数据ID堆叠label_idxs label_idxs[:, label_idxs[1,:].argsort()]idxs label_idxs[0,:]for i in range(clients):rand_set set(np.random.choice(shard_idxs, num_shards_per_client, replaceFalse)) shard_idxs list(set(shard_idxs) - rand_set)for rand in rand_set:client_dict[i] np.concatenate((client_dict[i], idxs[rand*shards_size:(rand1)*shards_size]), axis0) # 拼接return client_dictclass MNISTDataset(Dataset):def __init__(self, data,label):self.data dataself.label labeldef __getitem__(self, idx):imagenp.array(self.data[idx]).astype(float32)imagenp.reshape(image,[1,28,28])labelnp.array(self.label[idx]).astype(int64)return image, labeldef __len__(self):return len(self.label)
6. 模型训练
class ClientUpdate(object):def __init__(self, data, label, batch_size, learning_rate, epochs):dataset MNISTDataset(data,label)self.train_loader DataLoader(dataset,batch_sizebatch_size,shuffleTrue,drop_lastTrue)self.learning_rate learning_rateself.epochs epochsdef train(self, model):optimizerpaddle.optimizer.SGD(learning_rateself.learning_rate,parametersmodel.parameters())criterion nn.CrossEntropyLoss(reductionmean)model.train()e_loss []for epoch in range(1,self.epochs1):train_loss []for image,label in self.train_loader:# imagepaddle.to_tensor(image)# labelpaddle.to_tensor(label.reshape([label.shape[0],1]))outputmodel(image)loss criterion(output,label)# print(loss)loss.backward()optimizer.step()optimizer.clear_grad()train_loss.append(loss.numpy()[0])t_losssum(train_loss)/len(train_loss)e_loss.append(t_loss)total_losssum(e_loss)/len(e_loss)return model.state_dict(), total_losstrain_x np.array(Data)
train_y np.array(Label)BATCH_SIZE 32
# 通信轮数
rounds 100
# client比例
C 0.1
# clients数量
K 100
# 每次通信在本地训练的epoch
E 5
# batch size
batch_size 10
# 学习率
lr0.001
# 数据切分
iid_dict IID(mnist_data_train, 100)def training(model, rounds, batch_size, lr, ds,L, data_dict, C, K, E, plt_title, plt_color):global_weights model.state_dict()train_loss []start time.time()# clients与server之间通信for curr_round in range(1, rounds1):w, local_loss [], []m max(int(C*K), 1) # 随机选取参与更新的clientsS_t np.random.choice(range(K), m, replaceFalse)for k in S_t:# print(data_dict[k])sub_data ds[data_dict[k]]sub_y L[data_dict[k]]local_update ClientUpdate(sub_data,sub_y, batch_sizebatch_size, learning_ratelr, epochsE)weights, loss local_update.train(model)w.append(weights)local_loss.append(loss)# 更新global weightsweights_avg w[0]for k in weights_avg.keys():for i in range(1, len(w)):# weights_avg[k] (num[i]/sum(num))*w[i][k]weights_avg[k]weights_avg[k]w[i][k] weights_avg[k]weights_avg[k]/len(w)global_weights[k].set_value(weights_avg[k])# global_weights weights_avg# print(global_weights)#模型加载最新的参数model.load_dict(global_weights)loss_avg sum(local_loss) / len(local_loss)if curr_round % 10 0:print(Round: {}... \tAverage Loss: {}.format(curr_round, np.round(loss_avg, 5)))train_loss.append(loss_avg)end time.time()fig, ax plt.subplots()x_axis np.arange(1, rounds1)y_axis np.array(train_loss)ax.plot(x_axis, y_axis, tab:plt_color)ax.set(xlabelNumber of Rounds, ylabelTrain Loss,titleplt_title)ax.grid()fig.savefig(plt_title.jpg, formatjpg)print(Training Done!)print(Total time taken to Train: {}.format(end-start))return model.state_dict()#导入模型
mnist_cnn CNN()
mnist_cnn_iid_trained training(mnist_cnn, rounds, batch_size, lr, train_x,train_y, iid_dict, C, K, E, MNIST CNN on IID Dataset, orange)W0605 23:22:00.961916 10307 gpu_context.cc:278] Please NOTE: device: 0, GPU Compute Capability: 7.0, Driver API Version: 11.2, Runtime API Version: 10.1
W0605 23:22:00.966121 10307 gpu_context.cc:306] device: 0, cuDNN Version: 7.6.Round: 10... Average Loss: 0.033
Round: 20... Average Loss: 0.011
Round: 30... Average Loss: 0.012
Round: 40... Average Loss: 0.008
Round: 50... Average Loss: 0.003
Round: 60... Average Loss: 0.002
Round: 70... Average Loss: 0.001
Round: 80... Average Loss: 0.001
Round: 90... Average Loss: 0.001
