联邦学习-Tensorflow实现联邦模型AlexNet on CIFAR-10

目录

Client端

Server端

扩展

Client.py

Server.py

Dataset.py

Model.py


联邦学习-Tensorflow实现联邦模型AlexNet on CIFAR-10_第1张图片

分享一种实现联邦学习的方法,它具有以下优点:

不需要读写文件来保存、切换Client模型
不需要在每次epoch重新初始化Client变量
内存占用尽可能小(参数量仅翻一倍,即Client端+Server端)
切换Client只增加了一些赋值操作

学习的目标是一个更好的模型,由Server保管,Clients提供更新
数据(Data)由Clients保管、使用文章的代码环境、库依赖:

Python 3.7
Tensorflow v1.14.x
tqdm(一个Python模块)


接下来本文会分成Client端、Server端代码设计与实现进行讲解。懒得看讲解的可以直接拉到最后的完整代码章节,共有四个代码文件,运行python Server.py即可以立马体验原汁原味的(单机模拟)联邦学习。

Client端


明确一下Client端的任务,包含下面三个步骤:

将Server端发来的模型变量加载到模型上
用自己的所有数据更新当前模型
将更新后的模型变量发回给Server
在这些任务下,我们可以设计出Client代码需要具备的一些功能:

创建、训练Tensorflow模型(也就是计算图)
加载Server端发过来的模型变量值
提取当前模型的变量值,发送给Server
维护自己的数据集用于训练
其实,仔细一想也就比平时写的tf模型代码多了个加载、提取模型变量。假设Client类已经构建好了模型,那么sess.run()一下每个变量,即可得到模型变量的值了。下面的代码展示了部分Clients类的定义,get_client_vars函数将返回计算图中所有可训练的变量值:

class Clients:
    def __init__(self, input_shape, num_classes, learning_rate, clients_num):
        self.graph = tf.Graph()
        self.sess = tf.Session(graph=self.graph)
        
        """ 本函数未完待续... """
        
        
    def get_client_vars(self):
        """ Return all of the variables list """
        with self.graph.as_default():
            client_vars = self.sess.run(tf.trainable_variables())
        return client_vars


加载Server端发过来的global_vars到模型变量上,核心在于tf.Variable.load()函数,把一个Tensor的值加载到模型变量中,例如:

variable.load(tensor, sess)


将tensor(类型为tf.Tensor)的值赋值给variable(类型为tf.Varibale),sess是tf.Session。

如果要把整个模型中的变量值都加载,可以用tf.trainable_variables()获取计算图中的所有可训练变量(一个list),保证它和global_vars的顺序对应后,可以这样实现:

    def set_global_vars(self, global_vars):
        """ Assign all of the variables with global vars """
        with self.graph.as_default():
            all_vars = tf.trainable_variables()
            for variable, value in zip(all_vars, global_vars):
                variable.load(value, self.sess)


此外,Clients类还需要进行模型定义和训练。我相信这不是实现联邦的重点,因此在下面的代码中,我将函数体去掉只留下接口定义(完整代码在最后一个章节):

import tensorflow as tf
import numpy as np
from collections import namedtuple
import math

# 自定义的模型定义函数
from Model import AlexNet
# 自定义的数据集类
from Dataset import Dataset

# The definition of fed model
# 用namedtuple来储存一个模型,依次为:
# X: 输入
# Y: 输出
# DROP_RATE: 顾名思义
# train_op: tf计算图中的训练节点(一般是optimizer.minimize(xxx))
# loss_op: 顾名思义
# loss_op: 顾名思义
FedModel = namedtuple('FedModel', 'X Y DROP_RATE train_op loss_op acc_op')


class Clients:
    def __init__(self, input_shape, num_classes, learning_rate, clients_num):
        self.graph = tf.Graph()
        self.sess = tf.Session(graph=self.graph)

        # Call the create function to build the computational graph of AlexNet
        # `net` 是一个list,依次包含模型中FedModel需要的计算节点(看上面)
        net = AlexNet(input_shape, num_classes, learning_rate, self.graph)
        self.model = FedModel(*net)

        # initialize 初始化
        with self.graph.as_default():
            self.sess.run(tf.global_variables_initializer())

        # Load Cifar-10 dataset
        # NOTE: len(self.dataset.train) == clients_num
        # 加载数据集。对于训练集:`self.dataset.train[56]`可以获取56号client的数据集
        # `self.dataset.train[56].next_batch(32)`可以获取56号client的一个batch,大小为32
        # 对于测试集,所有client共用一个测试集,因此:
        # `self.dataset.test.next_batch(1000)`将获取大小为1000的数据集(无随机)
        self.dataset = Dataset(tf.keras.datasets.cifar10.load_data,
                        split=clients_num)

    def run_test(self, num):
        """
            Predict the testing set, and report the acc and loss
            预测测试集,返回准确率和loss
        
            num: number of testing instances
        """
        pass

    def train_epoch(self, cid, batch_size=32, dropout_rate=0.5):
        """
            Train one client with its own data for one epoch
            用`cid`号的client的数据对模型进行训练
            cid: Client id
        """
        pass
        
    def choose_clients(self, ratio=1.0):
        """
            randomly choose some clients
            随机选择`ratio`比例的clients,返回编号(也就是下标)
        """
        client_num = self.get_clients_num()
        choose_num = math.floor(client_num * ratio)
        return np.random.permutation(client_num)[:choose_num]
    
    def get_clients_num(self):
        """ 返回clients的数量 """
        return len(self.dataset.train)


细心的同学可能已经发现了,类名是Clients是复数,表示一堆Clients的集合。但模型self.model只有一个,原因是:不同Clients的模型实际上是一样的,只是数据不同;类成员self.dataset已经对数据进行了划分,需要不同client参与训练时,只需要用Server给的变量值把模型变量覆盖掉,再用下标cid找到该Client的数据进行训练就得了。

当然,这样实现的最重要原因,是避免构建那么多个Client的计算图。咱没那么多显存TAT
概括一下:联邦学习的Clients,只是普通TF训练模型代码上,加上模型变量的值提取、赋值功能。

Server端


按照套路,明确一下Server端代码的主要任务:

使用Clients:给一组模型变量给某个Client进行更新,把更新后的变量值拿回来
管理全局模型:每一轮更新,收集多个Clients更新后的模型进行归总,成为新一轮的模型
简单起见,我们Server端的代码不再抽象成一个类,而是以脚本的形式编写。首先,实例化咱们上面定义的Clients:

from Client import Clients

def buildClients(num):
    learning_rate = 0.0001
    num_input = 32  # image shape: 32*32
    num_input_channel = 3  # image channel: 3
    num_classes = 10  # Cifar-10 total classes (0-9 digits)

    #create Client and model
    return Clients(input_shape=[None, num_input, num_input, num_input_channel],
                  num_classes=num_classes,
                  learning_rate=learning_rate,
                  clients_num=num)

CLIENT_NUMBER = 100
client = buildClients(CLIENT_NUMBER)
global_vars = client.get_client_vars()


client变量储存着CLIENT_NUMBER个Clients的模型(实际上只有一个计算图)和数据。global_vars储存着Server端的模型变量值,也就是我们大名鼎鼎的训练目标,目前它只是Client端模型初始化的值。

接下来,对于Server的一个epoch,Server会随机挑选一定比例的Clients参与这轮训练,分别把当前的Server端模型global_vars交给它们进行更新,并分别收集它们更新后的变量。本轮参与训练的Clients都收集后,平均一下这些更新后的变量值,就得到新一轮的Server端模型,然后进行下一个epoch。下面是循环epoch更新的代码,仔细看注释哦:

def run_global_test(client, global_vars, test_num):
    """ 跑一下测试集,输出ACC和Loss """
    client.set_global_vars(global_vars)
    acc, loss = client.run_test(test_num)
    print("[epoch {}, {} inst] Testing ACC: {:.4f}, Loss: {:.4f}".format(
        ep + 1, test_num, acc, loss))


CLIENT_RATIO_PER_ROUND = 0.12  # 每轮挑选clients跑跑看的比例
epoch = 360  # epoch上限

for ep in range(epoch):
    # We are going to sum up active clients' vars at each epoch
    # 用来收集Clients端的参数,全部叠加起来(节约内存)
    client_vars_sum = None

    # Choose some clients that will train on this epoch
    # 随机挑选一些Clients进行训练
    random_clients = client.choose_clients(CLIENT_RATIO_PER_ROUND)

    # Train with these clients
    # 用这些Clients进行训练,收集它们更新后的模型
    for client_id in tqdm(random_clients, ascii=True):
        # Restore global vars to client's model
        # 将Server端的模型加载到Client模型上
        client.set_global_vars(global_vars)

        # train one client
        # 训练这个下标的Client
        client.train_epoch(cid=client_id)

        # obtain current client's vars
        # 获取当前Client的模型变量值
        current_client_vars = client.get_client_vars()

        # sum it up
        # 把各个层的参数叠加起来
        if client_vars_sum is None:
            client_vars_sum = current_client_vars
        else:
            for cv, ccv in zip(client_vars_sum, current_client_vars):
                cv += ccv

    # obtain the avg vars as global vars
    # 把叠加后的Client端模型变量 除以 本轮参与训练的Clients数量
    # 得到平均模型、作为新一轮的Server端模型参数
    global_vars = []
    for var in client_vars_sum:
        global_vars.append(var / len(random_clients))

    # run test on 1000 instances
    # 跑一下测试集、输出一下
    run_global_test(client, global_vars, test_num=600)


经过那么一些轮的迭代,我们就可以得到Server端的训练好的模型参数global_vars了。虽然它逻辑很简单,但我希望观众老爷们能注意到其中的两个联邦点:Server端代码没有接触到数据;每次参与训练的Clients数量相对于整体来说是很少的。

扩展


如果要更换模型,只需要实现新的模型计算图构造函数,替换Client端的AlexNet函数,保证它能返回那一系列的计算节点即可。

如果要实现Non-I.I.D.的数据分布,只需要修改Dataset.py中的数据划分方式。但是,我稍微试验了一下,目前这个模型+训练方式,不能应对极度Non-I.I.D.的情况。也反面证明了,Non-I.I.D.确实是联邦学习的一个难题。

如果要Clients和Server之间传模型梯度,需要把Client端的计算梯度和更新变量分开,中间插入和Server端的交互,交互内容就是梯度。这样说有点抽象,很多同学可能经常用Optimizer.minimize(文档在这),但并不知道它是另外两个函数的组合,分别为:compute_gradients()和apply_gradients()。前者是计算梯度,后者是把梯度按照学习率更新到变量上。把梯度拿到后,交给Server,Server返回一个全局平均后的梯度再更新模型。尝试过是可行的,但是并不能减少传输量,而且单机模拟实现难度大了许多。

如果要分布式部署,那就把Clients端代码放在flask等web后端服务下进行部署,Server端通过网络传输与Clients进行通信。需要注意,Server端发起请求的时候,可能因为参数量太大导致一些问题,考虑换个非HTTP协议。

完整代码
一共有四个代码文件,他们应当放在同一个文件目录下:

Client.py:Client端代码,管理模型、数据
Server.py:Server端代码,管理Clients、全局模型
Dataset.py:定义数据的组织形式
Model.py:定义TF模型的计算图
我也将它们传到了Github上,仓库链接:https://github.com/Zing22/tf-fed-demo。

下面开始分别贴出它们的完整代码,其中的注释只有我边打码边写的一点点,上文的介绍中补充了更多中文注释。运行方法非常简单:

python Server.py


Client.py


import tensorflow as tf
import numpy as np
from collections import namedtuple
import math

from Model import AlexNet
from Dataset import Dataset

# The definition of fed model
FedModel = namedtuple('FedModel', 'X Y DROP_RATE train_op loss_op acc_op')

class Clients:
    def __init__(self, input_shape, num_classes, learning_rate, clients_num):
        self.graph = tf.Graph()
        self.sess = tf.Session(graph=self.graph)

        # Call the create function to build the computational graph of AlexNet
        net = AlexNet(input_shape, num_classes, learning_rate, self.graph)
        self.model = FedModel(*net)

        # initialize
        with self.graph.as_default():
            self.sess.run(tf.global_variables_initializer())

        # Load Cifar-10 dataset
        # NOTE: len(self.dataset.train) == clients_num
        self.dataset = Dataset(tf.keras.datasets.cifar10.load_data,
                        split=clients_num)

    def run_test(self, num):
        with self.graph.as_default():
            batch_x, batch_y = self.dataset.test.next_batch(num)
            feed_dict = {
                self.model.X: batch_x,
                self.model.Y: batch_y,
                self.model.DROP_RATE: 0
            }
        return self.sess.run([self.model.acc_op, self.model.loss_op],
                             feed_dict=feed_dict)

    def train_epoch(self, cid, batch_size=32, dropout_rate=0.5):
        """
            Train one client with its own data for one epoch
            cid: Client id
        """
        dataset = self.dataset.train[cid]

        with self.graph.as_default():
            for _ in range(math.ceil(dataset.size / batch_size)):
                batch_x, batch_y = dataset.next_batch(batch_size)
                feed_dict = {
                    self.model.X: batch_x,
                    self.model.Y: batch_y,
                    self.model.DROP_RATE: dropout_rate
                }
                self.sess.run(self.model.train_op, feed_dict=feed_dict)

    def get_client_vars(self):
        """ Return all of the variables list """
        with self.graph.as_default():
            client_vars = self.sess.run(tf.trainable_variables())
        return client_vars

    def set_global_vars(self, global_vars):
        """ Assign all of the variables with global vars """
        with self.graph.as_default():
            all_vars = tf.trainable_variables()
            for variable, value in zip(all_vars, global_vars):
                variable.load(value, self.sess)

    def choose_clients(self, ratio=1.0):
        """ randomly choose some clients """
        client_num = self.get_clients_num()
        choose_num = math.ceil(client_num * ratio)
        return np.random.permutation(client_num)[:choose_num]

    def get_clients_num(self):
        return len(self.dataset.train)


Server.py


import tensorflow as tf
from tqdm import tqdm

from Client import Clients

def buildClients(num):
    learning_rate = 0.0001
    num_input = 32  # image shape: 32*32
    num_input_channel = 3  # image channel: 3
    num_classes = 10  # Cifar-10 total classes (0-9 digits)

    #create Client and model
    return Clients(input_shape=[None, num_input, num_input, num_input_channel],
                  num_classes=num_classes,
                  learning_rate=learning_rate,
                  clients_num=num)


def run_global_test(client, global_vars, test_num):
    client.set_global_vars(global_vars)
    acc, loss = client.run_test(test_num)
    print("[epoch {}, {} inst] Testing ACC: {:.4f}, Loss: {:.4f}".format(
        ep + 1, test_num, acc, loss))


#### SOME TRAINING PARAMS ####
CLIENT_NUMBER = 100
CLIENT_RATIO_PER_ROUND = 0.12
epoch = 360


#### CREATE CLIENT AND LOAD DATASET ####
client = buildClients(CLIENT_NUMBER)

#### BEGIN TRAINING ####
global_vars = client.get_client_vars()
for ep in range(epoch):
    # We are going to sum up active clients' vars at each epoch
    client_vars_sum = None

    # Choose some clients that will train on this epoch
    random_clients = client.choose_clients(CLIENT_RATIO_PER_ROUND)

    # Train with these clients
    for client_id in tqdm(random_clients, ascii=True):
        # Restore global vars to client's model
        client.set_global_vars(global_vars)

        # train one client
        client.train_epoch(cid=client_id)

        # obtain current client's vars
        current_client_vars = client.get_client_vars()

        # sum it up
        if client_vars_sum is None:
            client_vars_sum = current_client_vars
        else:
            for cv, ccv in zip(client_vars_sum, current_client_vars):
                cv += ccv

    # obtain the avg vars as global vars
    global_vars = []
    for var in client_vars_sum:
        global_vars.append(var / len(random_clients))

    # run test on 600 instances
    run_global_test(client, global_vars, test_num=600)


#### FINAL TEST ####
run_global_test(client, global_vars, test_num=10000)


Dataset.py


import numpy as np
from tensorflow.keras.utils import to_categorical


class BatchGenerator:
    def __init__(self, x, yy):
        self.x = x
        self.y = yy
        self.size = len(x)
        self.random_order = list(range(len(x)))
        np.random.shuffle(self.random_order)
        self.start = 0
        return

    def next_batch(self, batch_size):
        perm = self.random_order[self.start:self.start + batch_size]

        self.start += batch_size
        if self.start > self.size:
            self.start = 0

        return self.x[perm], self.y[perm]

    # support slice
    def __getitem__(self, val):
        return self.x[val], self.y[val]


class Dataset(object):
    def __init__(self, load_data_func, one_hot=True, split=0):
        (x_train, y_train), (x_test, y_test) = load_data_func()
        print("Dataset: train-%d, test-%d" % (len(x_train), len(x_test)))

        if one_hot:
            y_train = to_categorical(y_train, 10)
            y_test = to_categorical(y_test, 10)

        x_train = x_train.astype('float32') / 255
        x_test = x_test.astype('float32') / 255

        if split == 0:
            self.train = BatchGenerator(x_train, y_train)
        else:
            self.train = self.splited_batch(x_train, y_train, split)

        self.test = BatchGenerator(x_test, y_test)

    def splited_batch(self, x_data, y_data, split):
        res = []
        for x, y in zip(np.split(x_data, split), np.split(y_data, split)):
            assert len(x) == len(y)
            res.append(BatchGenerator(x, y))
        return res


Model.py


import tensorflow as tf
import numpy as np
from tensorflow.compat.v1.train import AdamOptimizer

#### Create tf model for Client ####

def AlexNet(input_shape, num_classes, learning_rate, graph):
    """
        Construct the AlexNet model.
        input_shape: The shape of input (`list` like)
        num_classes: The number of output classes (`int`)
        learning_rate: learning rate for optimizer (`float`)
        graph: The tf computation graph (`tf.Graph`)
    """
    with graph.as_default():
        X = tf.placeholder(tf.float32, input_shape, name='X')
        Y = tf.placeholder(tf.float32, [None, num_classes], name='Y')
        DROP_RATE = tf.placeholder(tf.float32, name='drop_rate')

        # 1st Layer: Conv (w ReLu) -> Lrn -> Pool
        # conv1 = conv(X, 11, 11, 96, 4, 4, padding='VALID', name='conv1')
        conv1 = conv(X, 11, 11, 96, 2, 2, name='conv1')
        norm1 = lrn(conv1, 2, 2e-05, 0.75, name='norm1')
        pool1 = max_pool(norm1, 3, 3, 2, 2, padding='VALID', name='pool1')

        # 2nd Layer: Conv (w ReLu)  -> Lrn -> Pool with 2 groups
        conv2 = conv(pool1, 5, 5, 256, 1, 1, groups=2, name='conv2')
        norm2 = lrn(conv2, 2, 2e-05, 0.75, name='norm2')
        pool2 = max_pool(norm2, 3, 3, 2, 2, padding='VALID', name='pool2')

        # 3rd Layer: Conv (w ReLu)
        conv3 = conv(pool2, 3, 3, 384, 1, 1, name='conv3')

        # 4th Layer: Conv (w ReLu) splitted into two groups
        conv4 = conv(conv3, 3, 3, 384, 1, 1, groups=2, name='conv4')

        # 5th Layer: Conv (w ReLu) -> Pool splitted into two groups
        conv5 = conv(conv4, 3, 3, 256, 1, 1, groups=2, name='conv5')
        pool5 = max_pool(conv5, 3, 3, 2, 2, padding='VALID', name='pool5')

        # 6th Layer: Flatten -> FC (w ReLu) -> Dropout
        # flattened = tf.reshape(pool5, [-1, 6*6*256])
        # fc6 = fc(flattened, 6*6*256, 4096, name='fc6')

        flattened = tf.reshape(pool5, [-1, 1 * 1 * 256])
        fc6 = fc_layer(flattened, 1 * 1 * 256, 1024, name='fc6')
        dropout6 = dropout(fc6, DROP_RATE)

        # 7th Layer: FC (w ReLu) -> Dropout
        # fc7 = fc(dropout6, 4096, 4096, name='fc7')
        fc7 = fc_layer(dropout6, 1024, 2048, name='fc7')
        dropout7 = dropout(fc7, DROP_RATE)

        # 8th Layer: FC and return unscaled activations
        logits = fc_layer(dropout7, 2048, num_classes, relu=False, name='fc8')

        # loss and optimizer
        loss_op = tf.reduce_mean(
            tf.nn.softmax_cross_entropy_with_logits_v2(logits=logits,
                                                        labels=Y))
        optimizer = AdamOptimizer(
            learning_rate=learning_rate)
        train_op = optimizer.minimize(loss_op)

        # Evaluate model
        prediction = tf.nn.softmax(logits)
        pred = tf.argmax(prediction, 1)

        # accuracy
        correct_pred = tf.equal(pred, tf.argmax(Y, 1))
        accuracy = tf.reduce_mean(
            tf.cast(correct_pred, tf.float32))

        return X, Y, DROP_RATE, train_op, loss_op, accuracy


def conv(x, filter_height, filter_width, num_filters,
            stride_y, stride_x, name, padding='SAME', groups=1):
    """Create a convolution layer.

    Adapted from: https://github.com/ethereon/caffe-tensorflow
    """
    # Get number of input channels
    input_channels = int(x.get_shape()[-1])

    # Create lambda function for the convolution
    convolve = lambda i, k: tf.nn.conv2d(
        i, k, strides=[1, stride_y, stride_x, 1], padding=padding)

    with tf.variable_scope(name) as scope:
        # Create tf variables for the weights and biases of the conv layer
        weights = tf.get_variable('weights',
                                    shape=[
                                        filter_height, filter_width,
                                        input_channels / groups, num_filters
                                    ])
        biases = tf.get_variable('biases', shape=[num_filters])

    if groups == 1:
        conv = convolve(x, weights)

    # In the cases of multiple groups, split inputs & weights and
    else:
        # Split input and weights and convolve them separately
        input_groups = tf.split(axis=3, num_or_size_splits=groups, value=x)
        weight_groups = tf.split(axis=3,
                                    num_or_size_splits=groups,
                                    value=weights)
        output_groups = [
            convolve(i, k) for i, k in zip(input_groups, weight_groups)
        ]

        # Concat the convolved output together again
        conv = tf.concat(axis=3, values=output_groups)

    # Add biases
    bias = tf.reshape(tf.nn.bias_add(conv, biases), tf.shape(conv))

    # Apply relu function
    relu = tf.nn.relu(bias, name=scope.name)

    return relu


def fc_layer(x, input_size, output_size, name, relu=True, k=20):
    """Create a fully connected layer."""

    with tf.variable_scope(name) as scope:
        # Create tf variables for the weights and biases.
        W = tf.get_variable('weights', shape=[input_size, output_size])
        b = tf.get_variable('biases', shape=[output_size])
        # Matrix multiply weights and inputs and add biases.
        z = tf.nn.bias_add(tf.matmul(x, W), b, name=scope.name)

    if relu:
        # Apply ReLu non linearity.
        a = tf.nn.relu(z)
        return a

    else:
        return z


def max_pool(x,
                filter_height, filter_width,
                stride_y, stride_x,
                name, padding='SAME'):
    """Create a max pooling layer."""
    return tf.nn.max_pool2d(x,
        ksize=[1, filter_height, filter_width, 1],
        strides=[1, stride_y, stride_x, 1],
        padding=padding,
        name=name)


def lrn(x, radius, alpha, beta, name, bias=1.0):
    """Create a local response normalization layer."""
    return tf.nn.local_response_normalization(x,
        depth_radius=radius,
        alpha=alpha,
        beta=beta,
        bias=bias,
        name=name)


def dropout(x, rate):
    """Create a dropout layer."""
    return tf.nn.dropout(x, rate=rate)

你可能感兴趣的:(2023,AI,neo4j)