Detailed instructions for constructing generative adversarial neural networks (GANs) using the example of two models implemented using the PyTorch deep learning framework.

使用通过PyTorch深度学习框架实现的两个模型的示例来构建生成对抗性神经网络(GAN)的详细说明。

Generative adversarial networks (abbreviated GAN) are neural networks that can generate images, music, speech, and texts similar to those that humans do. GANs have become an active research topic in recent years. Facebook AI Lab Director Yang Lekun called adversarial learning “the most exciting machine learning idea in the last 10 years.” Below we will explore how GANs work and create two models using the PyTorch deep learning framework.

生成对抗网络(简称GAN)是神经网络 ，可以生成与人类相似的图像，音乐，语音和文本。 GAN近年来已成为活跃的研究主题。 Facebook AI实验室总监Yang Lekun称对抗学习为“过去10年来最令人兴奋的机器学习理念”。 下面我们将探索GAN的工作原理，并使用PyTorch 深度学习框架创建两个模型。

什么是生成对抗网络？ (What is a Generative Adversarial Network?)

A generative adversarial network (GAN) is a machine learning model that can simulate a given data distribution. The model was first proposed in a 2014 NeurIPS paper by deep learning expert Ian Goodfellow and colleagues.

生成对抗网络 (GAN)是一种可以模拟给定数据分布的机器学习模型。 该模型首次在提出一个 2014 的NeurIPS纸深度学习专家伊恩·古德费洛和同事。

GANs consist of two neural networks, one of which is trained to generate data, and the other is trained to distinguish simulated data from real data (hence the “adversarial” nature of the model). Generative adversarial networks show impressive results in terms of image and video generation:

GAN由两个神经网络组成，其中一个被训练生成数据，另一个被训练以区分模拟数据与真实数据(因此具有模型的“对抗性”性质)。 生成的对抗网络在图像和视频生成方面显示出令人印象深刻的结果：

• transfer of styles (CycleGAN) — the transformation of one image in accordance with the style of other images (for example, paintings by a famous artist);

  样式转移( CycleGAN )-根据其他图像的样式转换一个图像(例如，著名画家的绘画)；

• Human Face Generation (StyleGAN), realistic examples are available at This Person Does Not Exist.

  人脸生成( StyleGAN )， 此人不存在时可以找到现实的示例。

GANs and other data-generating structures are called generative models as opposed to more widely studied discriminative models. Before diving into GANs, let’s look at the differences between the two types of models.

GAN和其他数据生成结构被称为生成模型，与更广泛研究的判别模型相反。 在探讨GAN之前，让我们看一下两种模型之间的差异。

判别式和生成式机器学习模型的比较 (Comparison of Discriminative and Generative Machine Learning models)

Discriminative models are used for most supervised learning problems for classification or regression. As an example of a classification problem, suppose you want to train a handwritten digit image recognition model. To do this, we can use a labeled dataset containing photographs of handwritten numbers to which the numbers themselves are correlated.

判别模型用于大多数监督学习问题的分类或回归 。 作为分类问题的示例，假设您想训练一个手写数字图像识别模型 。 为此，我们可以使用标记的数据集，其中包含与数字本身相关的手写数字的照片。

Training is reduced to setting the parameters of the model using a special algorithm that minimizes the loss function. The loss function is a criterion for the discrepancy between the true value of the estimated parameter and its expectation. After the learning phase, we can use the model to classify a new (previously not considered) handwritten digit image by matching the most likely digit to the input image.

使用一种特殊的算法将训练减少为设置模型的参数，该算法可使损失函数最小化 。 损失函数是估计参数的真实值与其期望之间差异的标准。 在学习阶段之后，我们可以使用模型通过将最可能的数字与输入图像进行匹配来对新的(以前未考虑的)手写数字图像进行分类。

The discriminative model uses training data to find the boundaries between classes. The found boundaries are used to distinguish new inputs and predict their class. Mathematically, discriminative models study the conditional probability P (y | x) of an observation y for a given input x.

判别模型使用训练数据来查找班级之间的界限。 找到的边界用于区分新输入并预测其类别。 在数学上，判别模型研究给定输入x的观察值y的条件概率 P(y | x) 。

Discriminative models are not only neural networks but also logistic regression and support vector machine (SVM).

判别模型不仅是神经网络，而且是逻辑回归和支持向量机(SVM) 。

While discriminative models are used for supervised learning, generative models typically use a RAW data set, that is, may be seen as a form of unsupervised learning. So, using a dataset of handwritten numbers, you can train a generative model to generate new images.

尽管判别模型用于监督学习，但生成模型通常使用RAW数据集，也就是说，可以将其视为无监督学习的一种形式。 因此，使用手写数字数据集，您可以训练生成模型以生成新图像。

In contrast to discriminative models, generative models study the properties of the probability function P (x) of the input data 𝑥 . As a result, they do not generate a prediction, but a new object with properties akin to the training dataset.

与判别模型相反，生成模型研究输入数据the 的概率函数 P(x)的性质 。 结果，它们不生成预测，而是生成具有类似于训练数据集属性的新对象。

Besides GAN, there are other generative architectures:

除了GAN，还有其他生成架构：

• Boltzmann machine

  玻尔兹曼机

• Autoencoder

  自动编码器

• Hidden Markov Model

  隐马尔可夫模型

• Models that predict the next word in a sequence, such as GPT-2

  预测序列中下一个单词的模型，例如GPT-2

GANs have gained a lot of attention recently for their impressive results in visual content generation. Let’s dwell on the device of generative adversarial networks in more detail.

GAN最近在视觉内容生成方面取得了令人印象深刻的结果，因此备受关注。 让我们更详细地讨论生成对抗网络的设备。

生成对抗神经网络的架构 (The architecture of generative adversarial neural networks)

A generative adversarial network, as we have already understood, is not one network, but two: a generator and a discriminator. The role of the generator is to generate a dataset based on a real sample that resembles real data. The discriminator is trained to estimate the probability that the sample is obtained from real data and not provided by a generator. Two neural networks play cat and mouse: the generator tries to trick the discriminator, and the discriminator tries to better identify the generated samples.

正如我们已经了解的那样，生成对抗网络不是一个网络，而是两个：生成器和鉴别器。 生成器的作用是根据类似于真实数据的真实样本生成数据集。 鉴别器经过训练，可以估计从真实数据中获得样本，而不是由发生器提供样本的可能性。 有两个神经网络发挥作用：生成器试图欺骗鉴别器，鉴别器试图更好地识别生成的样本。

To understand how GAN training works, consider a toy example with a dataset consisting of two-dimensional samples (x1, x2), with x1 ranging from 0 to 2π and x2=sin(x1).

要了解GAN训练的工作原理，请考虑一个玩具示例，其中的数据集由二维样本(x1，x2)组成 ， x1的范围是0到2π ， x2 = sin(x1) 。

The general structure of the GAN for generating pairs (x̃1, x̃2) resembling points from a dataset is shown in the following figure.

下图显示了用于从数据集中生成相似点对(x̃1，x̃2)的GAN的一般结构。

The generator receives as input pairs of random numbers (z1, z2), transforming them so that they resemble examples from a real sample. The structure of the neural network can be any, for example, a multilayer perceptron or a convolutional neural network. G G

生成器接收随机数对(z1，z2)作为输入，对其进行转换，使其类似于真实样本中的示例。 神经网络的结构可以是任何结构，例如多层感知器或卷积神经网络 。 G G

The discriminator inputs alternately samples from the training dataset and simulated samples provided by the generator. The role of the discriminator is to assess the likelihood that the input data belongs to a real dataset. That is, training is performed in such a way that it gives out, receiving a real sample, and for the generated sample. D G D 1 0

鉴别器交替输入训练数据集的样本和生成器提供的模拟样本。 鉴别器的作用是评估输入数据属于真实数据集的可能性。 即，以发出，接收真实样本并针对生成的样本的方式进行训练。 D G D 1 0

As in the case with the generator, you can choose any structure of the neural network, taking into account the sizes of the input and output data. In this example, the input is 2D and the output is a scalar ranging from 0 to 1. D

与生成器一样，您可以考虑输入和输出数据的大小来选择神经网络的任何结构。 在此示例中，输入为2D，输出为标量，范围为0到D

Mathematically, the GAN learning process consists of a minimax game of two players, in which it is adapted to minimize the error of the difference between the real and the generated sample, and adapted to maximize the probability of making an error. D G D

在数学上，GAN学习过程由两个玩家组成的极小极大游戏组成，其中，该学习过程适用于最小化真实样本与生成的样本之间的差异的误差，并且适用于最大化发生错误的可能性。 D G D

At each stage of training, the parameters of the models and are updated. To train, at each iteration, we mark a sample of real samples with ones and a sample of generated samples created with zeros. Thus, a normal supervised learning approach can be used to update the parameters as shown in the diagram. D G D G D

在训练的每个阶段，都会更新模型的参数。 为了进行训练，在每次迭代中，我们用1标记真实样本的样本，并用0标记生成的样本的样本。 因此，可以使用正常的监督学习方法来更新参数，如图所示。 D G D G D

For each batch of training data containing tagged real and generated samples, we update the set of model parameters D, minimizing the loss function. After the parameters are Dupdated, we train to Ggenerate better samples. The set of parameters is D"frozen" during the training of the generator.

对于每批包含标记的真实样本和生成样本的训练数据，我们更新模型参数D的集合，从而使损失函数最小化。 在参数D更新之后，我们训练G生成更好的样本。 该组参数是D发电机的训练中“雪藏”。

When it starts to generate samples so well that it is “fooled”, the output probability tends to one — it considers that all samples belong to the original sample. G D D

当它开始很好地生成样本以至于被“愚弄”时，输出概率趋向于一-它认为所有样本都属于原始样本。 G D D

Now that we know how GAN works, we are ready to implement our own neural network using PyTorch.

既然我们知道了GAN的工作原理，我们就可以使用PyTorch来实现我们自己的神经网络了。

您的第一个生成对抗网络 (Your first generative adversarial network)

As the first experiment with generative adversarial networks, we will implement the above example with a harmonic function. To work with the example, we will use the popular PyTorch library, which can be installed using the instructions. If you’re seriously interested in Data Science, you may have already used the Anaconda distribution and the conda package and environment management system . Note that the environment makes the installation process easier.

作为生成对抗网络的第一个实验，我们将使用谐波函数实现上述示例。 为了与例如工作中，我们将使用流行的PyTorch库，它可以使用安装说明 。 如果您对数据科学非常感兴趣，则可能已经使用了Anaconda发行版和conda软件包以及环境管理系统。 请注意，环境使安装过程更容易。

Installing PyTorch with, first create an environment and activate it: conda

安装PyTorch，首先创建一个环境并激活它： conda

$ conda create --name gan
$ conda activate gan

This creates an environment condanamed gan. Inside the created environment, you can install the necessary packages:

这将创建一个名为gan的环境conda 。 在创建的环境中，您可以安装必要的软件包：

$ conda install -c pytorch pytorch=1.4.0
$ conda install matplotlib jupyter

Since PyTorch is an actively developing environment, the API may change in new versions. Code examples have been verified for version 1.4.0.

由于PyTorch是一个积极开发的环境，因此API可能会在新版本中更改。 已经针对1.4.0版验证了代码示例。

We will use matplotlib to work with graphs.

我们将使用matplotlib处理图。

When using Jupyter Notebook, you need to register the environment so that you can create notebooks using this environment as a kernel. To do this, in the activated environment, run the following command: conda gan gan

使用Jupyter Notebook时，您需要注册环境，以便可以使用该环境作为内核来创建笔记本。 为此，请在激活的环境中运行以下命令： conda gan gan

$ python -m ipykernel install --user --name gan

Let’s start by importing the required libraries:

让我们从导入所需的库开始：

import torch
from torch import nnimport math
import matplotlib.pyplot as plt

Here we are importing the PyTorch (torc) library. We import the component separately from the library for more compact handling. The built-in library is only needed to get the value of the constant, and the tool mentioned above is for building dependencies. nn math pi matplotlib

在这里，我们正在导入PyTorch( torc )库。 我们从库中单独导入组件，以进行更紧凑的处理。 仅需要内置库来获取常量的值，并且上面提到的工具用于构建依赖项。 nn math pi matplotlib

It is good practice to temporarily secure the random number generator so that the experiment can be replicated on another machine. To do this in PyTorch, run the following code:

优良作法是暂时保护随机数生成器，以便可以在另一台计算机上复制实验。 要在PyTorch中执行此操作，请运行以下代码：

torch.manual_seed(111)

We use the 111 number to initialize the random number generator. We will need a generator to set the initial weights of the neural network. Despite the random nature of the experiment, its course will be reproducible.

我们使用111号来初始化随机数生成器。 我们将需要一个生成器来设置神经网络的初始权重。 尽管实验具有随机性，但其过程是可重现的。

为GAN训练准备数据 (Preparing data for GAN training)

The training set consists of pairs of numbers (x1, x2) — such that x2 corresponds to the sine value of x1 for x1 in the range from 0 to 2π. Training data can be obtained as follows:

训练集由数字对(x1，x2)组成 -使得x2对应于x1的x1的正弦值，范围从0到2π 。 培训数据可按以下方式获得：

train_data_length = 1024
train_data = torch.zeros((train_data_length, 2))
train_data[:, 0] = 2 * math.pi * torch.rand(train_data_length)
train_data[:, 1] = torch.sin(train_data[:, 0])
train_labels = torch.zeros(train_data_length)
train_set = [(train_data[i], train_labels[i]) for i in range(train_data_length)]

Here we compile a training dataset of 1024 pairs (x1, x2). Then we initialize with zeros — a matrix of 1024 rows and 2 columns. train_data

在这里，我们编译了1024对(x1，x2)的训练数据集。 然后，我们用零初始化-一个1024行2列的矩阵。 train_data

The first column is filled with random values ​​in the range from 0 to 2π. We calculate the values ​​of the second column as the sine of the first. train_data

第一列填充了0到2π范围内的随机值。 我们将第二列的值计算为第一列的正弦值。 train_data

We then formally need an array of labels, which we pass to the PyTorch data loader. Since the GAN implements unsupervised learning, the labels can be anything. train_labels

然后，我们正式需要一个标签数组，并将其传递给PyTorch数据加载器。 由于GAN实施无监督学习，因此标签可以是任何东西。 train_labels

Finally, we create a list of tuples from and. train_data train_labels train_set

最后，我们从和创建一个元组列表。 train_data train_labels train_set

Let’s display the data for training by plotting each point (x1, x2):

让我们通过绘制每个点(x1，x2)来显示训练数据：

plt.plot(train_data[:, 0], train_data[:, 1], ".")

Let’s create a data loader named train_loaderthat will shuffle data from train_set, returning packets of 32 samples ( batch_size) used to train the neural network:

让我们创建一个名为train_loader的数据加载器，它将从train_set随机train_set数据，并返回用于训练神经网络的32个样本( batch_size )的数据包：

batch_size = 32
train_loader = torch.utils.data.DataLoader(train_set, batch_size=batch_size, shuffle=True)

The data is ready, now you need to create the discriminator and GAN neural networks.

数据已经准备就绪，现在您需要创建鉴别器和GAN神经网络。

GAN鉴别器实施 (GAN Discriminator Implementation)

In PyTorch, neural network models are represented by classes that inherit from a class. If you are new to OOP, the article “An Introduction to Object-Oriented Programming (OOP) in Python” will suffice to understand what is happening. nn.Module

在PyTorch中，神经网络模型由继承自类的类表示。 如果您不熟悉OOP，则文章“ Python中的面向对象编程(OOP)简介”将足以了解正在发生的事情。 nn.Module

The discriminator is a two-dimensional input and one-dimensional output model. It takes a sample from real data or from a generator and provides the probability that the sample is from real training data. The code below shows how to create a discriminator class.

鉴别器是二维输入和一维输出模型。 它从真实数据或生成器中获取样本，并提供样本来自真实训练数据的概率。 下面的代码显示了如何创建区分类。

class Discriminator(nn.Module):def __init__(self):super().__init__()self.model = nn.Sequential(nn.Linear(2, 256),nn.ReLU(),nn.Dropout(0.3),nn.Linear(256, 128),nn.ReLU(),nn.Dropout(0.3),nn.Linear(128, 64),nn.ReLU(),nn.Dropout(0.3),nn.Linear(64, 1),nn.Sigmoid())def forward(self, x):output = self.model(x)return output

A standard class method is used to build a neural network model . Inside this method, we first call to run the corresponding method of the inherited class . A multilayer perceptron is used as the architecture of the neural network . Its structure is specified in layers using . The model has the following characteristics: __init__() super().__init__() __init__() nn.Module nn.Sequential()

使用标准的类方法来建立神经网络模型。 在此方法内部，我们首先调用以运行继承的类的相应方法。 多层感知器被用作神经网络的体系结构。 使用分层指定其结构。 该模型具有以下特征： __init__() super().__init__() __init__() super().__init__() __init__() nn.Module nn.Sequential()

• two-dimensional entrance;
  二维入口；

• the first hidden layer consists of 256 neurons and has a ReLU activation function ;

  第一隐蔽层256个包括神经元的，并且具有一个 RELU 激活功能 ;

• in the subsequent layers, the number of neurons decreases to 128 and 64. The output has a sigmoidal activation function, which is characteristic of representing the probability ( Sigmoid);

  在随后的层中，神经元的数量减少到128和64。输出具有S型激活函数，该函数表示概率( Sigmoid )。

• to avoid overfitting, after the first, second and third hidden layers, a part of the neurons is dropped ( Dropout).

  为了避免过度拟合，在第一，第二和第三隐藏层之后，将掉落一部分神经元( Dropout )。

For the convenience of inference, a method is also created in the class. Here corresponds to the input of the model. In this implementation, the output is obtained by feeding the input into the model we have defined without preprocessing. forward() x x

为了方便推断，在类中还创建了一个方法。 这里对应于模型的输入。 在此实现中，通过将输入输入我们定义的模型而无需预处理即可获得输出。 forward() x x

After declaring the discriminator class, create an instance of it:

在声明了鉴别器类之后，创建它的一个实例：

discriminator = Discriminator()

GAN生成器的实现 (GAN generator implementation)

In generative adversarial networks, a generator is a model that takes as input some sample from a space of hidden variables that resemble the data in the training set. In our case, this is a 2D input model that will receive random points (z1, z2), and a 2D output that produces points (x̃1, x̃2) that look like the points from the training data.

在生成对抗网络中，生成器是一个模型，该模型从类似于训练集中数据的隐藏变量空间中获取一些样本作为输入。 在我们的例子中，这是一个2D输入模型，它将接收随机点(z1，z2) ，并在一个2D输出中生成看起来像训练数据中的点的点(x̃1，x̃2) 。

The implementation is similar to what we wrote for the discriminator. First, you need to create a class that inherits from, then define the architecture of the neural network, and finally create an instance of the object : Generator nn.Module Generator

实现与我们为鉴别器编写的类似。 首先，您需要创建一个继承自的类，然后定义神经网络的体系结构，最后创建该对象的实例： Generator nn.Module Generator

class Generator(nn.Module):def __init__(self):super().__init__()self.model = nn.Sequential(nn.Linear(2, 16),nn.ReLU(),nn.Linear(16, 32),nn.ReLU(),nn.Linear(32, 2))def forward(self, x):output = self.model(x)return outputgenerator = Generator()

The generator includes two hidden layers with 16 and 32 neurons with the ReLU activation function, and at the output a layer with two neurons with a linear activation function. Thus, the output will consist of two elements ranging from −∞ to + ∞ , which will represent (x̃1 , x̃2). That is, initially we do not impose any restrictions on the generator — it must “learn everything by itself.”

生成器包括两个具有ReLU激活功能的具有16和32个神经元的隐藏层，并在输出端包含一个具有线性激活功能的两个神经元的层。 因此，输出将包括从-∞到+∞的两个元素，它们表示(x̃1，x̃2) 。 也就是说，起初我们不对生成器施加任何限制-它必须“自己学习一切”。

Now that we have defined the models for the discriminator and generator, we are ready to start training.

既然我们已经为鉴别器和生成器定义了模型，我们就可以开始训练了。

训练GAN模型 (Train GAN Models)

Before training the models, you need to configure the parameters that will be used in the training process:

在训练模型之前，您需要配置将在训练过程中使用的参数：

lr = 0.001
num_epochs = 300
loss_function = nn.BCELoss()

What’s going on here:

这里发生了什么：

1. We set the learning rate, which we will use to adapt the network weights. lr

   我们设置学习率，我们将使用它来调整网络权重。 lr

2. We set the number of epochs, which determines how many repetitions of the training process will be performed using the entire dataset. num_epochs

   我们设置时期的数量，该时期确定使用整个数据集将执行多少次重复训练过程。 num_epochs

3. To the variable, we assign the function of the logistic loss function (binary cross-entropy). This is the loss function that we will use to train the models. It is suitable both for training a discriminator (its task is reduced to a binary classification) and for a generator since it feeds its output to the input of the discriminator. loss_function BCELoss()

   为变量分配逻辑损失函数 (二进制交叉熵)的函数 。 这是我们用来训练模型的损失函数。 它既适合于训练鉴别器(将其任务简化为二进制分类)，也适合于生成器，因为它将其输出馈送到鉴别器的输入。 loss_function BCELoss()

The rules for updating weights (training the model) in PyTorch are implemented in a module. We will use Adam’s stochastic gradient descent algorithm to train discriminator and generator models. To create optimizers, run the following code: torch.optim torch.optim

在PyTorch中更新权重(训练模型)的规则在模块中实现。 我们将使用亚当的随机梯度下降算法来训练鉴别器和生成器模型。 要创建优化器，请运行以下代码： torch.optim torch.optim

optimizer_discriminator = torch.optim.Adam(discriminator.parameters(), lr=lr)
optimizer_generator = torch.optim.Adam(generator.parameters(), lr=lr)

Finally, it is necessary to implement a training cycle in which samples of the training sample are fed to the model input, and their weights are updated, minimizing the loss function:

最后，有必要实施一个训练周期，在该训练周期中，将训练样本的样本馈送到模型输入，并更新其权重，以最小化损失函数：

for epoch in range(num_epochs):for n, (real_samples, _) in enumerate(train_loader):# Data for descriminator trainingreal_samples_labels = torch.ones((batch_size, 1))latent_space_samples = torch.randn((batch_size, 2))generated_samples = generator(latent_space_samples)generated_samples_labels = torch.zeros((batch_size, 1))all_samples = torch.cat((real_samples, generated_samples))all_samples_labels = torch.cat((real_samples_labels, generated_samples_labels))# Discriminator trainingdiscriminator.zero_grad()output_discriminator = discriminator(all_samples)loss_discriminator = loss_function(output_discriminator, all_samples_labels)loss_discriminator.backward()optimizer_discriminator.step()# Data for generator traininglatent_space_samples = torch.randn((batch_size, 2))# Generator traininggenerator.zero_grad()generated_samples = generator(latent_space_samples)output_discriminator_generated = discriminator(generated_samples)loss_generator = loss_function(output_discriminator_generated, real_samples_labels)loss_generator.backward()optimizer_generator.step()# Output value of loss functionif epoch % 10 == 0 and n == batch_size - 1:print(f"Epoch: {epoch} Loss D.: {loss_discriminator}")print(f"Epoch: {epoch} Loss G.: {loss_generator}")

Here, at each training iteration, we update the discriminator and generator parameters. As is usually done for neural networks, the training process consists of two nested loops: the outer one for the training epochs, and the inner one for the packets within each epoch. In the inner loop, it all starts with preparing data for training the discriminator:

在这里，在每次训练迭代中，我们都会更新鉴别器和生成器参数。 就像神经网络通常所做的那样，训练过程包括两个嵌套循环：一个用于训练时期的外部循环，另一个用于在每个时期内的数据包的内部循环。 在内部循环中，一切都始于准备数据以训练鉴别器：

• We get real samples of the current batch from the data loader and assign them to a variable. Note that the first dimension in the array dimension has the number of elements equal to. This is the standard way of organizing data in PyTorch, where each tensor row represents one sample from the package. real_samples batch_size

  我们从数据加载器获取当前批次的真实样本，并将其分配给变量。 请注意，数组维度中的第一个维度的元素数量等于。 这是在PyTorch中组织数据的标准方法，其中每个张量行代表包装中的一个样品。 real_samples batch_size

• Use to create labels with a value of 1 for real samples and assign labels to a variable. torch.ones() real_samples_labels

  用于为实际样本创建值为1的标签，并将标签分配给变量。 torch.ones() real_samples_labels

• We generate samples by storing random data, which we then pass to the generator for receiving. We use zeros for the labels of the generated samples, which we store in latent_space_samples generate_samples torch.zeros() generate_samples_labels

  我们通过存储随机数据来生成样本，然后将其传递给生成器以进行接收。 我们使用零作为生成样本的标签，并将其存储在latent_space_samples generate_samples torch.zeros() generate_samples_labels

• It remains to combine the real and generated samples and labels and save respectively in andall_samples all_samples_labels

  仍然需要合并实际样本和生成的样本和标签，并分别保存在all_samples和all_samples_labels

In the next block, we train the discriminator:

在下一个步骤中，我们训练鉴别器：

• In PyTorch, it is important to clear the gradient values ​​at every step of the training. We do this using the method zero_grad()

  在PyTorch中，重要的是在训练的每个步骤中清除梯度值。 我们使用zero_grad()方法执行此操作

• We calculate the output of the discriminator using the training data all_samples

  我们使用训练数据all_samples计算鉴别器的输出

• Calculate the value of the loss function using the output at and labels output_discriminator all_samples_labels

  使用和处的输出计算损失函数的值，并标记output_discriminator all_samples_labels

• Calculate the gradients to update the weights withloss_discriminator.backward()

  计算梯度以使用loss_discriminator.backward()更新权重

• Find the updated discriminator weights by calling optimizer_discriminator.step()

  通过调用optimizer_discriminator.step()查找更新的鉴别器权重

• We prepare the data for training the generator. We use two columns to match the 2D data at the generator input. latent_space_samples batch_size

  我们准备数据来训练生成器。 我们使用两列来匹配生成器输入处的2D数据。 latent_space_samples batch_size

We train the generator:

我们训练发电机：

• We clean up the gradients using the method. zero_grad()

  我们使用该方法清理渐变。 zero_grad()

• We pass it on to the generator and save its output to latent_space_samples generate_samples

  我们将其传递给生成器，并将其输出保存到latent_space_samples generate_samples

• We pass the generator output to the discriminator and save its output, which will be used as the output of the entire model. output_discriminator_generated

  我们将生成器的输出传递给鉴别器，并保存其输出，该输出将用作整个模型的输出。 output_discriminator_generated

• Calculate the loss function using the output of the classification system stored in and labels equal to 1. output_discriminator_generated real_samples_labels

  使用存储在所述分类系统的输出计算所述损失函数和标签等于1 output_discriminator_generated real_samples_labels

• Calculating gradients and updating generator weights. Remember that when we train the generator, we are keeping the discriminator weights frozen.
  计算梯度并更新发电机权重。 请记住，在训练发电机时，我们要保持鉴别器的重量冻结。

Finally, in the last lines of the loop, the values ​​of the discriminator and generator loss functions are output at the end of every tenth epoch.

最后，在循环的最后几行中，在每个第十个周期的末尾输出鉴别器和发电机损耗函数的值。

检查GAN生成的样本 (Checking samples generated by GAN)

Generative adversarial networks are designed to generate data. Thus, after the training process is complete, we can call the generator to get new data:

生成对抗网络旨在生成数据。 因此，在训练过程完成之后，我们可以调用生成器以获取新数据：

latent_space_samples = torch.randn(100, 2)
generated_samples = generator(latent_space_samples)

Let’s plot the generated data and check how similar it is to the training data. Before plotting a graph for the generated samples, you need to apply a method detach()to get the necessary data from the PyTorch computational graph:

让我们绘制生成的数据，并检查它与训练数据的相似程度。 在为生成的样本绘制图形之前，您需要应用detach()方法从PyTorch计算图形中获取必要的数据：

generated_samples = generated_samples.detach()
plt.plot(generated_samples[:, 0], generated_samples[:, 1], ".")

The distribution of the generated data is very similar to real data — the original sine. The animation of the evolution of learning can be viewed here .

生成的数据的分布与真实数据(原始正弦)非常相似。 学习演变的动画可以在这里查看 。

At the beginning of the training process, the distribution of the generated data is very different from the real data. But as it learns, the generator learns the real data distribution, as if adjusting to it.

在训练过程开始时，生成的数据的分布与实际数据有很大不同。 但是，随着学习，生成器将学习实际的数据分布，就像对其进行调整一样。

Now that we have implemented the first model of a generative adversarial network, we can move on to a more practical example of generating images.

现在，我们已经实现了生成对抗网络的第一个模型，我们可以继续进行生成图像的更实际示例。

带有GAN的手写数字生成器 (Handwritten Digit Generator with GAN)

In the following example, we will use GAN to generate images of handwritten numbers. To do this, we will train the models using the MNIST dataset of handwritten numbers. This standard dataset is included in the package torchvision

在下面的示例中，我们将使用GAN生成手写数字的图像。 为此，我们将使用MNIST手写数字数据集训练模型。 这个标准数据集包含在torchvision软件包中

First, in the activated environment, you need to install : gan torchvision

首先，在激活的环境中，您需要安装： gan torchvision

$ conda install -c pytorch torchvision=0.5.0

Again, here we are specifying the specific version just like we did with PyTorch to ensure that the code examples run. torchvision

同样，在这里我们指定特定版本，就像我们使用PyTorch一样以确保代码示例运行。 torchvision

We start by importing the required libraries:

我们首先导入所需的库：

import torchvision
import torchvision.transforms as transformstorch.manual_seed(111)

In addition to the libraries that we imported earlier, we will also need to transform the information stored in image files. torchvision torchvision.transforms

除了我们先前导入的库之外，我们还需要转换存储在图像文件中的信息。 torchvision torchvision.transforms

Since the training set includes images in this example, the models will be more complex and the training will take significantly longer. When training in a central processing unit (CPU), one epoch will take about two minutes. It will take about 50 epochs to get an acceptable result, so the total training time using the processor is about 100 minutes.

由于在此示例中训练集包括图像，因此模型将更加复杂，并且训练将花费更长的时间。 在中央处理器( CPU )中进行训练时，一个历时大约需要2分钟。 获得可接受的结果大约需要50个纪元，因此使用处理器的总训练时间约为100分钟。

A graphics processing unit (GPU) can be used to reduce training time.

图形处理单元( GPU )可用于减少训练时间。

To make the code work regardless of the characteristics of the computer, let’s create an object that will point either to the central processor or (if available) to the graphics processor: device

为了使代码工作不管计算机的特性，让我们创建一个对象，要么指向中央处理器或(如果可用)到图形处理器： device

device = ""
if torch.cuda.is_available():device = torch.device("cuda")
else:device = torch.device("cpu")

The environment is configured, let’s prepare a dataset for training.

环境已配置，让我们准备训练数据集。

准备MNIST数据集 (Preparing the MNIST dataset)

The MNIST dataset consists of images of handwritten digits 0 through 9. The images are in grayscale and are 28 × 28 pixels in size. To use them with PyTorch, you need to do some transformations. To do this, we define the function used when loading data: transform

MNIST数据集由手写数字0到9的图像组成。图像为灰度图像 ，大小为28×28像素 。 要将它们与PyTorch一起使用，您需要进行一些转换。 为此，我们定义了加载数据时使用的函数： transform

transform = transforms.Compose([transforms.ToTensor(), transforms.Normalize((0.5,), (0.5,))])

The function has two parts:

该功能分为两部分：

1. transforms.ToTensor() converts the data into a PyTorch tensor.

   transforms.ToTensor()将数据转换为PyTorch张量。

2. transforms.Normalize() converts a range of tensor coefficients.

   transforms.Normalize()转换一系列张量系数。

The original coefficients are given by the function range from 0 to 1. Since the images have a black background, most of the coefficients are 0. transforms.ToTensor()

原始系数由0到1的函数范围给出。由于图像具有黑色背景，因此大多数系数为0。transforms.ToTensor transforms.ToTensor()

The function changes the range of coefficients to transforms.Normalize() [ - 1 , 1 ][−1,1] , subtracting 0.5 from the original odds and dividing the result by 0.5. The transformation reduces the number of elements in the input samples to zero. This helps in training the models.

该函数将系数的范围更改为transforms.Normalize() [-1，1] [-1,1] ，从原始几率中减去0.5，然后将结果除以0.5。 该变换将输入样本中的元素数量减少为零。 这有助于训练模型。

We can now load the training data by calling : torchvision.datasets.MNIST

现在，我们可以通过调用以下torchvision.datasets.MNIST加载训练数据： torchvision.datasets.MNIST

train_set = torchvision.datasets.MNIST(root=".", train=True, download=True, transform=transform)

The argument ensures that the first time you run the code, the MNIST dataset will be loaded and saved in the current directory as specified in the argument. download = True root

该参数确保第一次运行代码时，MNIST数据集将按照参数指定的方式加载并保存在当前目录中。 download = True root

We created so that we can create a data loader as we did before: train_set

我们创建后可以像以前一样创建数据加载器： train_set

batch_size = 32
train_loader = torch.utils.data.DataLoader(train_set, batch_size=batch_size, shuffle=True)

Let’s use matplotlibfor the selective plotting of data. Well suited as a palette cmap = gray_r. The numbers will be displayed in black on a white background:

让我们使用matplotlib选择性绘制数据。 非常适合用作调色板cmap = gray_r 。 数字将在白色背景上以黑色显示：

real_samples, mnist_labels = next(iter(train_loader))
for i in range(16):ax = plt.subplot(4, 4, i + 1)plt.imshow(real_samples[i].reshape(28, 28), cmap="gray_r")plt.xticks([])plt.yticks([])

As you can see, the dataset contains numbers with different handwriting. As the GAN learns the distribution of the data, it also generates numbers with different handwriting styles.

如您所见，数据集包含具有不同笔迹的数字。 GAN在学习数据分布时，还会生成具有不同笔迹样式的数字。

We have prepared training data, we can implement discriminator and generator models.

我们已经准备了训练数据，可以实现鉴别器和生成器模型。

鉴别器和生成器的实现 (Discriminator and generator implementation)

In this case, the discriminator is a multilayer perceptron neural network, which takes an image of 28 × 28 pixels and finds the probability that the image belongs to real training data.

在这种情况下，鉴别器是多层​​感知器神经网络，它拍摄28×28像素的图像，并找到该图像属于真实训练数据的概率。

class Discriminator(nn.Module):def __init__(self):super().__init__()self.model = nn.Sequential(nn.Linear(784, 1024),nn.ReLU(),nn.Dropout(0.3),nn.Linear(1024, 512),nn.ReLU(),nn.Dropout(0.3),nn.Linear(512, 256),nn.ReLU(),nn.Dropout(0.3),nn.Linear(256, 1),nn.Sigmoid(),)def forward(self, x):x = x.view(x.size(0), 784)output = self.model(x)return output

To introduce the image coefficients into the perceptron neural network, it is necessary to vectorize them so that the neural network receives a vector consisting of 784 coefficients (28 × 28 = 784).

要将图像系数引入感知器神经网络，必须对其进行矢量化处理，以使神经网络接收包含784个系数( 28×28 = 784 )的向量。

Vectorization occurs in the first line of the method — the call transforms the form of the input tensor. Initial tensor form forward() x.view() 𝑥, where 32 is the batch size. After transformation, the form 32 × 1 × 28 × 28 𝑥x becomes equal, with each row representing the image coefficients of the training set. 32 × 784

向量化发生在方法的第一行中-调用将转换输入张量的形式。 初始张量形式forward() x.view() 𝑥 ，其中32是批处理大小。 转换后，形式为32 × 1 × 28 × 28 𝑥x变得相等，每一行代表训练集的图像系数。 32 × 784

To run a discriminator model using a GPU, you need to instantiate it and associate it with a device object using the method : to()

要使用GPU运行鉴别器模型，您需要实例化它并使用以下方法to()其与设备对象关联： to()

discriminator = Discriminator().to(device=device)

The generator will create more complex data than the previous example. Therefore, it is necessary to increase the size of the input data used for initialization. Here we are using a 100-dimensional input and output with 784 coefficients. The result is organized as a 28x28 tensor representing the image.

生成器将创建比上一个示例更复杂的数据。 因此，有必要增加用于初始化的输入数据的大小。 在这里，我们使用具有784个系数的100维输入和输出。 结果组织为代表图像的28x28张量 。

class Generator(nn.Module):def __init__(self):super().__init__()self.model = nn.Sequential(nn.Linear(100, 256),nn.ReLU(),nn.Linear(256, 512),nn.ReLU(),nn.Linear(512, 1024),nn.ReLU(),nn.Linear(1024, 784),nn.Tanh(),)def forward(self, x):output = self.model(x)output = output.view(x.size(0), 1, 28, 28)return outputgenerator = Generator().to(device=device)

The output coefficients must be in the range from -1 to 1. Therefore, at the output of the generator, we use a hyperbolic activation function. In the last line, we instantiate the generator and associate it with the device object. Tanh()

输出系数必须在-1到1的范围内。因此，在生成器的输出处，我们使用双曲线激活函数。 在最后一行，我们实例化了生成器并将其与设备对象相关联。 Tanh()

It remains only to train the model.

它仅用于训练模型。

模型训练 (Model training)

To train models, you need to define training parameters and optimizers:

要训​​练模型，您需要定义训练参数和优化器：

lr = 0.0001
num_epochs = 50
loss_function = nn.BCELoss()optimizer_discriminator = torch.optim.Adam(discriminator.parameters(), lr=lr)
optimizer_generator = torch.optim.Adam(generator.parameters(), lr=lr)

We are reducing the learning rate compared to the previous example. To shorten the training time, set the number of epochs to 50.

与前面的示例相比，我们正在降低学习率。 要缩短训练时间，请将纪元数设置为50。

The learning loop is similar to the one we used in the previous example:

学习循环类似于上一个示例中使用的循环：

for epoch in range(num_epochs):for n, (real_samples, mnist_labels) in enumerate(train_loader):# Data for discriminator trainingreal_samples = real_samples.to(device=device)real_samples_labels = torch.ones((batch_size, 1)).to(device=device)latent_space_samples = torch.randn((batch_size, 100)).to(device=device)generated_samples = generator(latent_space_samples)generated_samples_labels = torch.zeros((batch_size, 1)).to(device=device)all_samples = torch.cat((real_samples, generated_samples))all_samples_labels = torch.cat((real_samples_labels, generated_samples_labels))# Discriminator trainingdiscriminator.zero_grad()output_discriminator = discriminator(all_samples)loss_discriminator = loss_function(output_discriminator, all_samples_labels)loss_discriminator.backward()optimizer_discriminator.step()# Data for generator traininglatent_space_samples = torch.randn((batch_size, 100)).to(device=device)# Generator traininggenerator.zero_grad()generated_samples = generator(latent_space_samples)output_discriminator_generated = discriminator(generated_samples)loss_generator = loss_function(output_discriminator_generated, real_samples_labels)loss_generator.backward()optimizer_generator.step()# Output value of loss functionif n == batch_size - 1:print(f"Epoch: {epoch} Loss D.: {loss_discriminator}")print(f"Epoch: {epoch} Loss G.: {loss_generator}")

检查生成的GAN样本 (Checking Generated GAN Samples)

Let’s generate some samples of “handwritten numbers”. To do this, we pass the generator an initiating set of random numbers:

让我们生成一些“手写数字”的样本。 为此，我们向生成器传递一组初始的随机数：

latent_space_samples = torch.randn(batch_size, 100).to(device=device)
generated_samples = generator(latent_space_samples)

To build the generated samples, you need to move the data back to the central processor, if it was processed on the GPU. To do this, just call the method cpu(). As before, before plotting the data, you need to call the method detach():

要构建生成的样本，您需要将数据移回中央处理器(如果已在GPU上处理过)。 为此，只需调用方法cpu() 。 和以前一样，在绘制数据之前，您需要调用detach()方法：

generated_samples = generated_samples.cpu().detach()for i in range(16):ax = plt.subplot(4, 4, i + 1)plt.imshow(generated_samples[i].reshape(28, 28), cmap="gray_r")plt.xticks([])plt.yticks([])

The output should be numbers that resemble training data.

输出应为类似于训练数据的数字。

After fifty epochs of learning, there are several numbers, as if written by a human hand. Results can be improved with longer training times (with more epochs). As in the previous example, you can visualize the evolution of training by using a fixed tensor of the input and feeding it to the generator at the end of each epoch (animation of the evolution of training).

在学习了五十个纪元之后，有几个数字，就好像是人手所写的一样。 训练时间更长(时间越长)，结果越好。 与前面的示例一样，您可以通过使用固定的输入张量并在每个纪元末尾将其馈送到生成器( 训练演变的动画 )来可视化训练的演变 。

At the beginning of the training process, the generated images are completely random. As it learns, the generator learns the distribution of real data, and after about twenty epochs some of the generated digit images already resemble real data.

在训练过程的开始，生成的图像是完全随机的。 在学习过程中，生成器学习了真实数据的分布，并且在大约二十个纪元后，一些生成的数字图像已经类似于真实数据。

结论 (Conclusion)

Congratulations! You have learned how to implement your own generative adversarial network. We first built a toy example to understand the structure of the GAN, and then looked at a network for generating images from the available sample data.

恭喜你！ 您已经了解了如何实现自己的生成对抗网络。 我们首先构建了一个玩具示例来了解GAN的结构，然后查看了一个网络，该网络用于从可用的样本数据生成图像。

Despite the complexity of the GAN topic, machine learning frameworks like PyTorch make implementation very easy.

尽管GAN主题很复杂，但是像PyTorch这样的机器学习框架使实现非常容易。