This article is aimed at people who want to learn or review how to build a basic Convolutional Neural Network in Keras. The dataset in which this article is based on is the Fashion-Mnist dataset.

本文针对想要学习或复习如何在Keras中构建基本卷积神经网络的人们。 本文所基于的数据集是Fashion-Mnist数据集 。

Along with this article, we will explain how:

与本文一起，我们将解释如何：

1. To build a basic CNN in Keras.在Keras中建立基本的CNN。
2. To run the neural networks.运行神经网络。
3. To save and load checkpoints.保存和加载检查点。

数据集描述 (Dataset description)

Fashion-MNIST is a dataset of Zalando’s article images — consisting of a training set of 60,000 examples and a test set of 10,000 examples. Each example is a 28x28 grayscale image, associated with a label from 10 classes. We intend Fashion-MNIST to serve as a direct drop-in replacement for the original MNIST dataset for benchmarking machine learning algorithms. It shares the same image size and structure of training and testing splits.

Fashion-MNIST是Zalando文章图片的数据集-包含60,000个示例的训练集和10,000个示例的测试集。 每个示例都是一个28x28灰度图像，与来自10个类别的标签相关联。 我们打算将Fashion-MNIST用作直接替代MNIST原始数据集的基准机器学习算法。 它具有相同的图像大小以及训练和测试分割的结构。

加载数据 (Loading the data)

In order to run the code displayed below, it is necessary to download the following files. Once downloaded the data, you can load it using the following code.

为了运行下面显示的代码，必须下载以下文件 。 下载数据后，您可以使用以下代码加载数据。

数据标准化 (Data normalization)

Then, we rescale the images from 0–255 to 0–1 by dividing the data by 255.

然后，通过将数据除以255，我们将图像从0-255缩放到0-1。

前处理 (Pre-processing)

Before loading the data inside the neural network, it is necessary to reshape the images to the correct format that Keras requires. When using a 2D convolution as the first layer in a model, the default shape is (batch_size, height, width, channels); (no_data, 128, 128, 3) for 128x128 RGB pictures.

在将数据加载到神经网络内部之前，有必要将图像重塑为Keras所需的正确格式。 当使用2D卷积作为模型的第一层时，默认形状为(batch_size，height，width，channels)； ( no_data，128，128，3 )用于128x128 RGB图片。

The images of our dataset are grayscale images, in which the value of each pixel is a single sample representing only an amount of light. Therefore, the shape of the training data has to be (no_data, 28, 28, 1).

我们的数据集的图像是灰度图像，其中每个像素的值是一个仅代表光量的单个样本。 因此，训练数据的形状必须为( no_data，28，28，1 )。

建立神经网络 (Building the neural network)

For this article, I built a neural network using two 2D convolutions layers and then two fully connected layers. When declaring the 2D convolutional layers, it is possible/necessary to indicate some parameters. Also, remember to recheck the input_shape. Most errors come from not declaring it right.

对于本文，我使用两个2D卷积层以及两个完全连接的层构建了一个神经网络。 在声明2D卷积层时，有可能/有必要指出一些参数。 另外，请记住要重新检查input_shape。 大多数错误来自未正确声明。

Arguments:

参数：

• filters: Integer, the dimensionality of the output space (i.e. the number of output filters in the convolution).

  过滤器 ：整数，输出空间的维数(即卷积中输出过滤器的数量)。

• kernel_size: An integer or tuple/list of 2 integers, specifying the height and width of the 2D convolution window. Can be a single integer to specify the same value for all spatial dimensions.

  kernel_size ：一个整数或2个整数的元组/列表，指定2D卷积窗口的高度和宽度。 可以是单个整数，以为所有空间尺寸指定相同的值。

• activation: Activation function to use. The typical activations used for all the layers but the last one is ‘relu’ activation. For the last layer, the most used activation is the ‘softmax’ activation.

  activation ：要使用的激活功能。 除最后一层外，所有层都使用的典型激活是“ relu”激活。 对于最后一层，最常用的激活是“ softmax”激活。

• pool_size: integer or tuple of 2 integers, window size over which to take the maximum. (2, 2) will take the max value over a 2x2 pooling window. If only one integer is specified, the same window length will be used for both dimensions.

  pool_size ：2个整数的整数或元组，最大的窗口大小。 (2, 2)将采用2x2合并窗口中的最大值。 如果仅指定一个整数，则两个尺寸将使用相同的窗口长度。

This is the architecture of the built neural network.

这是内置神经网络的体系结构。

训练神经网络 (Training the neural network)

For training the neural network, you can run the following code. Apart from training and evaluating the validation set, it will also save logs that can be afterward loaded into Tensorboard.

为了训练神经网络，您可以运行以下代码。 除了训练和评估验证集外，它还将保存日志，然后可以将其加载到Tensorboard中。

Besides, this code will save (1) the weights of the model for each epoch, and (2) the weights of the model with maximum accuracy.

此外，此代码还将节省(1)每个时期的模型权重，以及(2)以最大精度保存的模型权重。

Saving the best model is interesting because the last epoch is not always that one that performed best (e.g., the model is overfitting).

保存最佳模型很有趣，因为最后一个时期并不总是表现最好的那个时期(例如，模型过度拟合)。

加载并重新运行模型 (Loading and re-running the model)

Since we saved the models in the step before, now it is possible to load it and keep training the neural network.

由于我们在之前的步骤中保存了模型，因此现在可以加载模型并继续训练神经网络。

The code to load the model is the following.

下面是加载模型的代码。

To keep training the neural network from the point we left it before, just run the following code after loading the model.

为了从之前离开的地方继续训练神经网络，只需在加载模型后运行以下代码即可。

绘制结果 (Plotting the results)

The results are displayed in the following figures.

结果显示在下图中。

The classification results are displayed in the following table.

分类结果显示在下表中。