There are many ways to facilitate the creation of large-scale projects. One of the most commonly used methods is to create virtual machines that contain the program environment. However, software has recently been created to make this process even easier. One example is Docker, a software for automating the deployment and management of applications in an operating system-level virtualization environment. This paper discusses the Docker software, its features and benefits, which allows you to create images that contain the program and all the necessary components for its operation. The purpose of this work is to study the capabilities of Docker. And also, the creation of a container containing a software implementation of the neural network for recognition of various handwritten characters. Training and test data is a database of handwritten numbers and letters "MNIST" and "EMNIST". To teach the neural network to recognize numbers, a training set containing 60 thousand copies was used, and the test set includes 10 thousand copies. For letters - the training set contains 88800 copies, and the test set includes 14800 copies. The project was created on the basis of the image tensorflow downloaded from public Docker-registry Docker Hub. The program is written in Python3, using the service for interactive computing Jupyter Notebook.
Copyright © 2019 for the individual papers by the papers’ authors. Copying permitted for private and academic purposes. This volume is published and copyrighted by its editors.
In: K. E. Samouylov, L. A. Sevastianov, D. S. Kulyabov (eds.): Selected Papers of the IX Conference “Information and Telecommunication Technologies and Mathematical Modeling of High-Tech Systems”, Moscow, Russia, 19-Apr-2019, published at http://ceur-ws.org
In this paper we use Docker to automate the deployment and management of
applications. Docker is collection of programs used to run processes in an isolated
environment based on special images [
The main components of Docker are images, registry and containers.
Docker image is a read-only template. For example, an image might contain an operating system with installed applications. Images are used to create containers. Docker makes it easy to create a new image and update existing ones or download ready-only images.
Docker-registry stores images. There are public and private registries. One can download images from registry or upload images to registry. There is free public Docker registry called Docker Hub. It is accessible for all users.
Containers are similar to directories. They contain everything to run the application. Each container is created from an image, which forms an isolated and a secure platform for the application.
Docker container is created by following command $ docker run <attributes> <image> <command>
One of the advantages of Docker is the ability to create a container using Dockerfile and the docker build command. Dockerfile contains the base image. This image is used to build the desired container by applying commands, specified in Dockefile.
To demonstrate the capabilities of Docker, a project was created using a neural network for handwriting recognition.
Artificial neural network is a mathematical model built on the principle of biological
neural networks of nerve cells of a living organism. A neural network is constructed
in such a way that its nodes work like neurons in the human brain. The node collects
information, processes it, and passes it to the next node [
For the implementation of neural networks, there is a huge amount of software. The main diferences between them are their functionality. While some frameworks are used as shells to extend functionality and facilitate the writing of neural networks, others are full-fledged languages and can define neural networks of any level of complexity.
TensorFlow is library designed for machine learning. This framework is created by Google. It uses Python languages as back-end, but core parts are written in C++ [?,13,14]. We use TensorFlow to train and build a neural network for classifying and ifnding images that are close to human perception. All calculations in this environment are performed using stream data graphs, where nodes represent diferent mathematical operations and graph branches represent arrays of data.
The Keras library [
The Scikit-learn [
The convolutional neural network [
Convolutional neural network (CNN) consists of the following types of layers: convolutional layers, subsampling (or pooling) layers and perceptron layers. The first two types of layers, alternating with each other, form the input feature vector for the multilayer perceptron.
In convolutional neural network a convolution operation uses a limited matrix of weights of small size, which moves through the processed layer, forming after each shift activation signal for the next layer of the neuron with a similar position.
This weight matrix is called the convolution kernel. In a convolutional neural network, sets of weights encoding image elements are formed independently by training the network. After a convolutional layer comes the pooling layer. It also has maps, the number of which coincides with the previous layer. The purpose of this layer is to reduce the dimension of the maps. In the previous convolution operation, signs were identified that do not require such detail. Filtering already unnecessary parts helps the network avoid overtraining. After several repetitions of convolutional layers and pooling layers, a layer of the usual multilayer perceptron follows. The output layer is connected to all neurons of the previous layer and the number of neurons corresponds to the number of recognized classes. In the case of binary classification, a single neuron and hyperbolic tangent can be used as an activation function. Then, the output of a neuron with a value of 1 means belonging to a class, and the output of a neuron with a value of -1 means not belonging to a class.
The following architecture was used for the convolution network to recognize digits (fig. 1).
This network consists of 6 layers: 1. Convolutional layer with 75 feature maps, convolution kernel size: 5x5. 2. Pooling layer (MaxPooling) with 2x2 poolsize. 3. Another convolution layer with 100 feature maps, convolution kernel size: 5x5. 4. Second pooling layer with 2x2 poolsize. 5. Fully connected layer with 500 neurons 6. Fully connected output layer with 10 neuron, which correspond to the classes of handwritten digits from 0 to 9.
The activation function in hidden layers is ReLU, and the output layer is softmax.
As training and test data we use "MNIST" database of handwritten numbers. To train the neural network to recognize numbers we use a training set containing 60 thousand copies and the test set containing 10 thousand copies.
We use Matplotlib library to create an image of the symbols from the database 2. 3.
One of the concepts for describing metrics in terms of classification errors is the
error matrix. It is used if there are two classes and an algorithm predicting that each
object belongs to one of the classes, then the classification error matrix will look like
this [
One of the simplest metrics is accuracy — the proportion of true results (both positive and true negative) among the total number of cases considered, i.e. the probability that the class will be predicted correctly.
=
+ + + +
.
To assess the quality of the algorithm on each of the classes precision and recall metrics are used: =
+ , =
+ .
The accuracy of the classification of positive results (precision) is the proportion of positive results that are correctly identified by the classifier among the total number of considered cases.
Recall, also known as sensitivity, reflects the proportion of positive results that are correctly identified by the classifier.
Specificity — reflects the proportion of negative results that are correctly identified by the classifier.
To associate a precision with the fullness F-measure are used. F-measure is a harmonic mean of precision and completeness: = 2 · · + .
One way to evaluate the model is AUC-ROC — square under the curve of error in coordinates True Positive Rate (TPR) and False Positive Rate (FPR): =
+ , =
+ .
The neural network for handwriting recognition based on MNIST, showed an average eficiency of about 99.24%. The average training time of one era is about 160 seconds
Metrics were also calculated for each of the classes:
The following describes the process of downloading the created software package in Docker Hub. The first step is to create an image from the container that contains the program. We use the commit command to commit the changes to the new Docker image.
$ docker commit container_id repository/new_image_name
To show container’s id we use command docker ps –a. As repository name we use user’s Docker Hub login and assign a name to the image. This image is saved locally and to make sure that the new image is saved successfully, we use the command docker images.
Next, we upload our image to Docker Hub. To do this we enter the DockerHub account with command $ docker login -u docker-registry-username where docker-registry-username — Docker Hub user name. We enter the password and upload the image to DockerHub by the command $ docker push docker-registry-username/docker-image-name After successfully uploading the image, we can see it in the account toolbar (Fig. 6).
Thus, we can conclude that Keras library combined with Docker is a powerful toolset for the development software implementation of various machine learning methods. Keras is the most convenient library for writing neural networks, as it is easy to use and has a high speed of creating neural network models. The ease of use of Keras is due to the huge number of pre-installed functions designed to create diferent layers of the neural network. And Docker makes it easier to develop applications, because the installation of all the necessary libraries can be replaced by a single command — download the desired image and in addition, is a universal way to deliver the developed applications on local machines and run them in an isolated environment. 6.
The paper considered Docker software that allows you to create images that contain the program and all the necessary components for its operation.
Created image based on tensorflow containing the Jupyter notebook implemented in Python convolutional neural network using bibltoteki Keras.
This image was uploaded to the public image repository of Docker Hub containers.
The publication has been prepared with the support of the “RUDN University Program 5-100”.