Performance Evaluation of Integrated Deep Learning Web Platform for Dataset Training

Along with the complexity of recent web site, many users cannot get the benefits. We developed Integrated Deep Learning Web Platform to help researcher to prepare dataset training for Tensorflow. However, the quality of a web site needs to be assessed. This paper proposes an implementation of WebQual 4.0 for evaluating Integrated Deep Learning Platform for Training Dataset (INDEF) quality. This method used the WebQual model that has some instruments. The instruments grouped the WebQual questions to be three main categories: usability, information and service interaction. From the research conducted can be evaluated that all respondents agreed Integrated Deep Learning Platform for Training Dataset web site met all the WebQual characteristics.


INTRODUCTION
In recent years, there has been a rapid increase in building an object detection model with Tensorflow Object Detection API [1] for many smart applications. Instead of using a model provided by public dataset, many practitioners made their classifier. A model that is built with Tensorflow Object Detection API needed to pass a few steps to reach the best model. It consists of dataset download, image annotation, label map preparation, TF Record Creation, Pipeline Configuration, OMP Parameter Configuration, Training and Inference. Still, the processes are done with a separated different tool.
The first two steps which we called it dataset preparation is done using an image annotation tool where it's desktop version and not portable. And then the rest, training process, the program and configuration should be built by ourselves. INDEF (Integrated Deep Learning Platform for Training Dataset) provides all of those things and makes it be an integrated system. But, with the complexity of the recent web site, the quality of the website becomes the main goal for users who use it to feel the benefits.
WebQual [2] is a method for evaluating the quality of Web sites, developed by Baner & Vidgen (2003). WebQual 4.0 is an extended version from WebQual 3.0 where usability replaced site quality categories. WebQual has been used by many researchers and becomes the most popular method for assessing the quality of web sites (Barner & Vidgen, 2005).
The section in this paper has five purposes: Section 1 present the introduction about background problem, the related works and previous study of the research presents in section 2.
Section 3 offers the methodology of the research was conducted. Section 4 presents the results and discussion of our study. And finally, we closed this paper with Section 5 the conclusion. Section 3 offers the methodology of the research was conducted. Section 4 presents the results and discussion of our study. And finally, we closed this paper with Section 5 the conclusion.

TENSORFLOW MACHINE LEARNING DETECTION MODEL
The study and previous works have developed many models for object detection and machine learning using TensorFlow library. This solution has several steps to do which consist of eight steps: (1) Dataset Download, (2) Image Annotation, (3) Label Map Preparation, (4) TF Record Creation, (5) Pipeline Configuration, (6) OMP Parameter Configuration, (7) Training and (8) Inference. Several applications that used object detection from Tensorflow library and previously they have to create dataset training model in their own computer. The example of Object detection applications are Vehicle as a Mobile Sensor Networks (VaaMSN) [3][4] [5], Smart Environment Monitoring and Analytic in Real-time System (SEMAR) [6][7], Classification [8], and also fruit detection [9].

INTEGRATED DEEP LEARNING WEB PLATFORM FOR DATASET TRAINING
Integrated Deep Learning Web Platform for Dataset Training (INDEF) is a web application that run on Docker Engine [10]. It consists of several parts such as web service run on NginX webserver [11], Golang RESTful API [12] to control and run Tensorflow engine, Python RESTful API [13] to communicate between WebApps and also mongoDB [14]as a database server. The architecture can be seen on Figure 1. The process of web platform can be seen on Figure 2. First Data Labeling, this platform is built to give label to each image on dataset for Tensorflow Training process. The interface of data labeling is by draw a box on an image by click and drag the mouse, then user give a text label on that box. Second are Dataset Training, in this process web interface use RESTful API to run Tensorflow training process by Golang and Python scripts. Then the Tensorflow engine will run in the background process and periodically send a log message to a console and read it by web application interface.

MEASURE WEB QUALITY USING WEBQUAL
The previous study show the implementation of WebQual 4.0 as a method for assessing web site quality from the research conducted by Warjiyono and Corie Mei Hellyana explained that WebQual measurement is based on perception from the last user [15]. Meanwhile, the study published by Stuart J. Barnes and Richard T. Vidgen on an integrative approach to the assessment of e-commerce quality shows that this method can build a profile of qualities of an individual web site through WQ index. The quality of each category can be visualized with a radar chart which contains several categories.

EXPERIMENT
The key to getting a clear view of the quality of web site, qualitative and quantitative approaches were needed. The screenshot from Integrated Deep Learning Platform for Training Dataset can be seen from Figure 4 until Figure 10.

30
Web application ini dapat digunakan untuk mengubah dan menampil informasi personal user. Cara kerjanya adalah web application akan mengirimkan permintaan HTTP Update ke API untuk memperbarui data user di dalam database.
Informasi ini berupa nama proyek yang akan dibuat, gambar ilustrasi yang menggambarkan proyek tersebut, list label yang ingin digunakan, deskripsi dari proyek tersebut dan sebuah tombol add apabila user ingin menambah kontributor untuk membantu user melakukan labeling.   However, WebQual was adopted. The model of WebQual used is about getting the primer data from survey through a questionnaire. This survey can be achieved, whether it's offline or online. This research used the WebQual model, which consists of three instruments; usability, We run our web application on LAN 1Gbps and 802.11ac WLAN, we used server with specification Xeon 48 Core CPU, 96 GB RAM, NVIDIA GTX1060 GPU, and 3TB HDD. On the clients side, users used i5 CPU with 4GB with Chrome Browser web client.
In the Figure 11 shows that the different between labelling tool from labelimg [16] application and labelling from Integrated platform. The conventional way is labelimg, user have to create dataset label in their PC and re-manage all the images and send all the data to Tensorflow server or PC, but with our platform all of images already in the server and user able to directly run Tensorflow engine from web interface.

RESULT AND DISCUSSION
The data that was taken from questionnaire is qualitative data, and for the result, descriptive approached is used. The next step is to examine whether the factors are valid or not in order to measure the construction. In the measurement points above, maybe there is a point that not valid or reliable. Therefore, it should be removed or replaced by other questions. Validation is done by looking at a corrected item-total correlation in the analysis factor. Corrected item-total correlation number was from product-moment correlation between each question that will be test and sum of questions. As shown in Table 2, the average of strongly disagree and disagree are 0.14% and 1.8% it means that very small number that users doesn't agree. And for neutral average is 19.45%. The percentage for agree and strongly agree are 59.45% and 16.25%, it means that more than 75% users are agree. Therefore, the highest score is agreed.
It can be seen from Error! Reference source not found. that all the questions were grouped into three categories. The average of 24 question responses above, the results can be seen in Figure 12.

CONCLUSION
Based on Table 2 and Figure 12 can be concluded that all respondents agreed Integrated Deep Learning Platform for Training Dataset met all WebQual characteristics, which are usability, information quality and service interaction.