This is Techmeal TV. Today, we will talk about the two most commonly used deep learning frameworks, PyTorch and TensorFlow. The comparison between the two frameworks has been ongoing since their early days. A few days ago, Ryan O'Conner, a deep learning developer from the United States, shared an article. Through this article, we will summarize the strengths and weaknesses of PyTorch and TF in each category, looking ahead to 2022. Without further ado, let's get right into it. I divided the comparison into three main categories. Important terms will be used in English to avoid misinterpretation, but I will explain their meanings in Korean. The first category is Model Availability. In Korean, it can be translated as '모델 가용성.' As the field of deep learning grows rapidly every year, the most cutting-edge deep learning models, known as State-of-the-Art or SOTA, are emerging.

Approaching the overflowing SOTA models, in terms of convenience for immediate use, which framework shows higher scores between PyTorch and TF? To cut to the chase, PyTorch takes the lead, showing a significant gap. When looking at the platform HuggingFace, which gathers various state-of-the-art deep learning models, many models are based on PyTorch. The red-highlighted section represents the number of models implemented only in PyTorch. On the other hand, models implemented solely in TF are shown in the dark yellow in the middle. The far right indicates the total number of models implemented in both PyTorch and TF. Clearly, PyTorch dominates at a glance. Focusing on the Top-30 models most used within HuggingFace, TF falls drastically behind. Among the 30 models, there is not a single model implemented exclusively in TF.

This time, we will compare the number of research papers. In fact, when dividing the usage areas of PyTorch and TF into academia and business practical fields, the majority in academia still uses TF. So, I thought it would be quite different at least. However, as shown in the graph now, PyTorch has brought a significant increase in research paper share. The author of this article, Ryan O'Conner, claimed that it was due to the difficulties of TF 1, in other words, its complexity or obscurity. At that time, PyTorch stood out as an alternative for researchers. Even though TF 2 appeared in 2019 to solve various issues of the previous version, the trend had already shifted towards PyTorch. This graph, based on 2018, shows which framework researchers who used TF or PyTorch in the previous year switched to in 2019 to conduct their research.

Due to being a relative ratio, it cannot be compared in absolute numbers, but more than half of the existing TF users have switched to PyTorch. On the other hand, among the existing users of PyTorch, 85% have continued to use PyTorch. When looking at the total, the increase in PyTorch usage and the decrease in TF usage are evident. Even in 2019 when TF 2 was introduced, this trend continued. The second comparison point is Deployment Infrastructure. This refers to how well the deployment environment is set up. It can be understood as comparing the environment construction that can be loaded into services for microservices or actual business models. The winner in this aspect is TF. This was made possible by powerful tools like TF Lite and TF Extended. Through TFX and Serving, it is relatively easy to deploy models end-to-end on servers or in the cloud. This advantage allows for optimized deployment in various environments and ecosystems such as Google Cloud, Kubernetes, Docker, and others.

TF Lite is optimized for distribution, especially for mobile devices and IoT devices, covering not only Android and iOS but also microcontrollers and Linux, capturing both performance and scalability. Just because there is PyTorch, it doesn't mean such a tool doesn't exist. In 2020, AWS and former Facebook, now Meta, collaborated to create Torch Serve, which supports both REST API and gRPC API, but there are still some unstable parts compared to TF Serving. PyTorch Mobile, introduced in 2019, and PyTorch Live, introduced in early December, are also used for mobile deployment, but as they have been around for a short time, their usefulness needs to be further assessed. The last comparison point is Ecosystems. This is easy to translate. It is a part that compares how well the framework ecosystem is structured. As the overall deep learning ecosystem is developing widely and deeply, it is becoming important to see how much influence one framework holds in line with those changes.

The winner in this Daejeon is also TF. Google is building a deep learning workflow that works end-to-end from the very beginning to the end by investing tremendous capital, based on TF. They upload well-tuned and trained models on TF Hub, and through Extended, they can perform processes such as model loading, validation, analysis, transfer learning, and model deployment in one place. In 2020, with the introduction of Google Coral, building AI products became easier. Not only that, useful tools such as TF JS, Google Cloud, Colab, Vertex AI are very well equipped and continuously updated. Of course, in the PyTorch ecosystem, there are PyTorch Hub, Torch Elastic, Torch X SDK, PyTorch Lightning, but it has not yet reached the level of the TF ecosystem. It seems that the sudden increase in the number of existing PyTorch users was not adequately addressed by the ecosystem, which can be seen as an intermediate process that the ecosystem could not respond to.

Therefore, as the number of users increases, many tools will emerge in the future, making the ecosystem more robust. However, at the current moment looking ahead to 2022, it can be summarized that TF has a superior ecosystem. So, summarizing in the Model Availability section, PyTorch wins; in the Deployment Infrastructure section, TF wins; and in the Ecosystems section, TF wins. In total, TF is leading 2:1. Now, there is one question that arises: which framework should I use? Even though TF is leading, the number of PyTorch users is increasing, which is a point of concern. Therefore, in the next video of this series, I will organize which deep learning framework to choose, whether you are in business, academia, looking to change your career, or just starting as a hobby. If you subscribe and turn on notifications, you will be able to quickly access the next video.

Please hold on a moment, then we will continue in the next video.

2022년 딥러닝 프레임워크 고르는 방법

매일 아침 접하는 신선한 해외 테크 뉴스 & 트렌드

2022년 딥러닝 프레임워크 고르는 방법

2022년 딥러닝 프레임워크 고르는 방법

스크립트