Hello, in this video, I would like to introduce the Hugging Face tutorial. When we share code now, we tend to use models that have been trained like companies do. Hugging Face is widely used when sharing these trained models. The Korean tutorial for Hugging Face is very well done, so I would like to introduce that tutorial. First of all, when you go to the Hugging Face model, you will see that other people have used GPUs for training a lot, consuming a lot of electricity and time. You can download the trained models and retrain them, or use them to develop services. So, with these models and datasets available, if you check the documentation, you will find resources on how to utilize Hugging Face. Look below to find the run.

When you go to the Learn tab in the community, I would like to introduce a Korean tutorial among the NLP courses. If you go to the left side here, it's currently in English. If you select KO here, you can choose the Korean tutorial to watch. So, not all tutorials are available in Korean. If you look here now, there are tutorials numbered 0, 1, 2, 5, and 8 like this. There are tutorials that are not translated in the middle, you should switch back to the English tutorial to watch them. In the English tutorial now, there are tutorials up to number 9. Let's take a look at the Korean tutorial together. First of all, in the chapter, it says Welcome to the students in the lecture. In this lecture, we will focus on the Hugging Face environment's library, Transformer Dataset.

Tokenizer Accelerate AWA now, I will learn about natural language processing using the Hugging Face Hub. So, there is content now that is free lectures here, and there is a tutorial. The tutorial is really well done, so I often recommend it to those who are using Hugging Face for the first time, so I decided to make a video about it. So, not all chapters are translated here, but in chapters 1 to 4, the core concepts of the Transformer library are introduced, then you can understand the operation principle of the Transformer, use the model for fine-tuning with the dataset, and learn how to share the model with the Hub. Then, in chapters 5 to 8, before performing classic NLP tasks, dataset tokenizer.

When you learn about the basics, you will be able to solve general NLP problems on your own. Chapters 9 to 12 then explore how Transformer models are utilized not only for NLP problems but also for speech, computer vision. In this process, you will learn how to build and share model demos and optimize for production environments. If you don't have knowledge of Python, it is recommended to take basic courses in Python or in programs like FastAI, deep learning AI. There are also recommended links for further learning. Pre-knowledge of PyTorch or TensorFlow is not necessary but it is helpful if you are familiar with them. Let's quickly get started.

Now I'm introducing what the content is about. Natural language processing now refers to the field of machine learning that focuses on understanding everything related to human language. It includes sentence classification, word classification within sentences, text content generation, extracting answers from text, and generating new sentences from input text. In this tutorial, you can directly try these tasks with pre-trained models. The difficulty of this is also explained here. Computers and humans process information differently. When we read the sentence I'm hungry, we can understand its meaning immediately. Similarly, when given a pair of sentences like I'm hungry and I'm sad, we can easily determine how similar the two sentences are. However, machine learning models find it not as easy as humans do.

It is difficult because there are contents like this, and then there is the paradox of language that represents the ambiguous meaning, making natural language processing difficult. With the advent of Transformers, natural language processing has made significant advancements. So, what can we do with Transformers? We will have a practical exercise from here. You can directly open it in Colab. In Colab, although translations are not available, let's see how to use it effectively together. First, I will open it by saving a copy in the drive. Because when you save a copy, it will be saved in your Google Drive. I will generate it as a copy and show you. And first, with the Transformer.

Now, let's take a look at what has been introduced here. You can run it directly on Colab. Transformers are everywhere, so they are used to solve natural language processing problems. Many companies and organizations contribute to the community by sharing models, like Hugging Face and Transformer models as shown here. In the case of Hugging Face, even domestic companies like Kakao have uploaded their models, and companies like SK Telecom have also shared models trained using their resources and equipment. So, I will show you how we can find and use such models. Therefore, the Transformer library provides functions to use and build shared models.

When it comes to models like GPT-3, the cost of training is almost like, when you see it in papers, it's like, Oh, it cost around 5 billion, you know. So, when you look at those papers, they are now also considering issues related to carbon emissions in the future. To train a language model, it really costs a lot of money and time. For us individuals, it's difficult to train them. So, if pre-trained models are shared on platforms like Hugging Face, we can use those models, add additional datasets to train them, and use them. Then, there is a way to work with Hugging Face's pipeline. Here, on Hugging Face, there is a way to work with the pipeline, tutorials on working with this pipeline.

Let's try it together. There is a tutorial in the video now. So, if you watch the video, you can follow along. Let's try it together. First, load the transformer library and import the pipeline. By doing this, you can perform various tasks here. The simplest one here is sentiment analysis. By doing sentiment analysis, you can see if it is positive or negative, and then you can see what the score is, how much the probability is. So, for example, for a sentence in the classifier, this can be considered positive with a probability of 95%. Furthermore, you can input not only one sentence but also multiple sentences. So, the first sentence, the second one.

In each sentence, there are positive and negative, and then you can also see the probability values for these positive and negative. If you look to the right, there are tasks like zero-shot classification, text generation, inference, mask filling, named entity recognition, question answering, summarization, and translation. We will try these tasks together in Colab. First, let's establish the connection and then proceed with the necessary installations. So, we will install datasets, then huggingface transformers and sentencepiece. Next, we will import the transformer pipeline like this. Once the installation above is completed, we will try the tutorial in the example. We will bring sentiment analysis through the pipeline, put it into the pipeline, and then use the classifier of Hugging Face.

I have been eagerly waiting for this course in my life. When you see here, you can now insert not only one sentence but also several sentences together, as shown in this tutorial. So, if the installation is now complete, next, we will try sentiment analysis. If you look here now, you are downloading. So, after downloading, if you go to the link and open it, here it says DistilBERT-based SST-2 ReC. So, we bring a pre-trained model like this and try some sentiment analysis here. On the right side, it says I like you, I love, which is positive. Then, I will try the opposite emotion like I hate you and compute it.

Let's take a look at whether negatives are coming out. So, now you can try running this model on the site like this. When negatives came out as 999%, let's change it back to something positive like I like. When we compute it again, positive will come out as 100%. And here, it says Text Classification and it's written in English. Next, based on the DistilBERT model, the term Distill means to evaporate or distill, right? So, you can see that it's like they made the BERT model a bit lighter. Furthermore, you can also see additional information about the model used in papers written here. Then, when using this model, although we are currently using a pipeline here, you can see additional information about the model and when we use this model directly.

You can now copy and use the source code here as it is to further train the pre-trended model or to use it. Next, you can see that the evaluation values and result values are shared together. So, in this DistilBERT model based model, which is in English, when you try to use it in Korean, the performance may not be very good. For example, if you input I love you, a positive value of 0.825 should come out, which means it's positive with 82%. It might be a bit difficult if it comes out positive with 0.851; it should show a bit of negativity. For example, if you want to write a positive message but want to check it, if it gives 0.627, it means it's negative. Let's test it with a positive one again.

Let's try typing I am happy. Well, now 0.86, and now a positive one came out. In the case of Korean, the performance is now declining because this model is trained based on English. So, if you see it as such, when we do sentiment analysis here, in terms of tasks, the model is already loaded. Let me briefly look at the document here where we loaded the pipeline. When you look at this document, you can specify tasks or models and then use them. Next, when you look at this document, you can see that all the tasks are provided here. So, here, we will try using various tasks and running them. First, let's look at a few simple examples together. So, let's see how these models are on the site.

I have seen ClashFire in action, but I will try it again. When you enter ClashFire, it will tell you whether each sentence is positive or negative. Let's try entering happy first, then joyful, and difficult. Now, two positive and one negative. I will now add additional inputs. Then, positive, positive, positive, even though difficult is negative, it showed 0.76% positive. This is because the model used here is trained in English. So, below, let's also learn how to load a model trained in Korean and use it. Next, let's classify it with Zeroes. Here, you can see the label values such as education, politics, business, and various other label values.

Once the candidates are designated, they will now return the probability values of each label. So here, education came out with the highest probability of 0.84, so let's take a look at the document. When you look at the document, you can designate various label values such as agent, tablet computer, etc. So you can see to what extent they correspond to the probabilities of each label. If you change the view, for example, you can designate various things such as mobile website billing, account access, and more. So, you can also see things like cooking, dancing, and more. I will show you more about this below. So, when you look, it's day I.

I will see the sequence now. I put this sentence into a variable. Next, I put in the candidate labels, travel, cooking, and dancing. Here, travel came out with a probability of 99, the highest, followed by dancing and cooking. If you look here, when you set it to multi-label 2, then the highest probability does not come out as one, but you can see the probability of each. For example, travel is 99, exploration is 93, 0.93, like this, each probability value is output. Depending on the multi-label true/false, it will return accordingly. Then, here, how to use it with PyTorch. Which label belongs to which label. So, here, we call the pre-trained model like this.

You can now use that. So, let's take a look at this below. Next, you can also do text generation. By using GPT-2 from Hugging Face, you can now use it. Also, by clicking this link, you can check about the pre-trained GPT-2 model. Here, through the hosted Inference API, here is my name, it's a tutorial made by a person. So, when asked what my favorite thing is, I execute the next sentence like this. So, in the case of GPT-2, when you copy the model from here, if you look a bit more here, in the example, after generating text, it will show you How to and then something like an element, CSS selector, style, etc. This also gives different values each time you run it because when we ask a question to GPT, the answers are always different.

You know, the answers are now coming out a little differently each time, but if you seed here, you can fix it. Then you can specify a model, like the Hello Language model, then the Max Length, and then the Number of Return Sequences, and specify how many return sequences you want. So, when you run it, you will see one, two, three, four, five sentences generated. But here, when you specify the model, if you do it in Korean, the performance is not very good. So, let me try changing it to Korean once and run it. So, since it's trained in English, when I change it to Korean and run it, the results are a bit messed up here.

When we bring in models, the performance is not as good as expected, but I will show you how to bring in and use Korean now.

HuggingFace 처음이라면? 🤗 허깅페이스로 할 수 있는 일?분류, 번역, 요약, 개체명인식, 마스크채우기.. 그리고 또 뭐가 있더라? 트랜스포머 모델 활용법

🐍 Microsoft MVP(Python Developer Technologies)
🌱 인프런 지식공유자 : https://www.inflearn.com/users/91177/@todaycode
공공데이터 분석, 데이터 시각화, 캐글과 데이콘 머신러닝/딥러닝 튜토리얼
Pandas, Numpy, Scipy, scikit-learn, TensorFlow, Keras, Jupyter, Colaboratory, HuggingFace

📚＜모두의 한국어 텍스트 분석 with 파이썬 저자＞
교보문고 https://bit.ly/3MQ3cNi
Yes24 https://bit.ly/3MrUdk0
알라딘 https://bit.ly/3OsPTDQ
인터파크 https://bit.ly/3Ou97Je
영풍문고 https://bit.ly/3pTNHLd


👉 페이스북 : https://www.fb.com/todaycode​
👉 인스타그램 : https://www.instagram.com/_todaycode/​

* 아래 URL과 쿠폰코드로 할인금액으로 수강하실 수 있습니다.
🌱 인프런 📈 증권 데이터 수집과 분석으로 신호와 소음 찾기 🔎 전체 보기 👉 http://bit.ly/inflearn-finace-data ( 쿠폰코드: 3083-a12fb4ddd2d4 ) 
🌱 인프런 📊 공공데이터로 파이썬 데이터 분석 시작하기 💻 전체 보기 👉 https://bit.ly/inflearn-public-data-eda (쿠폰코드 : 3084-4e27fa1575a4 )
🌱 인프런 📝 단 두 장의 문서로 데이터 분석과 시각화 뽀개기⚡️ 전체 보기 👉 https://bit.ly/inflearn-pandas-tutorial  (쿠폰코드 : 3085-6040d593c7e9 )
🌱 인프런 📚 영화리뷰 감정 분석을 통한 파이썬 텍스트 분석과 자연어 처리 👉 https://bit.ly/inflearn-nlp-tutorial
