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


오늘코드todaycode

이 영상은 허깅페이스 튜토리얼을 소개하며, 사전 훈련된 모델을 활용하여 다양한 자연어 처리 작업을 수행하는 방법을 안내합니다.

허깅페이스 허브에서 모델을 다운로드하여 재학습하거나, 감성 분석, 텍스트 생성, 번역 등의 작업을 파이프라인을 통해 쉽게 수행할 수 있음을 보여줍니다.

특히 한국어 튜토리얼을 통해 트랜스포머 라이브러리의 핵심 개념을 배우고, 모델을 공유하는 방법까지 학습할 수 있음을 강조합니다.

Hugging Face 튜토리얼 소개 및 모델 공유의 중요성

Hugging Face는 학습된 모델 공유에 널리 사용되며, GPU를 사용한 학습 모델을 다운로드하여 재학습하거나 서비스 개발에 활용 가능.

Hugging Face 한국어 튜토리얼 소개

Hugging Face Learn 탭의 NLP 강좌 중 한국어 튜토리얼을 소개하며, 일부 튜토리얼은 영어로 제공됨.

Transformer 라이브러리의 핵심 개념, 모델 fine-tuning, 모델 공유 방법 등을 학습.

Transformer 모델이 NLP 뿐 아니라 음성, 컴퓨터 비전에도 활용되는 방법, 모델 데모 구축 및 공유, 프로덕션 환경 최적화 방법 등을 학습.

자연어 처리(NLP) 및 Transformer의 역할

NLP는 인간 언어 이해에 초점을 맞춘 머신러닝 분야이며, Transformer 등장으로 상당한 발전이 있었음.

Colab을 활용한 Transformer 실습 준비

Colab에서 Transformer를 직접 실행하고, Hugging Face와 Transformer 모델을 활용하는 방법 설명.

Hugging Face는 다양한 기업과 기관이 모델을 공유하는 플랫폼이며, Transformer 라이브러리는 공유 모델을 사용하고 구축하는 기능 제공.

사전 학습 모델의 중요성 및 Hugging Face 파이프라인 활용

개인이 학습하기 어려운 사전 학습 모델을 Hugging Face에서 활용하고, 추가 데이터셋으로 학습시키는 방법 소개.

Hugging Face 파이프라인을 이용한 감성 분석

파이프라인을 통해 감성 분석을 수행하고, 문장의 긍정/부정 여부와 확률을 확인하는 방법 설명.

다양한 NLP Task (zero-shot classification, text generation, inference, mask filling, named entity recognition, question answering, summarization, translation) 소개.

DistilBERT 모델을 이용한 감성 분석 실습

DistilBERT 기반의 사전 학습 모델을 사용하여 감성 분석을 수행하고, 긍정/부정 결과 확인.

모델에 대한 추가 정보 확인 및 소스 코드 활용 방법 안내.

한국어 감성 분석 시 성능 저하 및 개선 방안

영어 기반 모델을 한국어에 적용 시 성능이 저하될 수 있으며, 한국어로 학습된 모델을 사용하는 것이 좋음.

Hugging Face에서 제공하는 다양한 Task와 모델을 활용하는 방법 소개.

Zero-shot classification을 이용한 분류

Zero-shot classification을 사용하여 문장을 다양한 label 값으로 분류하고 확률 값을 확인하는 방법 설명.

Hugging Face의 GPT-2 모델을 사용하여 텍스트를 생성하고, 모델의 다양한 파라미터를 설정하는 방법 설명.

GPT 모델의 답변이 매번 달라지는 이유와 seed를 사용하여 결과를 고정하는 방법 설명.

한국어 텍스트 생성 시 문제점 및 해결 방안

영어 기반 GPT-2 모델을 한국어에 적용 시 성능이 저하될 수 있으며, 한국어로 학습된 모델을 사용하는 것이 좋음.

한국어 모델을 가져와 사용하는 방법 안내 예정.

이 비디오에서는 Hugging Face 튜토리얼을 소개합니다. Hugging Face는 **기업처럼 훈련된 모델을 공유하는 데 널리 사용**되며, 한국어 튜토리얼이 잘 되어 있어 이를 소개하고자 합니다. Hugging Face 모델을 통해 다른 사용자들이 GPU를 활용하여 많은 시간과 전력을 소비하며 훈련시킨 모델을 다운로드하여 재훈련하거나, 서비스를 개발하는 데 활용할 수 있습니다.

Hugging Face Hub에는 모델과 데이터셋이 공유되어 있으며, **문서를 통해 Hugging Face 활용법을 익힐 수 있습니다**.

📚 Hugging Face 한국어 튜토리얼 안내

Hugging Face 커뮤니티의 Learn 탭에서 NLP 강좌 중 한국어 튜토리얼을 선택할 수 있습니다. 하지만 **모든 튜토리얼이 한국어로 번역되어 있지는 않으므로**, 번역되지 않은 부분은 영어 튜토리얼로 전환해야 합니다. 튜토리얼은 Transformer 라이브러리, Dataset, Tokenizer, Accelerate 등을 다루며, Hugging Face Hub를 사용한 자연어 처리 학습을 목표로 합니다.

이 튜토리얼은 Hugging Face를 처음 사용하는 사람들에게 자주 추천되며, **Transformer 라이브러리의 핵심 개념을 소개하고 모델을 fine-tuning하는 방법과 Hub에 공유하는 방법**을 배울 수 있습니다. 또한, Transformer 모델이 NLP 문제뿐만 아니라 음성, 컴퓨터 비전에도 활용되는 방법을 탐구합니다.

🤖 Transformer 모델과 자연어 처리

자연어 처리(NLP)는 인간 언어 이해에 초점을 맞춘 머신 러닝 분야로, 문장 분류, 단어 분류, 텍스트 생성, 답변 추출, 새로운 문장 생성 등을 포함합니다. 이 튜토리얼에서는 **미리 훈련된 모델을 사용하여 이러한 작업을 직접 시도**해 볼 수 있습니다.

Transformer 모델의 등장으로 자연어 처리는 크게 발전했습니다. Hugging Face와 같은 플랫폼에서 **사전 훈련된 모델을 공유함으로써 개인도 이러한 모델을 활용**하고 추가 데이터셋으로 훈련하여 사용할 수 있게 되었습니다.

**GPT-3와 같은 모델은 훈련 비용이 매우 높기 때문에** Hugging Face와 같은 플랫폼에서 공유되는 사전 훈련된 모델을 활용하는 것이 효율적입니다.

Hugging Face의 pipeline을 사용하면 감성 분석, zero-shot 분류, 텍스트 생성, 마스크 채우기, 개체명 인식, 질문 답변, 요약, 번역 등 다양한 작업을 수행할 수 있습니다. **pipeline을 통해 문장의 긍정 또는 부정을 분석하고 확률 값을 확인**할 수 있습니다.

Colab에서 datasets, huggingface transformers, sentencepiece를 설치한 후, transformer pipeline을 import하여 Hugging Face의 classifier를 사용할 수 있습니다. **DistilBERT 기반의 사전 훈련된 모델을 가져와 감성 분석을 수행**할 수 있습니다.

예를 들어, "I love you"라는 문장을 입력하면 긍정적인 결과가 나오지만, **모델이 영어로 훈련되었기 때문에 한국어 문장에서는 성능이 저하될 수 있습니다**.

🔎 다양한 Task 실행 및 모델 탐색

Hugging Face 문서를 통해 다양한 task와 모델을 지정하여 사용할 수 있습니다. ClashFire를 사용하여 문장의 긍정/부정을 판단할 수 있지만, **영어 기반 모델이므로 한국어 문장에서는 정확도가 떨어질 수 있습니다**.

Zero-shot classification을 통해 교육, 정치, 비즈니스 등 다양한 label 값의 확률을 확인할 수 있습니다. 텍스트 생성 기능은 GPT-2 모델을 사용하여 텍스트를 생성하며, **매번 다른 답변이 나올 수 있지만 seed 값을 통해 결과를 고정**할 수 있습니다.

모델을 지정하고, 생성할 텍스트의 최대 길이, 반환할 시퀀스 수를 지정할 수 있습니다. 하지만 **한국어로 훈련된 모델이 아니면 성능이 좋지 않을 수 있습니다**.

한국어 모델을 가져와 사용하는 방법을 알아볼 수 있습니다.