koGPT.qmd

---
title: "chatGPT"
subtitle: "koGPT"
author:
  - name: 이광춘
    url: https://www.linkedin.com/in/kwangchunlee/
    affiliation: 한국 R 사용자회
    affiliation-url: https://github.com/bit2r
title-block-banner: true
#title-block-banner: "#562457"
format:
  html:
    css: css/quarto.css
    theme: flatly
    code-fold: true
    toc: true
    toc-depth: 3
    toc-title: 목차
    number-sections: true
    highlight-style: github    
    self-contained: false
filters:
   - lightbox
   - interview-callout.lua
lightbox: auto
link-citations: yes
knitr:
  opts_chunk: 
    message: false
    warning: false
    collapse: true
    comment: "#>" 
    R.options:
      knitr.graphics.auto_pdf: true
editor_options: 
  chunk_output_type: console
---


# R + hugging face

- R 패키지
    - [huggingfaceR](https://github.com/farach/huggingfaceR)
    - [text](https://github.com/OscarKjell/text)
- 블로그
    - [R, Reticulate, and Hugging Face Models](https://cengiz.me/posts/huggingface/)
    - [Hello Transformers from R](https://rpubs.com/eR_ic/transfoRmers)
    
## [huggingfaceR](https://github.com/farach/huggingfaceR) 설치 방법

`reticulate` 최신버전을 설치하고 나서, `miniconda`를 설치한다. 
기존 설치된 경우 `install_miniconda(force = TRUE)` 인자를 넣어 재설치한다.


```{r}
#| eval: false
remotes::install_github("rstudio/reticulate")
reticulate::install_miniconda(force = TRUE)
```

:::{.callout-note}

miniconda 설치에 어려움이 생긴경우 `rminiconda`가 대안이 될 수 있다.

- [`rminiconda`](https://github.com/hafen/rminiconda)

:::


## huggingfaceR 설치

```{r}
#| eval: false
devtools::install_github("farach/huggingfaceR")
```

## 헬로월드

`huggingfaceR` README.md 파일에 실린 헬로월드 텍스트 분류 모형을 돌려보자.

```{r}
library(huggingfaceR)
library(reticulate)

use_python("C:/Users/statkclee/AppData/Local/r-miniconda/envs/huggingfaceR/python.exe")
# hf_python_depends('transformers') # 빠진 라이브러리 설치

distilBERT <- hf_load_pipeline(
  model_id = "distilbert-base-uncased-finetuned-sst-2-english", 
  task = "text-classification")

distilBERT("I like you. I love you")
```



# 모형 활용통계

::: {.panel-tabset}

## 다운로드 횟수

```{r}
library(tidyverse)
library(huggingfaceR)
library(scales)

models <- huggingfaceR::models_with_downloads

models %>%
  ggplot(aes(x= downloads))+
  geom_histogram(alpha = 0.8, color = "black")+
  scale_x_continuous(trans = "log10", labels = comma)+
  theme_bw(base_family = "sans")+
  labs(title = "Log-transformed Histogram of Downloads per Model",
       subtitle = "The vast majority of models have received < 1k downloads",
       y = "Number of Models",
       x = "Number of Downloads",
       caption = paste0("Up to date as of: ", Sys.Date()))
```

## 작업분류별

```{r}
models %>%
  count(task, sort = TRUE)%>%
  mutate(task = stringr::str_to_title(task),
         task = stringr::str_replace_all(task, "-", " "))%>%
  na.omit()%>%
  ggplot(aes(y= reorder(task,n), x = n))+
  geom_col(fill = "#0f50d2")+
  theme_bw(base_family = "sans")+
  labs(title = "Models Saved on Hugging Face Hub by Task",
       subtitle = "Models without download information & NA's for task omitted",
       caption = paste0("Up to date as of: ", Sys.Date()),
       y = "Model task",
       x = "Number of models per task")
```

## 모형갯수와 다운로드 평균횟수

```{r}
models %>%
  mutate(task = forcats::fct_lump_n(task, n = 16))%>%
  na.omit()%>%
  group_by(task)%>%
  summarise(av_dl = mean(downloads), med_dl = median(downloads), n = n())%>%
  arrange(desc(av_dl))%>%
  ggplot(aes(y= av_dl, x = n))+
  geom_point()+
  geom_text(aes(label = stringr::str_to_title(stringr::str_replace_all(task, "-", " "))),
            check_overlap = TRUE, size =3, nudge_y = 350)+
  theme_bw()+
  labs(title = "Average Downloads per Task vs Number of Models")+
  labs(y = "Average Downloads", x = "Number of Models",
       caption = paste0("Up to date as of: ", Sys.Date()))+
  expand_limits(x = c(-10, 8500))
```

## 모형갯수와 다운로드 중위수

```{r}
models %>%
  mutate(task = forcats::fct_lump_n(task, n = 16))%>%
  na.omit()%>%
  group_by(task)%>%
  summarise(av_dl = mean(downloads), med_dl = median(downloads), n = n())%>%
  arrange(desc(av_dl))%>%
  ggplot(aes(y= med_dl, x = n))+
  geom_point()+
  geom_text(aes(label = stringr::str_to_title(stringr::str_replace_all(task, "-", " "))),
            check_overlap = FALSE, size = 3)+
  theme_bw()+
  labs(title = "Median Downloads per Task vs Number of Models")+
  labs(y = "Median Downloads", x = "Number of Models",
       caption = paste0("Up to date as of: ", Sys.Date()),
       subtitle = "Median is far below mean - models are a winner-takes-most game")+
  expand_limits(x = c(-100, 8500))
```

:::

## 인기 모형

다운로드 횟수가 많은 `hugginface` 모형은 다음과 같다.


```{r}
library(reactable)
models %>% 
  select(-private, -sha) %>% 
  reactable::reactable(
    searchable = TRUE, minRows = 10,
    columns = list(downloads = colDef(format = colFormat(separators  = TRUE)),
                   model = colDef(align = "center"),
                   task = colDef(align = "center")))
```


# 데이터셋

`huggingfaceR` 패키지 `hf_load_dataset()` 함수를 사용해서 [`emotion`](https://huggingface.co/datasets/emotion) 감성 데이터셋을 가져와서 
`tidvyerse` 작업흐름과 연계시킬 수 있다.


```{r}
#| eval: false

emo <- hf_load_dataset(
  dataset = "emotion", 
  split = "train"
  )

emo_model <- emo %>%
  sample_n(100) %>% 
  transmute(
    text,
    emotion_id = label,
    emotion_name = label_name,
    distilBERT_sent = distilBERT(text)
  ) %>%
  unnest_wider(distilBERT_sent)

glimpse(emo_model)
```