chatGPT

1 R + hugging face

• R 패키지
  • huggingfaceR
  • text
• 블로그
  • R, Reticulate, and Hugging Face Models
  • Hello Transformers from R

1.1 huggingfaceR 설치 방법

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

코드

remotes::install_github("rstudio/reticulate")
reticulate::install_miniconda(force = TRUE)

노트

miniconda 설치에 어려움이 생긴경우 rminiconda가 대안이 될 수 있다.

• rminiconda

1.2 huggingfaceR 설치

코드

devtools::install_github("farach/huggingfaceR")

1.3 헬로월드

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

코드

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")
#> [[1]]
#> [[1]]$label
#> [1] "POSITIVE"
#> 
#> [[1]]$score
#> [1] 0.9998739

2 모형 활용통계

다운로드 횟수

코드

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

코드

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

코드

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

코드

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))

2.1 인기 모형

다운로드 횟수가 많은 hugginface 모형은 다음과 같다.

코드

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")))

3 데이터셋

huggingfaceR 패키지 hf_load_dataset() 함수를 사용해서 emotion 감성 데이터셋을 가져와서 tidvyerse 작업흐름과 연계시킬 수 있다.

코드

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)