In jtrecenti/captchaTJSC:

library(magrittr)
devtools::load_all()

for (i in 1:50) {
  Sys.sleep(3)
  try({
    x <- baixar()
    d <- ler(x, salvar = TRUE)
  })
}
baixar() %>% decifrar()

Baixar imagens para treino

for(i in 1:1000) {
  download() %>% 
    classificar(path = 'data-raw/letras')
}

Ajuste do modelo

devtools::load_all(".")
library(caret)
library(dplyr)

d <- 'data-raw/letras' %>%
  dir(full.names = TRUE) %>%
  arrumar()

set.seed(1234567)
aff <- d %>% sample_n(4000)
d_train <- aff %>%
  mutate(y = factor(letra)) %>%
  select(-arq, -group, -letra)

m1 <- train(x = dplyr::select(as.data.frame(d_train), -y), 
            y = d_train$y, method = 'rf')

saveRDS(m1, 'data-raw/m_rf2.rds')
# m2 <- train(x = dplyr::select(as.data.frame(d_train), -y), y = d_train$y, method = 'svmRadial')
# saveRDS(m2, 'data-raw/m_svm.rds')
# m3 <- train(x = dplyr::select(as.data.frame(d_train), -y), y = d_train$y, method = 'gbm')
# saveRDS(m3, 'data-raw/m_gbm.rds')


m1 <- readRDS('data-raw/m_rf.rds')
# m2 <- readRDS('data-raw/m_svm.rds')
# m3 <- readRDS('data-raw/m_gbm.rds')

# d_final <- data.frame(pred1 = predict(m1), pred2 = predict(m2))
# m <- train(x = d_final, y = d_train$y, method = 'gbm', tuneLength = 1)

# acerto <- d %>%
#   dplyr::anti_join(aff, c('arq', 'group')) %>%
#   dplyr::mutate(y = factor(letra)) %>%
#   dplyr::select(-arq, -group, -letra) %>% {
#     dd <- .
#     ddy <- as.data.frame(dplyr::select(dd, -y))
#     pred1 <- predict(m1, newdata = ddy)
#     pred2 <- predict(m2, newdata = ddy)
#     d_pred <- data.frame(pred1, pred2)
#     dplyr::mutate(dd, yest = predict(m, newdata = d_pred))
#   } %>% 
#   with(table(y, yest))

acerto1 <- d %>%
  dplyr::anti_join(aff, c('arq', 'group')) %>%
  dplyr::mutate(y = factor(letra)) %>%
  dplyr::select(-arq, -group, -letra) %>% {
    dd <- .
    ddy <- as.data.frame(dplyr::select(dd, -y))
    dplyr::mutate(dd, yest = predict(m1, newdata = ddy))  
  } %>% 
  with(table(y, yest))

# sum(diag(acerto)) / sum(acerto)
sum(diag(acerto1)) / sum(acerto1)

jtrecenti/captchaTJSC documentation built on May 20, 2019, 3:16 a.m.