analyse_training_validation_loss.Rmd

---
title: "Training/validation loss"
output: html_notebook
---

```{r packages, include=FALSE, echo=FALSE}
if (!require(pacman)) install.packages("pacman")
pacman::p_load(gridExtra, here, tidyverse)
```

```{r define_doccano_colours}
doccano_colour_scheme_technical <- c(
  'ATTENTION'         = '#f44e3b',
  'EVENT'             = '#0062b1',
  'LOC'               = '#7b64ff',
  'MISC'              = '#a79c4f',
  'ORG'               = '#16a5a5',
  'PER'               = '#fcdc00',
  'Personal Bookmark' = '#fda1ff',
  'POSTCORR'          = '#9f0500',
  'TIME'              = '#73d8ff',
  '?'                 = '#cccccc'
)

doccano_colour_scheme_named <- c(
  'Event'             = '#0062b1',
  'Location'          = '#7b64ff',
  'Miscellaneous'     = '#a79c4f',
  'Organisation'      = '#16a5a5',
  'Person'            = '#fcdc00',
  'Time'              = '#73d8ff'
)
```

```{r, warning=FALSE}
data_file <- here::here("training-losses.xlsx")
labels <- data_file %>% readxl::excel_sheets()

# find y max
for (label in labels) {
  label_max <-
    data_file %>%
    readxl::read_excel(sheet = label, col_types = "numeric") %>%
    select(Epoch, `Training Loss`, `Validation Loss`) %>%
    select(-Epoch) %>%
    max()
  
  if (!exists("y_max")) {
    y_max <- label_max
  } else {
    y_max <- max(y_max, label_max)
  }
}

for (label in labels) {
  data <- 
    data_file %>%
    readxl::read_excel(sheet = label, col_types = "numeric") %>%
    select(Epoch, `Training Loss`, `Validation Loss`)
  
  line_plot <-
    data %>%
    
    ggplot(aes(x = Epoch)) +
    geom_line(aes(y = `Training Loss`, color = "Training Loss"), size = 1) +
    geom_line(aes(y = `Validation Loss`, color = "Validation Loss"), size = 1) +
    scale_color_manual(values = c("Training Loss" = "dodgerblue", "Validation Loss" = "tomato")) +
    labs(title = paste("Training and validation loss for", label, "over epochs"),
         x = "Epoch",
         y = "Loss",
         color = "Legend") +
    ylim(0, y_max) +
    theme_bw()
  
  table_plot <- 
    data %>%
    mutate_at(
      vars(`Training Loss`, `Validation Loss`), 
      ~format(round(., 5), nsmall = 5)
    ) %>%
    tableGrob(rows = NULL)
  
  plot <- grid.arrange(
    line_plot, table_plot,
    ncol = 1, heights = c(2, 1)
  )
  
  print(plot)
  
  ggsave(
    filename = here::here("plots", paste0("loss_", label, ".png")),
    plot = plot,
    create.dir = TRUE
  )
}

```