label selection script

data-raw/00_select_ml_training_data.R

@@ -0,0 +1,69 @@
+# compile a random dataset
+# balancing the different sheets / handwritten styles
+# the characters represented, the uncertainty of the
+# citizen science transcriptions and empty values
+#
+# ideally a subset of the full dataset (>350K) values
+# is made and restricted to a couple of thousand values
+# to fit within the repo for demo purposes, only to train
+# the full model on a separate data directory later
+
+library(dplyr)
+library(tidyr)
+library(readr)
+
+# read in the majority votes for the Jungle Weather
+# citizen science programme
+df <- readr::read_csv("data-raw/climate_data_majority_vote.csv")
+
+# only retain values with absolute concensus between
+# volunteers, can be relaxed to grow the dataset
+df <- df |>
+  filter(
+    consensus == 1
+  ) |>
+  select(
+    filename,
+    final_value,
+    folder,
+    col
+  )
+
+# how many images are retained
+message(sprintf("number of retained images: %s", nrow(df)))
+
+# too much snot in my head just do a simple
+# random sampling for now with 10k samples
+df <- df |>
+  sample_n(size = 10000) |>
+  mutate(
+    filename = gsub(".png","_dl.png", filename)
+  )
+
+readr::write_csv(df, "data/tmp/labels.csv")
+
+files <- list.files("/data/scratch/zooniverse/format_1_batch_1_dl/","*.png")
+
+# copy all files over to tmp directory
+file.copy(
+  file.path("/data/scratch/zooniverse/format_1_batch_1_dl/",df$filename),
+  "data/tmp/"
+)
+
+# split out individual characters
+# run some statistics on the prevalence of the characters
+# present in the dataset
+char_stats <- df |>
+  rowwise() |>
+  mutate(
+    character = strsplit(as.character(final_value[1]),"")
+  ) |>
+  unnest(cols = "character") |>
+  ungroup() |>
+  group_by(character) |>
+  summarize(
+    n = n()
+  ) |>
+  arrange(
+    n
+  )