In maju116/platypus: Tools for Computer Vision in R

Download images and annotations: BCCD dataset.

Split dataset into train, validation and test set

library(tidyverse)
library(platypus)
library(abind)
library(here)

BCCD_path <- here("development/BCCD/")
annot_path <- file.path(BCCD_path, "Annotations/")
images_path <- file.path(BCCD_path, "JPEGImages/")

c("train", "valid", "test") %>%
  walk(~ {
    dir.create(file.path(BCCD_path, .))
    dir.create(file.path(BCCD_path, ., "Annotations/"))
    dir.create(file.path(BCCD_path, ., "JPEGImages/"))
  })

annot_paths <- list.files(annot_path, full.names = TRUE)
images_paths <- list.files(images_path, full.names = TRUE)
n_samples <- length(annot_paths)
set.seed(111)
train_ids <- sample(1:n_samples, round(0.8 * n_samples))
valid_ids <- sample(setdiff(1:n_samples, train_ids), round(0.19 * n_samples))
test_ids <- setdiff(1:n_samples, c(train_ids, valid_ids))

walk2(c("train", "valid", "test"), list(train_ids, valid_ids, test_ids), ~ {
  annots <- annot_paths[.y]
  images <- images_paths[.y]
  dir_name <- .x
  annots %>% walk(~ file.copy(., gsub("BCCD", paste0("BCCD/", dir_name), .)))
  images %>% walk(~ file.copy(., gsub("BCCD", paste0("BCCD/", dir_name), .)))
})

Generate custom anchor boxes:

blood_labels <- c("Platelets", "RBC", "WBC")
n_class <- length(blood_labels)
net_h <- 416 # Must be divisible by 32
net_w <- 416 # Must be divisible by 32
anchors_per_grid <- 3

blood_anchors <- generate_anchors(
  anchors_per_grid = anchors_per_grid, # Number of anchors (per one grid) to generate
  annot_path = annot_path, # Annotations directory
  labels = blood_labels, # Class labels
  n_iter = 10, # Number of k-means++ iterations
  annot_format = "pascal_voc", # Annotations format
  seed = 55, # Random seed
  centroid_fun = mean # Centroid function
)

blood_anchors

Build YOLOv3 model (you can load YOLOv3 Darknet weights trained on COCO dataset. Download pre-trained weights from here):

blood_yolo <- yolo3(
  net_h = net_h, # Input image height
  net_w = net_w, # Input image width
  grayscale = FALSE, # Should images be loaded as grayscale or RGB
  n_class = n_class, # Number of object classes (80 for COCO dataset)
  anchors = blood_anchors # Anchor boxes
)
blood_yolo %>% load_darknet_weights(here("development/yolov3.weights")) # Optional

blood_yolo

Compile the model with correct loss and metrics:

blood_yolo %>% compile(
  optimizer = optimizer_adam(lr = 1e-5),
  loss = yolo3_loss(blood_anchors, n_class = n_class),
  metrics = yolo3_metrics(blood_anchors, n_class = n_class)
)

Create data generators:

train_blood_yolo_generator <- yolo3_generator(
  annot_path = file.path(BCCD_path, "train", "Annotations/"),
  images_path = file.path(BCCD_path, "train", "JPEGImages/"),
  net_h = net_h,
  net_w = net_w,
  batch_size = 16,
  shuffle = FALSE,
  labels = blood_labels
)

valid_blood_yolo_generator <- yolo3_generator(
  annot_path = file.path(BCCD_path, "valid", "Annotations/"),
  images_path = file.path(BCCD_path, "valid", "JPEGImages/"),
  net_h = net_h,
  net_w = net_w,
  batch_size = 16,
  shuffle = FALSE,
  labels = blood_labels
)

Fit the model:

blood_yolo %>%
  fit_generator(
    generator = blood_yolo_generator,
    epochs = 1000,
    steps_per_epoch = 19,
    validation_data = valid_blood_yolo_generator,
    validation_steps = 5,
    callbacks = list(callback_model_checkpoint("development/BCCD/blood_w.hdf5",
                                               save_best_only = TRUE,
                                               save_weights_only = TRUE)
    )
  )

Predict on new images:

blood_yolo <- yolo3(
  net_h = net_h,
  net_w = net_w,
  grayscale = FALSE,
  n_class = n_class,
  anchors = blood_anchors
)
blood_yolo %>% load_model_weights_hdf5(here("development/BCCD/blood_w.hdf5"))

test_blood_yolo_generator <- yolo3_generator(
  annot_path = file.path(BCCD_path, "test", "Annotations/"),
  images_path = file.path(BCCD_path, "test", "JPEGImages/"),
  net_h = net_h,
  net_w = net_w,
  batch_size = 4,
  shuffle = FALSE,
  labels = blood_labels
)

test_preds <- predict_generator(blood_yolo, test_blood_yolo_generator, 1)

test_boxes <- get_boxes(test_preds, blood_anchors, blood_labels,
                        obj_threshold = 0.6)

test_boxes

Plot / save images with predicted bounding boxes:

plot_boxes(
  images_paths = list.files(file.path(BCCD_path, "test", "JPEGImages/"), full.names = TRUE),
  boxes = test_boxes,
  labels = blood_labels,
  save_dir = BCCD_path)