RStudio AI Weblog: Prepare in R, run on Android: Picture segmentation with torch


In a way, picture segmentation isn’t that completely different from picture classification. It’s simply that as a substitute of categorizing a picture as a complete, segmentation leads to a label for each single pixel. And as in picture classification, the classes of curiosity depend upon the duty: Foreground versus background, say; various kinds of tissue; various kinds of vegetation; et cetera.

The current publish isn’t the primary on this weblog to deal with that matter; and like all prior ones, it makes use of a U-Web structure to attain its aim. Central traits (of this publish, not U-Web) are:

  1. It demonstrates the way to carry out knowledge augmentation for a picture segmentation job.

  2. It makes use of luz, torch’s high-level interface, to coach the mannequin.

  3. It JIT-traces the skilled mannequin and saves it for deployment on cell gadgets. (JIT being the acronym generally used for the torch just-in-time compiler.)

  4. It contains proof-of-concept code (although not a dialogue) of the saved mannequin being run on Android.

And in the event you assume that this in itself isn’t thrilling sufficient – our job right here is to search out cats and canines. What may very well be extra useful than a cell software ensuring you’ll be able to distinguish your cat from the fluffy couch she’s reposing on?

A cat from the Oxford Pet Dataset (Parkhi et al. (2012)).

Prepare in R

We begin by making ready the info.

Pre-processing and knowledge augmentation

As supplied by torchdatasets, the Oxford Pet Dataset comes with three variants of goal knowledge to select from: the general class (cat or canine), the person breed (there are thirty-seven of them), and a pixel-level segmentation with three classes: foreground, boundary, and background. The latter is the default; and it’s precisely the kind of goal we’d like.

A name to oxford_pet_dataset(root = dir) will set off the preliminary obtain:

# want torch > 0.6.1
# could must run remotes::install_github("mlverse/torch", ref = remotes::github_pull("713")) relying on whenever you learn this

dir <- "~/.torch-datasets/oxford_pet_dataset"

ds <- oxford_pet_dataset(root = dir)

Photographs (and corresponding masks) come in several sizes. For coaching, nevertheless, we’ll want all of them to be the identical dimension. This may be completed by passing in rework = and target_transform = arguments. However what about knowledge augmentation (mainly at all times a helpful measure to take)? Think about we make use of random flipping. An enter picture will probably be flipped – or not – in response to some likelihood. But when the picture is flipped, the masks higher had be, as effectively! Enter and goal transformations will not be impartial, on this case.

An answer is to create a wrapper round oxford_pet_dataset() that lets us “hook into” the .getitem() technique, like so:

pet_dataset <- torch::dataset(
  inherit = oxford_pet_dataset,
  initialize = perform(..., dimension, normalize = TRUE, augmentation = NULL) {
    self$augmentation <- augmentation
    input_transform <- perform(x) {
      x <- x %>%
        transform_to_tensor() %>%
      # we'll make use of pre-trained MobileNet v2 as a function extractor
      # => normalize with a view to match the distribution of photographs it was skilled with
      if (isTRUE(normalize)) x <- x %>%
        transform_normalize(imply = c(0.485, 0.456, 0.406),
                            std = c(0.229, 0.224, 0.225))
    target_transform <- perform(x) {
      x <- torch_tensor(x, dtype = torch_long())
      x <- x[newaxis,..]
      # interpolation = 0 makes certain we nonetheless find yourself with integer lessons
      x <- transform_resize(x, dimension, interpolation = 0)
    self$cut up <- cut up
      rework = input_transform,
      target_transform = target_transform
  .getitem = perform(i) {
    merchandise <- tremendous$.getitem(i)
    if (!is.null(self$augmentation)) 
      listing(x = merchandise$x, y = merchandise$y[1,..])

All we have now to do now’s create a customized perform that lets us determine on what augmentation to use to every input-target pair, after which, manually name the respective transformation capabilities.

Right here, we flip, on common, each second picture, and if we do, we flip the masks as effectively. The second transformation – orchestrating random modifications in brightness, saturation, and distinction – is utilized to the enter picture solely.

c(224, 224),
                        augmentation = augmentation)
valid_ds <- pet_dataset(root = dir,
                        cut up = "legitimate",
                        dimension = c(224, 224))

train_dl <- dataloader(train_ds, batch_size = 32, shuffle = TRUE)
valid_dl <- dataloader(valid_ds, batch_size = 32)

Mannequin definition

The mannequin implements a traditional U-Web structure, with an encoding stage (the “down” cross), a decoding stage (the “up” cross), and importantly, a “bridge” that passes options preserved from the encoding stage on to corresponding layers within the decoding stage.


First, we have now the encoder. It makes use of a pre-trained mannequin (MobileNet v2) as its function extractor.

The encoder splits up MobileNet v2’s function extraction blocks into a number of phases, and applies one stage after the opposite. Respective outcomes are saved in a listing.

encoder <- nn_module(
  initialize = perform() {
    mannequin <- model_mobilenet_v2(pretrained = TRUE)
    self$phases <- nn_module_list(listing(

    for (par in self$parameters) {

  ahead = perform(x) {
    options <- listing()
    for (i in 1:size(self$phases)) {
      x <- self$phases[[i]](x)
      options[[length(features) + 1]] <- x


The decoder is made up of configurable blocks. A block receives two enter tensors: one that’s the results of making use of the earlier decoder block, and one which holds the function map produced within the matching encoder stage. Within the ahead cross, first the previous is upsampled, and handed via a nonlinearity. The intermediate result’s then prepended to the second argument, the channeled-through function map. On the resultant tensor, a convolution is utilized, adopted by one other nonlinearity.

decoder_block <- nn_module(
  initialize = perform(in_channels, skip_channels, out_channels) {
    self$upsample <- nn_conv_transpose2d(
      in_channels = in_channels,
      out_channels = out_channels,
      kernel_size = 2,
      stride = 2
    self$activation <- nn_relu()
    self$conv <- nn_conv2d(
      in_channels = out_channels + skip_channels,
      out_channels = out_channels,
      kernel_size = 3,
      padding = "similar"
  ahead = perform(x, skip) {
    x <- x %>%
      self$upsample() %>%

    enter <- torch_cat(listing(x, skip), dim = 2)

    enter %>%
      self$conv() %>%

The decoder itself “simply” instantiates and runs via the blocks:

decoder <- nn_module(
  initialize = perform(
    decoder_channels = c(256, 128, 64, 32, 16),
    encoder_channels = c(16, 24, 32, 96, 320)
  ) {

    encoder_channels <- rev(encoder_channels)
    skip_channels <- c(encoder_channels[-1], 3)
    in_channels <- c(encoder_channels[1], decoder_channels)

    depth <- size(encoder_channels)

    self$blocks <- nn_module_list()
    for (i in seq_len(depth)) {
        in_channels = in_channels[i],
        skip_channels = skip_channels[i],
        out_channels = decoder_channels[i]

  ahead = perform(options) {
    options <- rev(options)
    x <- options[[1]]
    for (i in seq_along(self$blocks)) {
      x <- self$blocks[[i]](x, options[[i+1]])

High-level module

Lastly, the top-level module generates the category rating. In our job, there are three pixel lessons. The score-producing submodule can then simply be a last convolution, producing three channels:

mannequin <- nn_module(
  initialize = perform() {
    self$encoder <- encoder()
    self$decoder <- decoder()
    self$output <- nn_sequential(
      nn_conv2d(in_channels = 16,
                out_channels = 3,
                kernel_size = 3,
                padding = "similar")
  ahead = perform(x) {
    x %>%
      self$encoder() %>%
      self$decoder() %>%

Mannequin coaching and (visible) analysis

With luz, mannequin coaching is a matter of two verbs, setup() and match(). The training charge has been decided, for this particular case, utilizing luz::lr_finder(); you’ll doubtless have to vary it when experimenting with completely different types of knowledge augmentation (and completely different knowledge units).

mannequin <- mannequin %>%
  setup(optimizer = optim_adam, loss = nn_cross_entropy_loss())

fitted <- mannequin %>%
  set_opt_hparams(lr = 1e-3) %>%
  match(train_dl, epochs = 10, valid_data = valid_dl)

Right here is an excerpt of how coaching efficiency developed in my case:

# Epoch 1/10
# Prepare metrics: Loss: 0.504                                                           
# Legitimate metrics: Loss: 0.3154

# Epoch 2/10
# Prepare metrics: Loss: 0.2845                                                           
# Legitimate metrics: Loss: 0.2549


# Epoch 9/10
# Prepare metrics: Loss: 0.1368                                                           
# Legitimate metrics: Loss: 0.2332

# Epoch 10/10
# Prepare metrics: Loss: 0.1299                                                           
# Legitimate metrics: Loss: 0.2511

Numbers are simply numbers – how good is the skilled mannequin actually at segmenting pet photographs? To search out out, we generate segmentation masks for the primary eight observations within the validation set, and plot them overlaid on the photographs. A handy approach to plot a picture and superimpose a masks is supplied by the raster package deal.

Pixel intensities must be between zero and one, which is why within the dataset wrapper, we have now made it so normalization may be switched off. To plot the precise photographs, we simply instantiate a clone of valid_ds that leaves the pixel values unchanged. (The predictions, however, will nonetheless must be obtained from the unique validation set.)

valid_ds_4plot <- pet_dataset(
  root = dir,
  cut up = "legitimate",
  dimension = c(224, 224),
  normalize = FALSE

Lastly, the predictions are generated in a loop, and overlaid over the photographs one-by-one:

indices <- 1:8

preds <- predict(fitted, dataloader(dataset_subset(valid_ds, indices)))

png("pet_segmentation.png", width = 1200, top = 600, bg = "black")

par(mfcol = c(2, 4), mar = rep(2, 4))

for (i in indices) {
  masks <- as.array(torch_argmax(preds[i,..], 1)$to(gadget = "cpu"))
  masks <- raster::ratify(raster::raster(masks))
  img <- as.array(valid_ds_4plot[i][[1]]$permute(c(2,3,1)))
  cond <- img > 0.99999
  img[cond] <- 0.99999
  img <- raster::brick(img)
  # plot picture
  raster::plotRGB(img, scale = 1, asp = 1, margins = TRUE)
  # overlay masks
  plot(masks, alpha = 0.4, legend = FALSE, axes = FALSE, add = TRUE)
Learned segmentation masks, overlaid on images from the validation set.

Now onto working this mannequin “within the wild” (effectively, kind of).

JIT-trace and run on Android

Tracing the skilled mannequin will convert it to a kind that may be loaded in R-less environments – for instance, from Python, C++, or Java.

We entry the torch mannequin underlying the fitted luz object, and hint it – the place tracing means calling it as soon as, on a pattern statement:

m <- fitted$mannequin
x <- coro::gather(train_dl, 1)

traced <- jit_trace(m, x[[1]]$x)

The traced mannequin might now be saved to be used with Python or C++, like so:

traced %>% jit_save("")

Nevertheless, since we already know we’d prefer to deploy it on Android, we as a substitute make use of the specialised perform jit_save_for_mobile() that, moreover, generates bytecode:

# want torch > 0.6.1
jit_save_for_mobile(traced_model, "")

And that’s it for the R facet!

For working on Android, I made heavy use of PyTorch Cell’s Android instance apps, particularly the picture segmentation one.

The precise proof-of-concept code for this publish (which was used to generate the under image) could also be discovered right here: (Be warned although – it’s my first Android software!).

In fact, we nonetheless must attempt to discover the cat. Right here is the mannequin, run on a tool emulator in Android Studio, on three photographs (from the Oxford Pet Dataset) chosen for, firstly, a variety in problem, and secondly, effectively … for cuteness:

Where’s my cat?

Thanks for studying!

Parkhi, Omkar M., Andrea Vedaldi, Andrew Zisserman, and C. V. Jawahar. 2012. “Cats and Canine.” In IEEE Convention on Laptop Imaginative and prescient and Sample Recognition.

Ronneberger, Olaf, Philipp Fischer, and Thomas Brox. 2015. “U-Web: Convolutional Networks for Biomedical Picture Segmentation.” CoRR abs/1505.04597.


Leave a Reply

Your email address will not be published. Required fields are marked *