Loading data

knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>",
  eval = identical(Sys.getenv("TORCH_TEST", unset = "0"), "1"),
  purl = FALSE
)
library(torch)

Datasets and data loaders

Central to data ingestion and preprocessing are datasets and data loaders. A dataset is an object that holds the data to use, while a data loader is an object that will load the data from a dataset providing a way to access subsets of the data. By using datasets and data loaders you will have a process for clearly organizing your data and passing it to other components of the torch package, such as model training.

Built into torch are premade datasets that are commonly used in machine learning, such as the MNIST handwriting dataset (mnist_dataset()). Most of the prebuilt datasets relate to image recognition and natural language processing.

Below is an example of how you would use the MNIST dataset with a dataloader. First, the minst_dataset() function is used to create ds which is a Dataset object. Then a dataloader dl is created to query that data. Finally, that dataloader is used in a coro::loop() to iterate over batches of that data:

# Create a dataset from included data
ds <- mnist_dataset(
  dir, 
  download = TRUE, 
  transform = function(x) {
    x <- x$to(dtype = torch_float())/256
    x[newaxis,..]
  }
)

# Create the loader to query the data in batches
dl <- dataloader(ds, batch_size = 32, shuffle = TRUE)

coro::loop(for (b in dl)) {
# use the data from each batch `b` here
# ...
})

See vignettes/examples/mnist-cnn.R for a complete example.

In the more common situation where you have a unique set of data that isn't included with the package you'll need to make a custom Dataset subclass by using the dataset() function. The custom Dataset subclass is an abstract R6 container for the data. It will need to know some information about the particular dataset, such as how to iterate over it.

At a minimum, when using dataset() to create a custom Dataset class you'll want to define the following:

Example of using a custom Dataset

While this may sound complicated the base logic is only a few steps--the complexity often comes from the data itself and how involved your preprocessing is. Here we show how to create your own Dataset class to train on Allison Horst\'s penguins.

| Component | Dataset R6 class | Dataset object | DataLoader object | batch | |:----|:---------:|:---:|:---:|:---:| | Description | Output of dataset(). When calling dataset() it should have at least a name, initialize, .getitem, and .length. Output is a Dataset generator. | Object created by using the custom Dataset generator. Actually stores the data | Object that queries the Dataset object to pull batches of data | The subsample of data used for things like model training | | Penguin example | penguins_dataset | tuxes | dl | b|

library(palmerpenguins)
library(magrittr)

penguins

In addition, any number of helper functions can be defined.

Here, we assume the penguins have already been loaded, and all preprocessing consists in removing lines with NA values, transforming factors to numbers starting from 0, and converting from R data types to torch tensors.

In .getitem, we essentially decide how this data is going to be used: All variables besides species go into x, the predictor, and species will constitute y, the target. Predictor and target are returned in a list, to be accessed as batch[[1]] and batch[[2]] during training.

penguins_dataset <- dataset(

  name = "penguins_dataset",

  initialize = function() {
    self$data <- self$prepare_penguin_data()
  },

  .getitem = function(index) {

    x <- self$data[index, 2:-1]
    y <- self$data[index, 1]$to(torch_long())

    list(x, y)
  },

  .length = function() {
    self$data$size()[[1]]
  },

  prepare_penguin_data = function() {

    input <- na.omit(penguins) 
    # conveniently, the categorical data are already factors
    input$species <- as.numeric(input$species)
    input$island <- as.numeric(input$island)
    input$sex <- as.numeric(input$sex)

    input <- as.matrix(input)
    torch_tensor(input)
  }
)

Let's create the dataset , query for it's length, and look at its first item:

tuxes <- penguins_dataset()
tuxes$.length()
tuxes$.getitem(1)

To be able to iterate over tuxes, we need a data loader (we override the default batch size of 1):

dl <- tuxes %>% dataloader(batch_size = 8)

Calling .length() on a data loader (as opposed to a dataset) will return the number of batches we have:

dl$.length()

And we can create an iterator to inspect the first batch:

iter <- dl$.iter()
b <- iter$.next()
b

To train a network, we can use coro::loop() to iterate over batches.

Training with data loaders

Our example network is very simple. (In reality, we would want to treat island as the categorical variable it is, and either one-hot-encode or embed it.)

net <- nn_module(
  "PenguinNet",
  initialize = function() {
    self$fc1 <- nn_linear(7, 32)
    self$fc2 <- nn_linear(32, 3)
  },
  forward = function(x) {
    x %>% 
      self$fc1() %>% 
      nnf_relu() %>% 
      self$fc2() %>% 
      nnf_log_softmax(dim = 1)
  }
)

model <- net()

We still need an optimizer:

optimizer <- optim_sgd(model$parameters, lr = 0.01)

And we're ready to train:

for (epoch in 1:10) {

  l <- c()

  coro::loop(for (b in dl) {
    optimizer$zero_grad()
    output <- model(b[[1]])
    loss <- nnf_nll_loss(output, b[[2]])
    loss$backward()
    optimizer$step()
    l <- c(l, loss$item())
  })

  cat(sprintf("Loss at epoch %d: %3f\n", epoch, mean(l)))
}

Through this example we have trained a deep learning model using dataset() to define a custom class and then loaded it in batches with a data loader. By using the dataset and data loader we were able to write code that split the data preprocessing and setup from the model training itself.

Notes on efficiency

When using datasets and data loaders you may find that under certain conditions your code is running more slowly than you'd expect. In some situations the overhead of using dataloaders and datasets can impact overall performance. This may change in time as the R/C++ integration of Torch improves, but for now there are some workarounds:

Use .getbatch() instead of .getitem()

By default a dataloader will use the .getitem() member function to pull each single datapoint individually. You can speed this up by switching to using .getbatch() which will pull all the datapoints in a batch at once:

penguins_dataset_batching <- dataset(

  name = "penguins_dataset_batching",

  initialize = function() {
    self$data <- self$prepare_penguin_data()
  },

  # the only change is that this went from .getitem to .getbatch
  .getbatch = function(index) {

    x <- self$data[index, 2:-1]
    y <- self$data[index, 1]$to(torch_long())

    list(x, y)
  },

  .length = function() {
    self$data$size()[[1]]
  },

  prepare_penguin_data = function() {

    input <- na.omit(penguins) 
    # conveniently, the categorical data are already factors
    input$species <- as.numeric(input$species)
    input$island <- as.numeric(input$island)
    input$sex <- as.numeric(input$sex)

    input <- as.matrix(input)
    torch_tensor(input)
  }
)

In many instances the only change is to exactly replace just .getitem with .getbatch since often the .getitem function is written to handle vectors of indices. In this penguins example the .getitem function used the index to select the rows, which will work fine with a vector instead

Remove dataset dataloader and manually define the function calls

If switching to .getbatch does not provide the benefit you were expecting you could also remove the dataset entirely and manually pass the data. At this point you are trading readability of your code and convenience for speed.

input <- na.omit(penguins) 
# conveniently, the categorical data are already factors
input$species <- as.numeric(input$species)
input$island <- as.numeric(input$island)
input$sex <- as.numeric(input$sex)

input <- as.matrix(input)
input <- torch_tensor(input)

data_x <- input[, 2:-1]
data_y <- input[, 1]$to(torch_long())

batch_size <- 8
num_data_points <- data_y$size(1)
num_batches <- floor(num_data_points/batch_size)

for(epoch in 1:10){

  # rearrange the data each epoch
  permute <- torch_randperm(num_data_points) + 1L
  data_x <- data_x[permute]
  data_y <- data_y[permute]

  # manually loop through the batches
  for(batch_idx in 1:num_batches){

    # here index is a vector of the indices in the batch
    index <- (batch_size*(batch_idx-1) + 1):(batch_idx*batch_size)

    x <- data_x[index]
    y <- data_y[index]$to(torch_long())

    optimizer$zero_grad()
    output <- model(x)
    loss <- nnf_nll_loss(output, y)
    loss$backward()
    optimizer$step()
    l <- c(l, loss$item())
  }

  cat(sprintf("Loss at epoch %d: %3f\n", epoch, mean(l)))
}


Try the torch package in your browser

Any scripts or data that you put into this service are public.

torch documentation built on May 29, 2024, 9:54 a.m.