In rstudio/keras: R Interface to 'Keras'

Introduction

A callback is a powerful tool to customize the behavior of a Keras model during training, evaluation, or inference. Examples include keras.callbacks.TensorBoard to visualize training progress and results with TensorBoard, or keras.callbacks.ModelCheckpoint to periodically save your model during training.

In this guide, you will learn what a Keras callback is, what it can do, and how you can build your own. We provide a few demos of simple callback applications to get you started.

Setup

library(keras3)

Keras callbacks overview

All callbacks subclass the keras.callbacks.Callback class, and override a set of methods called at various stages of training, testing, and predicting. Callbacks are useful to get a view on internal states and statistics of the model during training.

You can pass a list of callbacks (as the keyword argument callbacks) to the following model methods:

• fit()
• evaluate()
• predict()

An overview of callback methods

Global methods

on_(train|test|predict)_begin(logs = NULL)

Called at the beginning of fit/evaluate/predict.

on_(train|test|predict)_end(logs = NULL)

Called at the end of fit/evaluate/predict.

Batch-level methods for training/testing/predicting

on_(train|test|predict)_batch_begin(batch, logs = NULL)

Called right before processing a batch during training/testing/predicting.

on_(train|test|predict)_batch_end(batch, logs = NULL)

Called at the end of training/testing/predicting a batch. Within this method, logs is a named list containing the metrics results.

Epoch-level methods (training only)

on_epoch_begin(epoch, logs = NULL)

Called at the beginning of an epoch during training.

on_epoch_end(epoch, logs = NULL)

Called at the end of an epoch during training.

A basic example

Let's take a look at a concrete example. To get started, let's import tensorflow and define a simple Sequential Keras model:

# Define the Keras model to add callbacks to
get_model <- function() {
  model <- keras_model_sequential()
  model |> layer_dense(units = 1)
  model |> compile(
    optimizer = optimizer_rmsprop(learning_rate = 0.1),
    loss = "mean_squared_error",
    metrics = "mean_absolute_error"
  )
  model
}

Then, load the MNIST data for training and testing from Keras datasets API:

# Load example MNIST data and pre-process it
mnist <- dataset_mnist()

flatten_and_rescale <- function(x) {
  x <- array_reshape(x, c(-1, 784))
  x <- x / 255
  x
}

mnist$train$x <- flatten_and_rescale(mnist$train$x)
mnist$test$x  <- flatten_and_rescale(mnist$test$x)

# limit to 1000 samples
n <- 1000
mnist$train$x <- mnist$train$x[1:n,]
mnist$train$y <- mnist$train$y[1:n]
mnist$test$x  <- mnist$test$x[1:n,]
mnist$test$y  <- mnist$test$y[1:n]

Now, define a simple custom callback that logs:

• When fit/evaluate/predict starts & ends
• When each epoch starts & ends
• When each training batch starts & ends
• When each evaluation (test) batch starts & ends
• When each inference (prediction) batch starts & ends

show <- function(msg, logs) {
  cat(glue::glue(msg, .envir = parent.frame()),
      "got logs: ", sep = "; ")
  str(logs); cat("\n")
}

callback_custom <- Callback(
  "CustomCallback",
  on_train_begin         = \(logs = NULL)        show("Starting training", logs),
  on_epoch_begin         = \(epoch, logs = NULL) show("Start epoch {epoch} of training", logs),
  on_train_batch_begin   = \(batch, logs = NULL) show("...Training: start of batch {batch}", logs),
  on_train_batch_end     = \(batch, logs = NULL) show("...Training: end of batch {batch}",  logs),
  on_epoch_end           = \(epoch, logs = NULL) show("End epoch {epoch} of training", logs),
  on_train_end           = \(logs = NULL)        show("Stop training", logs),


  on_test_begin          = \(logs = NULL)        show("Start testing", logs),
  on_test_batch_begin    = \(batch, logs = NULL) show("...Evaluating: start of batch {batch}", logs),
  on_test_batch_end      = \(batch, logs = NULL) show("...Evaluating: end of batch {batch}", logs),
  on_test_end            = \(logs = NULL)        show("Stop testing", logs),

  on_predict_begin       = \(logs = NULL)        show("Start predicting", logs),
  on_predict_end         = \(logs = NULL)        show("Stop predicting", logs),
  on_predict_batch_begin = \(batch, logs = NULL) show("...Predicting: start of batch {batch}", logs),
  on_predict_batch_end   = \(batch, logs = NULL) show("...Predicting: end of batch {batch}", logs),
)

Let's try it out:

model <- get_model()
model |> fit(
  mnist$train$x, mnist$train$y,
  batch_size = 128,
  epochs = 2,
  verbose = 0,
  validation_split = 0.5,
  callbacks = list(callback_custom())
)

res <- model |> evaluate(
  mnist$test$x, mnist$test$y,
  batch_size = 128, verbose = 0,
  callbacks = list(callback_custom())
)

res <- model |> predict(
  mnist$test$x,
  batch_size = 128, verbose = 0,
  callbacks = list(callback_custom())
)

Usage of logs list

The logs named list contains the loss value, and all the metrics at the end of a batch or epoch. Example includes the loss and mean absolute error.

callback_print_loss_and_mae <- Callback(
  "LossAndErrorPrintingCallback",

  on_train_batch_end = function(batch, logs = NULL)
    cat(sprintf("Up to batch %i, the average loss is %7.2f.\n",
                batch,  logs$loss)),

  on_test_batch_end = function(batch, logs = NULL)
    cat(sprintf("Up to batch %i, the average loss is %7.2f.\n",
                batch, logs$loss)),

  on_epoch_end = function(epoch, logs = NULL)
    cat(sprintf(
      "The average loss for epoch %2i is %9.2f and mean absolute error is %7.2f.\n",
      epoch, logs$loss, logs$mean_absolute_error
    ))
)


model <- get_model()
model |> fit(
  mnist$train$x, mnist$train$y,
  epochs = 2, verbose = 0, batch_size = 128,
  callbacks = list(callback_print_loss_and_mae())
)

res = model |> evaluate(
  mnist$test$x, mnist$test$y,
  verbose = 0, batch_size = 128,
  callbacks = list(callback_print_loss_and_mae())
)

For more information about callbacks, you can check out the Keras callback API documentation.

Usage of self$model attribute

In addition to receiving log information when one of their methods is called, callbacks have access to the model associated with the current round of training/evaluation/inference: self$model.

Here are of few of the things you can do with self$model in a callback:

• Set self$model$stop_training <- TRUE to immediately interrupt training.
• Mutate hyperparameters of the optimizer (available as self$model$optimizer), such as self$model$optimizer$learning_rate.
• Save the model at period intervals.
• Record the output of model |> predict() on a few test samples at the end of each epoch, to use as a sanity check during training.
• Extract visualizations of intermediate features at the end of each epoch, to monitor what the model is learning over time.
• etc.

Let's see this in action in a couple of examples.

Examples of Keras callback applications

Early stopping at minimum loss

This first example shows the creation of a Callback that stops training when the minimum of loss has been reached, by setting the attribute self$model$stop_training (boolean). Optionally, you can provide an argument patience to specify how many epochs we should wait before stopping after having reached a local minimum.

callback_early_stopping() provides a more complete and general implementation.

callback_early_stopping_at_min_loss <- Callback(
  "EarlyStoppingAtMinLoss",
  `__doc__` =
    "Stop training when the loss is at its min, i.e. the loss stops decreasing.

    Arguments:
        patience: Number of epochs to wait after min has been hit. After this
        number of no improvement, training stops.
    ",

  initialize = function(patience = 0) {
    super$initialize()
    self$patience <- patience
    # best_weights to store the weights at which the minimum loss occurs.
    self$best_weights <- NULL
  },

  on_train_begin = function(logs = NULL) {
    # The number of epoch it has waited when loss is no longer minimum.
    self$wait <- 0
    # The epoch the training stops at.
    self$stopped_epoch <- 0
    # Initialize the best as infinity.
    self$best <- Inf
  },

  on_epoch_end = function(epoch, logs = NULL) {
    current <- logs$loss
    if (current < self$best) {
      self$best <- current
      self$wait <- 0L
      # Record the best weights if current results is better (less).
      self$best_weights <- get_weights(self$model)
    } else {
      add(self$wait) <- 1L
      if (self$wait >= self$patience) {
        self$stopped_epoch <- epoch
        self$model$stop_training <- TRUE
        cat("Restoring model weights from the end of the best epoch.\n")
        model$set_weights(self$best_weights)
      }
    }
  },

  on_train_end = function(logs = NULL)
    if (self$stopped_epoch > 0)
      cat(sprintf("Epoch %05d: early stopping\n", self$stopped_epoch + 1))
)
`add<-` <- `+`


model <- get_model()
model |> fit(
  mnist$train$x,
  mnist$train$y,
  epochs = 30,
  batch_size = 64,
  verbose = 0,
  callbacks = list(callback_print_loss_and_mae(),
                   callback_early_stopping_at_min_loss())
)

Learning rate scheduling

In this example, we show how a custom Callback can be used to dynamically change the learning rate of the optimizer during the course of training.

See keras$callbacks$LearningRateScheduler for a more general implementations (in RStudio, press F1 while the cursor is over LearningRateScheduler and a browser will open to this page).

callback_custom_learning_rate_scheduler <- Callback(
  "CustomLearningRateScheduler",
  `__doc__` =
  "Learning rate scheduler which sets the learning rate according to schedule.

    Arguments:
        schedule: a function that takes an epoch index
            (integer, indexed from 0) and current learning rate
            as inputs and returns a new learning rate as output (float).
    ",

  initialize = function(schedule) {
    super$initialize()
    self$schedule <- schedule
  },

  on_epoch_begin = function(epoch, logs = NULL) {
    ## When in doubt about what types of objects are in scope (e.g., self$model)
    ## use a debugger to interact with the actual objects at the console!
    # browser()

    if (!"learning_rate" %in% names(self$model$optimizer))
      stop('Optimizer must have a "learning_rate" attribute.')

    # # Get the current learning rate from model's optimizer.
    # use as.numeric() to convert the keras variablea to an R numeric
    lr <- as.numeric(self$model$optimizer$learning_rate)
    # # Call schedule function to get the scheduled learning rate.
    scheduled_lr <- self$schedule(epoch, lr)
    # # Set the value back to the optimizer before this epoch starts
    optimizer <- self$model$optimizer
    optimizer$learning_rate <- scheduled_lr
    cat(sprintf("\nEpoch %03d: Learning rate is %6.4f.\n", epoch, scheduled_lr))
  }
)

LR_SCHEDULE <- tibble::tribble(
  ~start_epoch, ~learning_rate,
             0,            0.1,
             3,           0.05,
             6,           0.01,
             9,          0.005,
            12,          0.001,
  )

last <- function(x) x[length(x)]
lr_schedule <- function(epoch, learning_rate) {
  "Helper function to retrieve the scheduled learning rate based on epoch."
  with(LR_SCHEDULE, learning_rate[last(which(epoch >= start_epoch))])
}

model <- get_model()
model |> fit(
  mnist$train$x,
  mnist$train$y,
  epochs = 14,
  batch_size = 64,
  verbose = 0,
  callbacks = list(
    callback_print_loss_and_mae(),
    callback_custom_learning_rate_scheduler(lr_schedule)
  )
)

Built-in Keras callbacks

Be sure to check out the existing Keras callbacks by reading the API docs. Applications include logging to CSV, saving the model, visualizing metrics in TensorBoard, and a lot more!

rstudio/keras documentation built on May 17, 2024, 9:23 p.m.