tf_function and autograph

[mohamed-180]

from the advanced tutorial in Advanced Quickstart
the part of :

training_loop <- tf_function(autograph(function(train_ds, test_ds) {
  
  for (b1 in train_ds) {
    train_step(b1$x, b1$y)
  }
  
  for (b2 in test_ds) {
    test_step(b2$x, b2$y)
  }
  
  tf$print("Acc", train_accuracy$result(), "Test Acc", test_accuracy$result())
  
  train_loss$reset_states()
  train_accuracy$reset_states()
  test_loss$reset_states()
  test_accuracy$reset_states()
  
}))

can we just use tf_function with out the autograph function based on tf_function have argument to use autograph
see :

tensorflow/R/eager.R

Lines 126 to 145 in e3791c4

	tf_function <- function(f,
	input_signature = NULL,
	autograph = TRUE,
	...) {
	if (!is.function(f))
	stop("`f` must be an R function")
	if (!(isTRUE(autograph) || isFALSE(autograph)))
	stop("`autograph` must be TRUE or FALSE")
	if (autograph) {
	# Can't register tfautograph in Imports yet due to circular dependency
	if(!requireNamespace("tfautograph", quietly=TRUE))
	stop('"tfautograph" package required if autograph=TRUE. Please run install.packages("tfautograph")')
	f <- tfautograph::autograph(f)
	}
	args <- list(py_func(f), input_signature, FALSE, ...)
	do.call(tf$`function`, args)
	}

and use autograph function only in case of executing eagerly.

[t-kalinowski]

Hi, Thanks for asking. Yes, as of version 2.6 tf_function() calls autograph() by default for you and there is no need to call it yourself. The example in the "Advanced Quickstart" needs to be updated. Though, it's harmless to call autograph() on a function twice, so the example should still work even if it is a little verbose.