Why does the TensorFlow Estimator API take inputs as a lambda?

Question

The tf.estimator API takes input "input functions" that return Datasets. For example, Estimator.train() takes an input_fn (documentation).

In the examples I've seen, whenever this function is supplied manually, it is an argumentless lambda.

Doesn't that mean that the function always returns the same value? Or is it invoked multiple times with no arguments? I wasn't able to find documentation about this. Why don't functions like train() just take input as a Dataset explicitly?

Answer 1

Dataset objects are also backed by nodes in the computational graph. The Estimator constructs the computational graph on each all to train(), evaluate() etc. By doing so the Estimator API ensures that operations on different Estimator objects are isolated and Tensor/Dataset into independent graphs.

Some code pointers (for TensorFlow 1.7) in case you're interested:

• Estimator.train() eventually invokes this
• Which in turn invokes input_fn

Hope that helps.

Answer 2

According to Tensorflow documentation:

"Estimators expect an input_fn to take no arguments. To work around this restriction, we use lambda to capture the arguments and provide the expected interface."

https://www.tensorflow.org/guide/datasets_for_estimators