Executing the graph in a TensorFlow session

Until now, we have only created tensor objects and added them to a TensorFlow graph for later execution. In this recipe, we will learn how to create a TensorFlow session that can be used to execute (or run) the TensorFlow graph.

Getting ready

Before we run the graph, we should have TensorFlow installed and loaded in R. The installation details can be found in Chapter 1, Getting Started.

How to do it...

1. Load the tensorflow library and import the numpy package:

library(tensorflow)
np <- import("numpy")

2. Reset or remove any existing default_graph:

tf$reset_default_graph()

3. Start an InteractiveSession:

sess <- tf$InteractiveSession()

4. Initialize the global_variables:

sess$run(tf$global_variables_initializer())

5. Run iterations to perform optimization (training):

# Train the model
train_batch_size = 128L
for (i in 1:100) {
spls <- sample(1:dim(train_data$images)[1],train_batch_size)
if (i %% 10 == 0) {
train_accuracy <- accuracy$eval(feed_dict = dict(
x ...