To solve the problem of
"ValueError: GraphDef cannot be larger than 2GB."

tf.slice(data, indexTensor, sizeTensor) seems to occupy the size of data in the graph.

"Make big numpy arrays and convert to tensor" results in big graph while "make big tensors directly with tf.constant()" doesn't.

Making nodes of graph within a loop results in multiple copies of the nodes in the graph.

Checkpoints saves only the variable values such as weights. Writing summary saves the graph, which can be seen by Tensorboard.