Difference between tf.train.Checkpoint and tf.train.Saver

Question

I found there are different ways to save/restore models and variables in Tensorflow. These ways including:

• tf.saved_model.simple_save
• tf.train.Checkpoint
• tf.train.Saver

In tensorflow's documentations, I found some differences between them:

1. tf.saved_model is a thin wrapper around tf.train.Saver
2. tf.train.Checkpoint support eager execution but tf.train.Saver not.
3. tf.train.Checkpoint not creating .meta file but still can load graph structure (here is a big question! how it can do that?)

How tf.train.Checkpoint can load graph without .meta file? or more generally What is the difference between tf.train.Saver and tf.train.Checkpoint?

Answer 1

According to Tensorflow docs:

Checkpoint.save and Checkpoint.restore write and read object-based checkpoints, in contrast to tf.train.Saver which writes and reads variable.name based checkpoints. Object-based checkpointing saves a graph of dependencies between Python objects (Layers, Optimizers, Variables, etc.) with named edges, and this graph is used to match variables when restoring a checkpoint. It can be more robust to changes in the Python program, and helps to support restore-on-create for variables when executing eagerly. Prefer tf.train.Checkpoint over tf.train.Saver for new code.