Save the date! Google I/O returns May 18-20 Register now


A checkpointable object whose NumPy array attributes are saved/restored.

Example usage:

arrays = numpy_storage.NumpyState()
checkpoint = tf.train.Checkpoint(numpy_arrays=arrays)
arrays.x = np.ones([3, 4])
directory = self.get_temp_dir()
prefix = os.path.join(directory, 'ckpt')
save_path =
arrays.x[:] = 0.
assert (arrays.x == np.zeros([3, 4])).all()
assert (arrays.x == np.ones([3, 4])).all()

second_checkpoint = tf.train.Checkpoint(
# Attributes of NumpyState objects are created automatically by restore()
assert (second_checkpoint.numpy_arrays.x == np.ones([3, 4])).all()

Note that NumpyState objects re-create the attributes of the previously saved object on restore(). This is in contrast to TensorFlow variables, for which a Variable object must be created and assigned to an attribute.

This snippet works both when graph building and when executing eagerly. On save, the NumPy array(s) are fed as strings to be saved in the checkpoint (via a placeholder when graph building, or as a string constant when executing eagerly). When restoring they skip the TensorFlow graph entirely, and so no restore ops need be run. This means that restoration always happens eagerly, rather than waiting for checkpoint.restore(...).run_restore_ops() like TensorFlow variables when graph building.