Saving and loading custom models with BatchNormalization layers in TensorFlow

Question

I am training a custom model that includes BatchNormalization layers in TensorFlow 2.12. Along training, I save some checkpoints like this:

checkpoint = tf.train.Checkpoint(step=tf.Variable(0), optimizer=optimizer, model=model)
manager    = tf.train.CheckpointManager(checkpoint, os.path.join(output_folder,"ckpts"), max_to_keep=3)
manager.save(checkpoint_number=int(checkpoint.step)+1)

After the model is trained, I load the checkpoints like this

checkpoint.restore(manager.latest_checkpoint).expect_partial()

The weights are loaded into the model successfully. However, the non-trainable variables moving_variance and moving_mean of the BatchNormalization layers are reset to their default values, 1 and 0 respectively.

Why are the checkpoints not saving the values of moving_variance and moving_mean calculated during training?