Restart training from checkpoints

[mixmixmix]

With semi-frequent eculid restarts it would be good if we can restart training from a checkpoint.
I think it used to work, and maybe I have deactivated it at some point. It looks like checkpoints are saved but not used when starting training.

@ctorney could you take a look? Branch master.

[mixmixmix]

https://www.tensorflow.org/api_docs/python/tf/keras/callbacks/ModelCheckpoint

• save weights only needs to be true,
• the path to the cehckpoints must be different so it saves in hd5 format
• seems that it needs to be loaded manually, save_weights_on_restart is now deprecated