Created by: zdevito
It seems to cause problems:
line 254, in <lambda>
self._notify_manager(lambda: self._io.write(b))
OSError: [Errno 14] Bad address
But it isn't even needed for saving while also doing training. We can just use concurrent.futures.ThreadPoolExecutor to do the copies from another thread.
Note: how do I verify this doesn't upset the training process timings?