Always saves a network file upon starting a new training run, handling both seeding and full initializations. Each newly saved network is tagged with "init", but then overwrites the tagged "current" file under the output name.
Currently, this has the (slightly) odd behavior that it will overwrite the pure initialization if training is stopped before the first testing iteration. This behavior is ultimately due to the learning curve loading functionality, which also does not keep consistent records unless training passes the first testing iteration. If this is something we'd like to change, we can add a commit to handle the lc.get_last_it() functionality better.
Also fixes a typo error, and adds an explicit default argument to find_load_net to make the argument parsing logic more clear.
Always saves a network file upon starting a new training run, handling both seeding and full initializations. Each newly saved network is tagged with "init", but then overwrites the tagged "current" file under the output name.
Currently, this has the (slightly) odd behavior that it will overwrite the pure initialization if training is stopped before the first testing iteration. This behavior is ultimately due to the learning curve loading functionality, which also does not keep consistent records unless training passes the first testing iteration. If this is something we'd like to change, we can add a commit to handle the lc.get_last_it() functionality better.
Also fixes a typo error, and adds an explicit default argument to find_load_net to make the argument parsing logic more clear.