How to restart a training session that has completed

[bwanab]

I've written a reversi (Othello) game interface. For params.jl, I used a modified version of the mancala params.jl. For my first run, I set num_iters to 5. The results are somewhat encouraging, so I'd like to continue with num_iters set much higher, but when I invoke Scipts.train(exp), I get the message: "Training completed".

I tried changing num_iters in params.jl and in the session params.json, but same result in both cases.

Is there a way to get the training to resume from the point it has reached, or do I just have to delete the session and start over?

[jonathan-laurent]

The reason is that when loading a session, hyperparameters are loaded from "params.data" and not "params.json" (you can look at the source in session.jl). The former is not written in human-readable format. The latter is human-readable but not used to load sessions.

I agree this is not very elegant. The reason I had to do this is that last I checked, JSON3.jl was pretty good to at dumping complex data into JSON without much boilerplate but not as good at parsing back this data. See this issue.

Don't hesitate to contribute an improvement here.

[bwanab]

Thanks. I looked at the code and see what you mean. But, from your answer I take it that there's no other reason why one couldn't restart a training session assuming that some reasonably elegant mechanism exited to set the value of env.num_iters to a higher value.

[smart-fr]

There is a way, though not very elegant: #118 (comment)