Write out preliminary .nn models

[lostandsupine]

Is your feature request related to a problem? Please describe.
There can be a lot of iteration necessary with tuning the hyper-parameters and setting up correct rewards. When using executable environments it would be really helpful to see how the agents are performing during training, whether they are narrowing in on some degenerate or unintended behavior, etc. Especially when trying to learn a moderately complex game on a slow computer.

Describe the solution you'd like
A command line option or parameter in the yaml config to write out preliminary .nn models at set steps, e.g. "write_steps: 1e5"

Describe alternatives you've considered
Stopping early and restarting with --load, but this has other problems. Leaving graphics on, but that would just slow things down and can be hard to actually see what is happening with the increased frame rate. Manually running export_model() on a checkpoint in python, but I haven't tried and it wouldn't be as convenient.

Additional context
Or maybe I am just impatient and can't wait to see what the agents are doing :)

[awjuliani]

Hello @lostandsupine

Thank you for making this request. We have actually had a number of people ask for this recently, and we are looking into a way of making this possible. Currently there is some difficulty based on how TensorFlow deals with "freezing" models (which is required to produce a .nn file), but we know how useful it would be for people, so we want to make it happen.

We are internally tracking this request with MLA-315.

[chriselion]

.nn files are now saved at the same time as checkpoints (see #4127). This will be in the next release (sometime in August).

[github-actions]

This thread has been automatically locked since there has not been any recent activity after it was closed. Please open a new issue for related bugs.