Closed roboserg closed 4 years ago
hi @roboserg - I think it makes sense to be able to create a frozen graph pb file from the meta, index, and data files that are created with each save frequency. And then create a .nn file to be used with the inference engine. we have it in our backlog. would you be interested to work on this and create a PR? If so, I can ensure we provide support to review it. Otherwise, might take some time for us to work on this given some other priorities.
nice, thanks! Is the source for the mlagents-learn.exe in this repo? That's the executable that does the training and saving the models, right? (sorry for naive questions)
correct, you can start here in the trainers - https://github.com/Unity-Technologies/ml-agents/blob/develop/ml-agents/mlagents/trainers
If you make any PRs, please use the develop branch.
Thank you for submitting this request. I’ve added it to our internal tracker with the ID MLA-81. I’m going to close this issue for now, but we’ll ping back with any updates
.nn files are now saved at the same time as checkpoints (see https://github.com/Unity-Technologies/ml-agents/pull/4127). This will be in the next release (sometime in August).
You are awesome, thanks
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.
Right now I have to stop the training with CTRL+C in the console to get the current brain model. See how the model performs and continue training. Two features that would save time:
Please add an option to automatically save the .nn model during training. Currently you already save some checkpoint information every 50.000 steps. Please save a .nn model as well.
If the feature above is implemented it would be more then awesome to have a separate instance of the environment running during training at real time speed showing agents with the latest brain. For example when training with 6 instances of the environment the agents train with 100x speed. It would be very helpful if there would be a separate instance run in real time with the lastest saved *.nn brain so that the behavior can be observed.
Right now I have to constantly stop the training to observe the agents, it's very time consuming. Thanks.