Question mlagents not generating nn file

mlagents is not generating a nn file at the end of training. Is there a way to create one from a checkpoint or an onnx file?

 

As of Release 10, the default behavior is to use pytorch for trainers, which only produce .onnx files. You can still force tensorflow for training (which will produce .nn files, and .onnx files if you have tf2onnx installed) by passing "--tensorflow" on the commandline to mlagents-learn, or setting "framework: tensorflow" in the configuration file. Note that we'll likely be removing tensorflow support in the next release or two.

There is no way to create a .nn file from .onnx files or checkpoints, and we do not plan to add one.

Can I ask why you don't want to use the .onnx file?

 

I didn't realize that you could. A lot of the tutorials and documentation seems to focus on the older tf implementation. Thanks for the help.

 

I think we're a little behind about updating the examples and documentation. Not much we can do about the existing tutorials, though. But they should behave the same.