How to use trained model in pytorch or tensorflow

I want to use a trained model (.nn file) and load it in pytorch or tensorflow. Is there any script to convert .nn files to pytorch or tensorflow model in ml-agents ? if not what would the best way to do it ?

One solution I tried was to load the checkpoints and create a tf graph but how to create a tensorflow model from this graph ?

I use pytorch but don't have any expertise in tensorflow so any help regarding building the tensorflow model from a tensorflow graph would be really helpful since I can then translate it to creating a pytorch model.

 

Thanks for your reply !!
I am using an offroad vehicle model in Unity and using ml-agents to train a path following controller on it.
I trained ppo using ml-agents and got the checkpoints and frozen graph. We also have the real vehicle in our lab and the next goal is to test this controller on the real vehicle.

I finally used the frozen graph and did inference on it using tensorflow-1 and tensorflow-2. The future goal of the project is to train the policy further on the real vehicle and I need policy model and parameters to do that.

Since I am more comfortable with Pytorch I am looking for ways to create a pytorch policy using the checkpoints or the frozen graph.

One more issue that I faced while doing inference on the graph using tf-1 and tf-2 was that I get different values for the same input unless I initialize the session before every inference, to avoid this I need the details about the policy so that I can prune dropout nodes. I just started using tensorflow so I am not sure if this will actually solve the issue.

Is there any way of saving the tensorflow model in h5 format (I was looking at trainer.py file because I think this is the process that saves checkpoints) since I can than use it to create a pytorch model (Not sure though) and do inference in evaluation mode which will not effect the computational graph and give same output for same input.

I was able to control the simulation using openai-gym interface provided by ml-agents so this will be my last resort to train my own PPO using Pytorch.

One more thing I noticed was during training I get the message that tensorflow-2 is being used by ml-agents but the checkpoints are being saved in tf-1 format I don't know why this is happening.