[SOLVED] Unity machine learning using internal brain for pre-trained networks results in crash

Unity machine learning using internal brain for pre-trained networks results in crash

I try to get the machine learning project of Unity running. The project can be found here: https://github.com/Unity-Technologies/ml-agents

Following all the necessary steps to set everything up, everything works fine until I try to use a pre-trained model in an internal brain, e.g in the 3D-Ball example. On pressing play in the editor, it says to have stopped working and the editor crashes. Same goes for a successfully compiled executable.

Since I can't find any crash logs, I absolutely don't know what's the cause to this. Anybody had the same issue and solved it, or any ideas on this?

I'll just add the steps I made to set my project up, in case somebody wants to reproduce or check for mistakes I made:

Download zipped project from github
Open project with Unity 2018.1.1f1
Add ENABLE_TENSORFLOW to predefined symbols
Download and install TensorflowSharp Plugin
In the 3DBall Scene select the brain and switch BrainType to Internal
Run in editor or compile and run, results both times in a crash

EDIT

What's making me wonder is that I got it working just fine training an external brain in python using a compiled executable. So my thoughts were it should be much more easy to just forward pass a trained net, if it's working doing forward pass and back propagation. But it says though that external brain still is experimental, so not sure if that just will not work until some future patches.

EDIT 2

So I tried using an internal brain in an older version of the githubs project, which then worked quite fine. Also I used an older version of the TensforflowSharp Plugin. I'll try to figure out, which version still works and gonna post my results here ;)

Solution

To conclude, the problem was actually the recent version of the TensorflowSharp Plugin. Using an older version, which I found somewhere worked fine. I don't have the link to that version anymore, but in case someone experiences the exact same problem, I could search for it or upload it somewhere.