Output of .nn model

Hello,

In my attempt to training my own RL algorithm in Python, I must create a model that takes as input the agent observation and outputs the action(s) that it believes is the best.
However, what output exactly does the Tensorflow model (and thus the .nn model) needs to have in order to be « understood » in Unity ? is it the probabilities of each actions ? The log probs ? The actions selected directly?
I have searched through the docs but didn’t find the answer.

Thank you.

 

Thank you for your response.
I was wondering about the exact output shapes (as well as inputs) that the Tensorflow model should have in order to be, once converted to a .nn file, correctly interpreted by Unity. After digging through the code of mlagents-learn, here is what I found for those interested :

As inputs, the model should have :

• a placeholder named "vector_observation" of shape [None, vec_obs_size]
• a placeholder named "action_masks" of shape [None, num_act], where num_act is the total number of actions possible (every branch included)

As output, the model should have :

• a node called "action" which contains the log probs for each actions

Also, the model should contains the following constants :

• "version_number", of type tf.int32
• "memory_size" of type tf.int32
• "is_continuous_control" of type tf.int32
  
• "action_output_shape" of type tf.int32 (total number of actions possible, every branch included)

If you want an example of a simple model built with Tensorflow, then converted to a .nn file using the script tensorflow-to-barracuda.py, here's a simple Jupyter notebook : https://github.com/Procuste34/Unity...ow-to-barracuda/tensorflowModelCreation.ipynb

 

No, I was trying (and succedeed) to create .nn model that can be understood by Unity directly. I found out the different necessary nodes that need to be in the model to be understood by Unity (elsewise you get errors in Unity).

 

Hi Procuste,

Can you please help me find the resource regarding the necessary nodes or just tips on how to create a model suitable for a .nn file and Unity. That would be of great help.

 

I am also interested, I tried to train a model using stable baselines but I didn't manage to successfully convert the tensorflow graph into a .nn using tensorflow-to-barracuda.py. I always get errors.