Performance issues about tf.function

[DLPerf]

Hello! Our static bug checker has found a performance issue in dota2/omninight/learner_dota.py: create_host is repeatedly called in a for loop, but there are tf.function decorated functions inference and agent_inference defined and called in create_host.

In that case, when create_host is called in a loop, the function inference and agent_inference will create a new graph every time, and that can trigger tf.function retracing warning.

Similar issues in: dota2/shadowfiend/learner_dota.py, dr-derks-mutant-battlegrounds/learner_2.py and dr-derks-mutant-battlegrounds/learner_1.py.

Here is the tensorflow document to support it.

Briefly, for better efficiency, it's better to use:

@tf.function
def inner():
    pass

def outer():
    inner()  

than:

def outer():
    @tf.function
    def inner():
        pass
    inner()

Looking forward to your reply.

[DLPerf]

However, inference_device is called here, is there any good idea? @kimbring2

[kimbring2]

@DLPerf Hello, sorry for late response. Thank you for sharing good idea about the tf.function. Actually, I should train the model using CPU because of constant leaking of memory.

I could not solve that issue before. However, it seems like moving the tf.function inside of function could solve that issue.

Actually, I just borrow the train code from the Google Seed_RL project. Therefore, I am not sure I can change the code like your advice.

Lucklily, I could train the agent without GPU because the observation of Dota2 does not have any image frame.

[kimbring2]

@DLPerf Anyway, could you run the Dota2 environment successfully? I only tested it on my workspace. Therefore, I am not sure the environment works well also in the another workspace.

I found that recent update of Dota2 raises issue in the the GRPC part of Dotaservice. That is why I upload the previous Dota2 client version in Google Drive.