close display for robothor

[GracefulMan]

how can i close the display of unity, since it will create many display window when training!

[Lucaweihs]

Hi Hongying,

Unfortunately I don't think this is currently possible as Unity needs to render to a display when training. Unity does have a -batchmode option that allows for running without a display but some preliminary testing by me suggests that the images it returns when using this mode are blank. @ekolve would you know of workaround so that things don't get cluttered? One (slighly hacky): depending on your OS would it be possible to start training in its own separate screen/workspace so that all of the unity windows have their dedicated area?

[ekolve]

Could you tell me what platform you are on?

[GracefulMan]

I run the code on ubuntu 20.04, it will open many unity window.

[GracefulMan]

i want to know whether it will delay the trainning process? now there are a lot of display window when train the agent.

[Lucaweihs]

Hi Hongying, it shouldn't have a large impact on the training speed. If you tell me about your computer's hardware (# processes + # gpus) and which command you're running I should be able to give you a ballpark figure of what FPS you should be seeing during training (the FPS is printed every time AllenAct logs training metrics / losses).

[GracefulMan]

Hi, my hardware information:
Processor: Intel® Core™ i7-8700K CPU @ 3.70GHz × 12
Graphics: NVIDIA Corporation TU104 [GeForce RTX 2080] , only 1 GPU.
the command is from tutorial:

python main.py object_nav_ithor_ppo_one_object -b projects/tutorials -o one_ppo

the following is screenshot when run the command:

[ekolve]

Run the following from the shell:

sudo apt-get install xdotool

Then run the following script after starting your training process:

#!/bin/bash

for w in $(xdotool search --name '^AI2-Thor$'); do 
    xdotool windowminimize "$w" 
done

[Lucaweihs]

Hi @GracefulMan,

Did @ekolve's suggestion work for you? The FPS rates you're getting (~300) seem quite reasonable for a single GPU. The highest FPS we managed (with an 8 GPU machine) is around 1200 FPS for navigation.