Data interpretation logged by tensorboard_log

[HarukiKozukapenguin]

Thank you for interesting simulator!

I checked run_vision_ppo.py by following command.
python3 -m python.run_vision_ppo --render 0 --train 1
And, I found data when they train in the envtest/python/saved directory (e.g. PPO_1, PPO_2). I found some policies when they're training(/policy), and Test Trajectory(/TestTraj).
The questions I would like to ask is as follows.

1. where is the logging of reward transition when they train?
2. what does each axis means in the graph of TestTraj/Plots?
3. which code define the parameter of plotting or logging?

[yun-long]

hi,

1. the training reward is logged in tensorboard. you can to to saved and run

tensorboard --logdir=./

2. the plots in the first row are position [x, y, z] and the plots in the second row are velocity [x, y, z]

3. the plotting is done here

[HarukiKozukapenguin]

1. Which directory should I run this command (tensorboard --logdir=./), or should I run this command b/f I run python3 -m python.run_vision_ppo --render 0 --train 1?

[HarukiKozukapenguin]

I move to a directory of saved/PPO_(num) and I run tensorboard --logdir=./ then I can seer transition of each reward.
Thank you!

[HarukiKozukapenguin]

@yun-long

I have one question about interpretation of TensorBoard.
I can see rewards transition when I learns in the simulation, but I do not know how what does it means.

1. I think the reward is a sum of reward of each episode, Is it correct?
2. Is this rewards a rewards when agent trains? or rewards when agent evaluate?

[yun-long]

Hi,

the reward you see on Tensorboard is from here.

In summary, contains the sum reward and each individual reward component. The reward is a training reward, not an evaluation reward.

[HarukiKozukapenguin]

@yun-long
Which code writes these rewards to TensorBoard?

[yun-long]

hi

the code is here

[HarukiKozukapenguin]

Thank you!