Hi all,
I am using tune.fit to train an RL agent, but in each iteration, ray prints out a lot of information making tracking and debugging a bit difficult.
The default verbose in tune.fit(verbose=1) is 1 which prints out a lot of info shown in the following image.
If you set verbose to 0, it does not print anything, but it is not what I want.
I want the tune.fit to only print the following table, not the other information.
I wonder if anyone knows how to make it. Thanks 
This is internal to tune. RLlib onyl returns a results dictionary and tune will decide what to print from it. I’m moving this to the Ray AIR folks
Hi @deepgravity,
usually you should get the expected output with verbose=1. The default is 2, which matches the output you show.
We are refactoring the output engine, and you can try out the new layout by setting the environment variable RAY_AIR_NEW_OUTPUT=1. This works best on the nightly wheels. Do you want to give that a try and test it with AirVerbosity.DEFAULT?
For more details, see
https://docs.ray.io/en/latest/ray-air/experimental-features.html
If all of this does not lead to your goal, you can also write a custom callback to filter out metrics that you don’t need.
Hi @kai,
Many thanks for your reply.
Yes, that worked for me. Indeed:
import os
os.environ['RAY_AIR_NEW_OUTPUT'] = '1'
from ray.tune.experimental.output import AirVerbosity, get_air_verbosity
run_config = air.RunConfig(verbose=get_air_verbosity(AirVerbosity.DEFAULT))
Now, I have the following log:
Only prints the most important measurements.
I think here episodes_total is equivalent to iters that we had before right?
I also have a suggestion. I would even be happier to have only one dynamic table that changes after like every 10 episodes; rather than having multiple new static tables.
This is because I am already tracking the progress on tensorboard or wandb. So I do not need to have the whole run history in the terminal. The terminal is good only for quick debugging. So, having only one dynamic table would be enough.
Thanks anyway
Thanks @deepgravity, we’re actually working on this with out sticky layout. You can try it out already (RAY_AIR_RICH_LAYOUT=1) - but please note that it’s very experimental right now and probably has a few bugs
