[Bug] _init_callback in base_class does not handle verbose parameter when creating EvalCallback

[BurakDmb]

🐛 Bug

Hi,

I just noticed that in the _init_callback method, the verbose parameter is not used when creating EvalCallback. Since the default verbose value is 1 in the EvalCallback, logging could not be disabled when only using the default verbose value.

Code reference:

stable-baselines3/stable_baselines3/common/base_class.py

Lines 389 to 396 in c4f54fc

	eval_callback = EvalCallback(
	eval_env,
	best_model_save_path=log_path,
	log_path=log_path,
	eval_freq=eval_freq,
	n_eval_episodes=n_eval_episodes,
	)
	callback = CallbackList([callback, eval_callback])

stable-baselines3/stable_baselines3/common/callbacks.py

Line 307 in c4f54fc

verbose: int = 1,

To Reproduce

I would like to set verbose=0 in my studies, and therefore, when I enable evaluation, I cannot decide whether to print the evaluation results or not.

Below, you can see my example code:

model = PPO(..., verbose=0, ...)

model.learn(..., eval_env=eval_env,
            eval_freq=100,
            n_eval_episodes=5,
            eval_log_path="eval")

Expected behavior

I think this problem could be solved easily by passing the self.verbose variable into the EvalCalback constructor.

I can volunteer for this PR.

if eval_env is not None:
    eval_callback = EvalCallback(
        eval_env,
        best_model_save_path=log_path,
        log_path=log_path,
        eval_freq=eval_freq,
        n_eval_episodes=n_eval_episodes,
        verbose=self.verbose,
    )
    callback = CallbackList([callback, eval_callback])

System Info

({'OS': 'Linux-5.13.0-40-generic-x86_64-with-glibc2.17 #45~20.04.1-Ubuntu SMP Mon Apr 4 09:38:31 UTC 2022', 'Python': '3.8.13', 'Stable-Baselines3': '1.6.1a0', 'PyTorch': '1.11.0', 'GPU Enabled': 'True', 'Numpy': '1.22.3', 'Gym': '0.25.0'}, 'OS: Linux-5.13.0-40-generic-x86_64-with-glibc2.17 #45~20.04.1-Ubuntu SMP Mon Apr 4 09:38:31 UTC 2022\nPython: 3.8.13\nStable-Baselines3: 1.6.1a0\nPyTorch: 1.11.0\nGPU Enabled: True\nNumpy: 1.22.3\nGym: 0.25.0\n')

Checklist

• I have checked that there is no similar issue in the repo (required)
• I have read the documentation (required)
• I have provided a minimal working example to reproduce the bug (required)

[qgallouedec]

I see your point. If we want to be rigorous, you are right, verbose=0 should not permit to write anything to stdout.

However, I think many people like the format with evaluation log without training log (currently supported by verbose=0).

I need more feedback on this.

[araffin]

Hello,
good point, verbose value should be passed.
Also related: I was thinking about removing the eval_env parameter completely (#925) as its usage is limited and bring complexity to the codebase.
A better alternative is to create the eval callback outside the model and pass it to the learn() method.

[qgallouedec]

I can volunteer for this PR.

So yes, please go ahead.