Can't tune hyperparameters with CustomSACPolicy - multiple values for keyword argument 'layers'

[PierreExeter]

Describe the bug
Running hyperparameter tuning with SAC and CustomSACPolicy returns

TypeError: __init__() got multiple values for keyword argument 'layers'.

Note, hyperparameter tuning is working fine with MlpPolicy and normal training is working fine with CustomSACPolicy. The issue seems to be coming from Tensorflow.

Code example

After a recent git clone and using the defaults hyperparameters in hyperparameters/sac.yml:

python train.py --algo sac --env HopperBulletEnv-v0 -n 50000 -optimize --n-trials 100 --n-jobs 1

Full traceback:

Traceback (most recent call last):
  File "***/bin/anaconda3/lib/python3.7/site-packages/optuna/study.py", line 648, in _run_trial
    result = func(trial)
  File "***/rl-baselines-zoo/utils/hyperparams_opt.py", line 88, in objective
    model = model_fn(**kwargs)
  File "train.py", line 373, in create_model
    verbose=0, **kwargs)
  File "***/bin/anaconda3/lib/python3.7/site-packages/stable_baselines/sac/sac.py", line 125, in __init__
    self.setup_model()
  File "***/bin/anaconda3/lib/python3.7/site-packages/stable_baselines/sac/sac.py", line 145, in setup_model
    **self.policy_kwargs)
  File "***/rl-baselines-zoo/utils/utils.py", line 71, in __init__
    feature_extraction="mlp")
TypeError: __init__() got multiple values for keyword argument 'layers'

System Info

• stable baselines: 2.10.0 (installed with pip)
• rl-baselines-zoo commit: 645ea17
• Python 3.7.4
• Tensorflow: 1.14.0
• Gym: 0.15.4
• Pybullet: 2.5.8
• Ubuntu 18.04
• GPU: GeForce GTX 1060
• CUDA: 10.2

[araffin]

Hello,
This is normal. If you do hyperparameter tuning, you should set policy='MlpPolicy' otherwise you will get the mentioned error, as the CustomSACPolicy is already custom in term of number of layers, would be nice to change CustomSACPolicy to MlpPolicy but with policy_kwargs="dict(layers=[256,256])"

[PierreExeter]

Ok thanks for your very quick reply.

[PierreExeter]

Just a doubt, is it OK then to tune the hyperparameters with policy='MlpPolicy' and then to train the model with CustomSACPolicy? Does it not defeat the purpose of tuning in the first place? i.e. would hyperparameters optimised with one policy be also optimal for another policy?

[araffin]

Does it not defeat the purpose of tuning in the first place? i.e. would hyperparameters optimised with one policy be also optimal for another policy?

If in your hyperparameter optimization you allow architecture search:

rl-baselines-zoo/utils/hyperparams_opt.py

Lines 245 to 250 in 645ea17

	net_arch = trial.suggest_categorical('net_arch', ["small", "medium", "big"])
	net_arch = {
	'small': [64, 64],
	'medium': [256, 256],
	'big': [400, 300],

then it does make sense to have policy='MlpPolicy'.
However, if you fix the architecture (by commenting the lines above), then you can use CustomSACPolicy (or in a equivalent way, MlpPolicy + policy_kwargs="dict(layers=[256,256])")

[PierreExeter]

Ok thanks a lot for your help, I'm closing this issue now.