# Trainable class from SACConfig.build()?
**URL:** https://discuss.ray.io/t/trainable-class-from-sacconfig-build/9249
**Category:** Ray Tune stopping condition & comparisons
**Created:** [February 5, 2023, 5:50am UTC](https://discuss.ray.io/t/trainable-class-from-sacconfig-build/9249 "2023-02-05T05:50:13Z")
**Posts on this page:** 3
**Page:** 1
### Author:  [@Teenforever](https://discuss.ray.io/u/Teenforever)
#### Post date: [February 5, 2023, 5:50am UTC](https://discuss.ray.io/t/trainable-class-from-sacconfig-build/9249/1 "2023-02-05T05:50:13Z")
I am in the process of migrating a Tensorflow SAC trainer written from scratch to RLLIB, but struggling with some ideas in RLLIB. I try to use Ray Tune for a SAC algorithm and plugged in some config parameters like:
```auto
register_env("Env", Env)
sac_config = SACConfig().environment("Env")
sac_config = sac_config.training(
clip_actions = True,
gamma = 0.99,
optimization_config = {
"actor_learning_rate": 0.01,
"critic_learning_rate": 0.01,
"entropy_learning_rate": 0.01,
},
policy_model_config = policy_model_config,
q_model_config = q_model_config,
store_buffer_in_checkpoints = True,
target_network_update_freq = 1,
tau = 0.005,
train_batch_size = 256,
twin_q = True,
)
```
the I run:
`sac = sac_config.build()`
I first thought `sac` would be a trainable class that Tune can take in under trainable class API, since build function in Algorithm\_config ([ray/algorithm\_config.py at master · ray-project/ray · GitHub](https://github.com/ray-project/ray/blob/master/rllib/algorithms/algorithm_config.py)) spits out an Algorithm, and Algorithm seems to be a subclass of Trainable ([ray/algorithm.py at master · ray-project/ray · GitHub](https://github.com/ray-project/ray/blob/master/rllib/algorithms/algorithm.py)). So as the next step I start Tune:
```auto
tuner = tune.Tuner(sac,
param_space = param_space, #for now, param_space is just a grid search of tau
tune_config = TuneConfig(
scheduler = ASHAScheduler(metric = "agg_reward", mode = "max"), #agg_reward is reported in step of environment
num_samples = 1,
),
run_config = RunConfig(
stop = {"training_iteration": 100},
local_dir = ospath.join(ospath.dirname( __file__ ),"ray_results"),
verbose = 1,
log_to_file = True,
checkpoint_config = CheckpointConfig(
checkpoint_frequency = 10
),
),
)
```
Then the error pops up:
` raise TuneError("Improper 'run' - not string nor trainable.")`
If I register sac as a trainable via `register_trainable("sac_trainable", sac())`, it says `TypeError: 'SAC' object is not callable`.
I suspect Tune is telling me that `sac` is not a trainable class, but if not, what is the type of object that `build` blurts out? Could someone please point out what went wrong in the above and how I can make Tune work? Thank you!
---
### Author:  [@mannyv](https://discuss.ray.io/u/mannyv)
#### Post date: [February 5, 2023, 12:23pm UTC](https://discuss.ray.io/t/trainable-class-from-sacconfig-build/9249/2 "2023-02-05T12:23:14Z")
Hi @Teenforever
> [@Teenforever](#):
>
> `register_trainable("sac_trainable", sac())`,
I think this is telling you it should look like this:
`register_trainable("sac_trainable", sac)`
In your case I think the more canonical approach would be to put this in the sac\_config:
`tau=tune.grid_search([...])`
Do not build but run like this:
```auto
tuner = tune.Tuner("SAC",
param_space = sac_config,
...
```
---
### Author:  [@Teenforever](https://discuss.ray.io/u/Teenforever)
#### Post date: [February 6, 2023, 1:55pm UTC](https://discuss.ray.io/t/trainable-class-from-sacconfig-build/9249/3 "2023-02-06T13:55:36Z")
Works like a breeze, thanks @mannyv .
For whoever in the future wondering if it’s necessary to convert the sac\_config to a dict using .dict() - I tried and this isn’t necessary. In fact, if one tries to convert the config to a dict and if the config contains gridsearch (and probably other tune instructions), it blurts out an error.