[Bug] ConfigurationSpace throws error with BOHB

[rohan-gt]

Passing a ConfigurationSpace after setting search_optimization='bohb' in TuneSearchCV throws the following error:

/usr/local/lib/python3.6/dist-packages/tune_sklearn/tune_basesearch.py in fit(self, X, y, groups, **fit_params)
    383                               "To show process output, set verbose=2.")
    384 
--> 385             result = self._fit(X, y, groups, **fit_params)
    386 
    387             if not ray_init and ray.is_initialized():

/usr/local/lib/python3.6/dist-packages/tune_sklearn/tune_basesearch.py in _fit(self, X, y, groups, **fit_params)
    331         config["n_jobs"] = self.sk_n_jobs
    332 
--> 333         self._fill_config_hyperparam(config)
    334         analysis = self._tune_run(config, resources_per_trial)
    335 

/usr/local/lib/python3.6/dist-packages/tune_sklearn/tune_search.py in _fill_config_hyperparam(self, config)
    336         samples = 1
    337         all_lists = True
--> 338         for key, distribution in self.param_distributions.items():
    339             if isinstance(distribution, list):
    340                 import random

AttributeError: 'ConfigurationSpace' object has no attribute 'items'

[rohan-gt]

The _get_bohb_config_space() function doesn't support integer spaces using UniformIntegerHyperparameter

[Yard1]

@rohan-gt Pass a list of CS hyperparameters instead. I'll think about fixing this later

[arainboldt]

Getting the same error using tune-sklearn version 0.2.1

Param Space: 
Configuration space object:
  Hyperparameters:
    boosting_type, Type: Categorical, Choices: {gbdt, dart, rf}, Default: gbdt
    colsample_bytree, Type: UniformFloat, Range: [0.3, 0.7], Default: 0.5
    learning_rate, Type: UniformFloat, Range: [1e-10, 0.1], Default: 0.05
    max_depth, Type: UniformInteger, Range: [1, 10], Default: 6
    n_estimators, Type: UniformInteger, Range: [10, 200], Default: 105
    num_leaves, Type: UniformInteger, Range: [10, 100], Default: 55
    reg_alpha, Type: UniformFloat, Range: [0.0, 1.0], Default: 0.5
    reg_lambda, Type: UniformFloat, Range: [0.0, 1.0], Default: 0.5
    subsample, Type: UniformFloat, Range: [0.3, 0.8], Default: 0.55
    subsample_freq, Type: UniformInteger, Range: [3, 7], Default: 5

---------------------------------------------------------------------------
AttributeError                            Traceback (most recent call last)
<ipython-input-42-5be89034031c> in <module>
     35 base_est = LGBMClassifier(class_weight='balanced',random_state=42,n_jobs=-2,silent=True,objective='multiclass')
     36 
---> 37 model = param_search(base_est,gen_config_space(lgbm_config_dists),train_data,y)
     38 full_glb_preds = pd.DataFrame(model.predict_proba(pred_data),index=pred_data.index )\
     39                              .rename(codes_to_labels,axis=1)

<ipython-input-41-35f8c014e1bd> in param_search(estimator, param_space, X, y, scorer, n_iters, n_folds, search_method, model_name)
     93 
     94 
---> 95     param_model.fit(X,y)
     96 
     97     print(f'Best Model Score: {param_model.score(X,y)}')

~/lib/python3.8/site-packages/tune_sklearn/tune_basesearch.py in fit(self, X, y, groups, **fit_params)
    662                                     "To show process output, set verbose=2.")
    663 
--> 664             result = self._fit(X, y, groups, **fit_params)
    665 
    666             if not ray_init and ray.is_initialized():

~/lib/python3.8/site-packages/tune_sklearn/tune_basesearch.py in _fit(self, X, y, groups, **fit_params)
    563 
    564         self._fill_config_hyperparam(config)
--> 565         analysis = self._tune_run(config, resources_per_trial)
    566 
    567         self.cv_results_ = self._format_results(self.n_splits, analysis)

~/lib/python3.8/site-packages/tune_sklearn/tune_search.py in _tune_run(self, config, resources_per_trial)
    652             search_space = None
    653             override_search_space = True
--> 654             if self._is_param_distributions_all_tune_domains():
    655                 run_args["config"].update(self.param_distributions)
    656                 override_search_space = False

~/lib/python3.8/site-packages/tune_sklearn/tune_search.py in _is_param_distributions_all_tune_domains(self)
    440     def _is_param_distributions_all_tune_domains(self):
    441         return all(
--> 442             isinstance(v, Domain) for k, v in self.param_distributions.items())
    443 
    444     def _get_bohb_config_space(self):

AttributeError: 'ConfigurationSpace' object has no attribute 'items'

Getting the below error using the suggestion of @Yard1 to pass a list of cs hyperparams instead of the configspace object:

Param Space: [boosting_type, Type: Categorical, Choices: {gbdt, dart, rf}, Default: gbdt, num_leaves, Type: UniformInteger, Range: [10, 100], Default: 55, max_depth, Type: UniformInteger, Range: [1, 10], Default: 6, learning_rate, Type: UniformFloat, Range: [1e-10, 0.1], Default: 0.05, n_estimators, Type: UniformInteger, Range: [10, 200], Default: 105, subsample, Type: UniformFloat, Range: [0.3, 0.8], Default: 0.55, subsample_freq, Type: UniformInteger, Range: [3, 7], Default: 5, colsample_bytree, Type: UniformFloat, Range: [0.3, 0.7], Default: 0.5, reg_alpha, Type: UniformFloat, Range: [0.0, 1.0], Default: 0.5, reg_lambda, Type: UniformFloat, Range: [0.0, 1.0], Default: 0.5]

---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
<ipython-input-38-5be89034031c> in <module>
     35 base_est = LGBMClassifier(class_weight='balanced',random_state=42,n_jobs=-2,silent=True,objective='multiclass')
     36 
---> 37 model = param_search(base_est,gen_config_space(lgbm_config_dists),train_data,y)
     38 full_glb_preds = pd.DataFrame(model.predict_proba(pred_data),index=pred_data.index )\
     39                              .rename(codes_to_labels,axis=1)

<ipython-input-37-e819165f716b> in param_search(estimator, param_space, X, y, scorer, n_iters, n_folds, search_method, model_name)
     71 
     72 
---> 73     param_model = TuneSearchCV(estimator, 
     74                       param_space,
     75                       early_stopping=True,

~/lib/python3.8/site-packages/tune_sklearn/tune_search.py in __init__(self, estimator, param_distributions, early_stopping, n_trials, scoring, n_jobs, refit, cv, verbose, random_state, error_score, return_train_score, local_dir, max_iters, search_optimization, use_gpu, loggers, pipeline_auto_early_stop, stopper, time_budget_s, sk_n_jobs, **search_kwargs)
    323         if isinstance(param_distributions, list):
    324             if search_optimization != "random":
--> 325                 raise ValueError("list of dictionaries for parameters "
    326                                  "is not supported for non-random search")
    327 

ValueError: list of dictionaries for parameters is not supported for non-random search

[Yard1]

@arainboldt Thanks for the report, this will be fixed. In the meantime, can you pass a dictionary of Ray Tune hyperparameters? Also, please update to 0.3.0