Merge branch 'master' into 17-adding-new-datasets

code/bolsonaro/models/kmeans_forest_regressor.py

+from bolsonaro.utils import tqdm_joblib
+
+from sklearn.ensemble import RandomForestRegressor
+from sklearn.metrics import mean_squared_error
+from sklearn.base import BaseEstimator
+from sklearn.cluster import KMeans
+from abc import abstractmethod, ABCMeta
+import numpy as np
+from scipy.stats import mode
+from joblib import Parallel, delayed
+from tqdm import tqdm
+
+
+class KMeansForestRegressor(BaseEstimator, metaclass=ABCMeta):
+    """
+    On extreme pruning of random forest ensembles for ral-time predictive applications', by Khaled Fawagreh, Mohamed Medhat Gaber and Eyad Elyan.
+    """
+
+    def __init__(self, models_parameters, score_metric=mean_squared_error):
+        self._models_parameters = models_parameters
+        self._estimator = RandomForestRegressor(**self._models_parameters.hyperparameters,
+            random_state=self._models_parameters.seed, n_jobs=-1)
+        self._extracted_forest_size = self._models_parameters.extracted_forest_size
+        self._score_metric = score_metric
+
+    @property
+    def models_parameters(self):
+        return self._models_parameters
+
+    def fit(self, X_train, y_train, X_val, y_val):
+        self._estimator.fit(X_train, y_train)
+
+        predictions = list()
+        for tree in self._estimator.estimators_:
+            predictions.append(tree.predict(X_train))
+        predictions = np.array(predictions)
+
+        kmeans = KMeans(n_clusters=self._extracted_forest_size, random_state=self._models_parameters.seed).fit(predictions)
+        labels = np.array(kmeans.labels_)
+
+        # For each cluster select the best tree on the validation set
+        extracted_forest_sizes = list(range(self._extracted_forest_size))
+        with tqdm_joblib(tqdm(total=self._extracted_forest_size, disable=True)) as prune_forest_job_pb:
+            pruned_forest = Parallel(n_jobs=-1)(delayed(self._prune_forest_job)(prune_forest_job_pb,
+                extracted_forest_sizes[i], labels, X_val, y_val, self._score_metric)
+                for i in range(self._extracted_forest_size))
+
+        self._estimator.estimators_ = pruned_forest
+
+    def _prune_forest_job(self, prune_forest_job_pb, c, labels, X_val, y_val, score_metric):
+        index = np.where(labels == c)[0]
+        with tqdm_joblib(tqdm(total=len(index), disable=True)) as cluster_job_pb:
+            cluster = Parallel(n_jobs=-1)(delayed(self._cluster_job)(cluster_job_pb, index[i], X_val, 
+                y_val, score_metric) for i in range(len(index)))
+        best_tree_index = np.argmax(cluster)
+        prune_forest_job_pb.update()
+        return self._estimator.estimators_[index[best_tree_index]]
+
+    def _cluster_job(self, cluster_job_pb, i, X_val, y_val, score_metric):
+        y_val_pred = self._estimator.estimators_[i].predict(X_val)
+        tree_pred = score_metric(y_val, y_val_pred)
+        cluster_job_pb.update()
+        return tree_pred
+
+    def predict(self, X):
+        return self._estimator.predict(X)
+
+    def score(self, X, y):
+        predictions = list()
+        for tree in self._estimator.estimators_:
+            predictions.append(tree.predict(X))
+        predictions = np.array(predictions)
+        mean_predictions = np.mean(predictions, axis=0)
+        score = self._score_metric(mean_predictions, y)
+        return score
+
+    def predict_base_estimator(self, X):
+        return self._estimator.predict(X)

code/bolsonaro/models/model_factory.py

@@ -2,6 +2,7 @@ from bolsonaro.models.omp_forest_classifier import OmpForestBinaryClassifier, Om
 from bolsonaro.models.omp_forest_regressor import OmpForestRegressor
 from bolsonaro.models.model_parameters import ModelParameters
 from bolsonaro.models.similarity_forest_regressor import SimilarityForestRegressor
+from bolsonaro.models.kmeans_forest_regressor import KMeansForestRegressor
 from bolsonaro.data.task import Task
 
 from sklearn.ensemble import RandomForestRegressor, RandomForestClassifier
@@ -22,9 +23,11 @@ class ModelFactory(object):
             elif model_parameters.extraction_strategy == 'random':
                 return RandomForestClassifier(n_estimators=model_parameters.extracted_forest_size,
                     random_state=model_parameters.seed)
-            else:
+            elif model_parameters.extraction_strategy == 'none':
                 return RandomForestClassifier(n_estimators=model_parameters.hyperparameters['n_estimators'],
                     random_state=model_parameters.seed)
+            else:
+                raise ValueError('Invalid extraction strategy')
         elif task == Task.REGRESSION:
             if model_parameters.extraction_strategy == 'omp':
                 return OmpForestRegressor(model_parameters)
@@ -33,15 +36,21 @@ class ModelFactory(object):
                     random_state=model_parameters.seed)
             elif model_parameters.extraction_strategy == 'similarity':
                 return SimilarityForestRegressor(model_parameters)
-            else:
+            elif model_parameters.extraction_strategy == 'kmeans':
+                return KMeansForestRegressor(model_parameters)
+            elif model_parameters.extraction_strategy == 'none':
                 return RandomForestRegressor(n_estimators=model_parameters.hyperparameters['n_estimators'],
                     random_state=model_parameters.seed)
+            else:
+                raise ValueError('Invalid extraction strategy')
         elif task == Task.MULTICLASSIFICATION:
             if model_parameters.extraction_strategy == 'omp':
                 return OmpForestMulticlassClassifier(model_parameters)
             elif model_parameters.extraction_strategy == 'random':
                 return RandomForestClassifier(n_estimators=model_parameters.extracted_forest_size,
                     random_state=model_parameters.seed)
-            else:
+            elif model_parameters.extraction_strategy == 'none':
                 return RandomForestClassifier(n_estimators=model_parameters.hyperparameters['n_estimators'],
                     random_state=model_parameters.seed)
+            else:
+                raise ValueError('Invalid extraction strategy')

code/bolsonaro/models/similarity_forest_regressor.py

@@ -3,6 +3,7 @@ from sklearn.metrics import mean_squared_error
 from sklearn.base import BaseEstimator
 from abc import abstractmethod, ABCMeta
 import numpy as np
+from tqdm import tqdm
 
 
 class SimilarityForestRegressor(BaseEstimator, metaclass=ABCMeta):
@@ -10,56 +11,69 @@ class SimilarityForestRegressor(BaseEstimator, metaclass=ABCMeta):
     https://www.ncbi.nlm.nih.gov/pmc/articles/PMC2822360/
     """
 
-    def __init__(self, models_parameters):
+    def __init__(self, models_parameters, score_metric=mean_squared_error):
         self._models_parameters = models_parameters
-        self._regressor = RandomForestRegressor(n_estimators=self._models_parameters.hyperparameters['n_estimators'],
-            random_state=models_parameters.seed)
+        self._estimator = RandomForestRegressor(**self._models_parameters.hyperparameters,
+            random_state=self._models_parameters.seed, n_jobs=-1)
         self._extracted_forest_size = self._models_parameters.extracted_forest_size
+        self._score_metric = score_metric
 
     @property
     def models_parameters(self):
         return self._models_parameters
 
-    def fit(self, X_train, y_train, X_val, y_val, score_metric=mean_squared_error):
+    def fit(self, X_train, y_train, X_val, y_val):
+        self._estimator.fit(X_train, y_train)
 
-        self._regressor.fit(X_train, y_train)
-
-        y_val_pred = self._regressor.predict(X_val)
-        forest_pred = score_metric(y_val, y_val_pred)
-        forest = self._regressor.estimators_
+        y_val_pred = self._estimator.predict(X_val)
+        forest_pred = self._score_metric(y_val, y_val_pred)
+        forest = self._estimator.estimators_
         selected_trees = list()
-        tree_list = list(self._regressor.estimators_)
+        tree_list = list(self._estimator.estimators_)
+
+        val_scores = list()
+        with tqdm(tree_list) as tree_pred_bar:
+            tree_pred_bar.set_description('[Initial tree predictions]')
+            for tree in tree_pred_bar:
+                val_scores.append(tree.predict(X_val))
+            tree_pred_bar.update(1)
 
-        for _ in range(self._extracted_forest_size):
+        with tqdm(range(self._extracted_forest_size), disable=True) as pruning_forest_bar:
+            pruning_forest_bar.set_description(f'[Pruning forest s={self._extracted_forest_size}]')
+            for i in pruning_forest_bar:
                 best_similarity = 100000
                 found_index = 0
-            for i in range(len(tree_list)):
-                lonely_tree = tree_list[i]
-                del tree_list[i]
-                val_list = list()
-                for tree in tree_list:
-                    val_pred = tree.predict(X_val)
-                    val_list.append(val_pred)
-                val_list = np.array(val_list)
-                val_mean = np.mean(val_list, axis=0)
-                val_score = score_metric(val_mean, y_val)
+                with tqdm(range(len(tree_list)), disable=True) as tree_list_bar:
+                    tree_list_bar.set_description(f'[Tree selection s={self._extracted_forest_size} #{i}]')
+                    for j in tree_list_bar:
+                        lonely_tree = tree_list[j]
+                        del tree_list[j]
+                        val_mean = np.mean(np.asarray(val_scores), axis=0)
+                        val_score = self._score_metric(val_mean, y_val)
                         temp_similarity = abs(forest_pred - val_score)
                         if (temp_similarity < best_similarity):
-                    found_index = i
+                            found_index = j
                             best_similarity = temp_similarity
-                tree_list.insert(i, lonely_tree)
+                        tree_list.insert(j, lonely_tree)
+                        val_scores.insert(j, lonely_tree.predict(X_val))
+                        tree_list_bar.update(1)
                 selected_trees.append(tree_list[found_index])
                 del tree_list[found_index]
+                del val_scores[found_index]
+                pruning_forest_bar.update(1)
 
         pruned_forest = list(set(forest) - set(selected_trees))
-        self._regressor.estimators_ = pruned_forest
+        self._estimator.estimators_ = pruned_forest
 
     def score(self, X, y):
         test_list = list()
-        for mod in self._regressor.estimators_:
+        for mod in self._estimator.estimators_:
             test_pred = mod.predict(X)
             test_list.append(test_pred)
         test_list = np.array(test_list)
         test_mean = np.mean(test_list, axis=0)
-        score = mean_squared_error(test_mean, y)
+        score = self._score_metric(test_mean, y)
         return score
+
+    def predict_base_estimator(self, X):
+        return self._estimator.predict(X)

code/compute_results.py

@@ -400,23 +400,51 @@ if __name__ == "__main__":
             xlabel='Number of trees extracted',
             ylabel=experiments_score_metric,
             title='Loss values of {}\nusing best params of previous stages'.format(args.dataset_name))
+    elif args.stage == 5:
+        # Retreive the extracted forest sizes number used in order to have a base forest axis as long as necessary
+        extracted_forest_sizes_number = retreive_extracted_forest_sizes_number(args.models_dir, args.experiment_ids[1])
 
-        # experiment_weights
-        #Plotter.weight_density(experiment_weights, output_path + os.sep + 'weight_density.png')
+        # base_with_params
+        logger.info('Loading base_with_params experiment scores...')
+        base_with_params_train_scores, base_with_params_dev_scores, base_with_params_test_scores, \
+            base_with_params_experiment_score_metric = \
+            extract_scores_across_seeds_and_forest_size(args.models_dir, args.results_dir, args.experiment_ids[0],
+            extracted_forest_sizes_number)
+        # random_with_params
+        logger.info('Loading random_with_params experiment scores...')
+        random_with_params_train_scores, random_with_params_dev_scores, random_with_params_test_scores, \
+            with_params_extracted_forest_sizes, random_with_params_experiment_score_metric = \
+            extract_scores_across_seeds_and_extracted_forest_sizes(args.models_dir, args.results_dir, args.experiment_ids[1])
+        # omp_with_params
+        logger.info('Loading omp_with_params experiment scores...')
+        omp_with_params_train_scores, omp_with_params_dev_scores, omp_with_params_test_scores, _, \
+            omp_with_params_experiment_score_metric = extract_scores_across_seeds_and_extracted_forest_sizes(
+                args.models_dir, args.results_dir, args.experiment_ids[2])
+        # omp_with_params
+        logger.info('Loading kmeans_with_params experiment scores...')
+        kmeans_with_params_train_scores, kmeans_with_params_dev_scores, kmeans_with_params_test_scores, _, \
+            kmeans_with_params_experiment_score_metric = extract_scores_across_seeds_and_extracted_forest_sizes(
+                args.models_dir, args.results_dir, args.experiment_ids[3])
+        
+        # Sanity check on the metrics retreived
+        if not (base_with_params_experiment_score_metric == random_with_params_experiment_score_metric
+            == omp_with_params_experiment_score_metric == kmeans_with_params_experiment_score_metric):
+            raise ValueError('Score metrics of all experiments must be the same.')
+        experiments_score_metric = base_with_params_experiment_score_metric
+
+        output_path = os.path.join(args.results_dir, args.dataset_name, 'stage5_kmeans')
+        pathlib.Path(output_path).mkdir(parents=True, exist_ok=True)
+
+        Plotter.plot_stage2_losses(
+            file_path=output_path + os.sep + 'losses.png',
+            all_experiment_scores=[base_with_params_test_scores, random_with_params_test_scores, omp_with_params_test_scores,
+                kmeans_with_params_test_scores],
+            all_labels=['base', 'random', 'omp', 'kmeans'],
+            x_value=with_params_extracted_forest_sizes,
+            xlabel='Number of trees extracted',
+            ylabel=experiments_score_metric,
+            title='Loss values of {}\nusing best params of previous stages'.format(args.dataset_name))
     else:
         raise ValueError('This stage number is not supported yet, but it will be!')
 
     logger.info('Done.')
-
-    """
-    TODO:
-    For each dataset:
-    Stage 1) [DONE for california_housing] A figure for the selection of the best base forest model hyperparameters (best vs default/random hyperparams)
-    Stage 2) [DONE for california_housing] A figure for the selection of the best combination of normalization: D normalization vs weights normalization (4 combinations)
-    Stage 3) [DONE for california_housing] A figure for the selection of the most relevant subsets combination: train,dev vs train+dev,train+dev vs train,train+dev
-    Stage 4) A figure to finally compare the perf of our approach using the previous selected
-        parameters vs the baseline vs other papers using different extracted forest size
-        (percentage of the tree size found previously in best hyperparams search) on the abscissa.
-
-    IMPORTANT: Compare experiments that used the same seeds among them (except for stage 1).
-    """

code/train.py

@@ -21,7 +21,7 @@ import numpy as np
 import shutil
 
 
-def process_job(seed, parameters, experiment_id, hyperparameters):
+def seed_job(seed_job_pb, seed, parameters, experiment_id, hyperparameters, verbose):
     """
     Experiment function.
 
@@ -34,7 +34,6 @@ def process_job(seed, parameters, experiment_id, hyperparameters):
     """
     logger = LoggerFactory.create(LOG_PATH, 'training_seed{}_ti{}'.format(
         seed, threading.get_ident()))
-    logger.info('seed={}'.format(seed))
 
     seed_str = str(seed)
     experiment_id_str = str(experiment_id)
@@ -55,13 +54,31 @@ def process_job(seed, parameters, experiment_id, hyperparameters):
     trainer = Trainer(dataset)
 
     if parameters['extraction_strategy'] != 'none':
-        for extracted_forest_size in parameters['extracted_forest_size']:
-            logger.info('extracted_forest_size={}'.format(extracted_forest_size))
-            sub_models_dir = models_dir + os.sep + 'extracted_forest_sizes' + os.sep + str(extracted_forest_size)
-            pathlib.Path(sub_models_dir).mkdir(parents=True, exist_ok=True)
+        with tqdm_joblib(tqdm(total=len(parameters['extracted_forest_size']), disable=not verbose)) as extracted_forest_size_job_pb:
+            Parallel(n_jobs=-1)(delayed(extracted_forest_size_job)(extracted_forest_size_job_pb, parameters['extracted_forest_size'][i],
+                models_dir, seed, parameters, dataset, hyperparameters, experiment_id, trainer)
+                for i in range(len(parameters['extracted_forest_size'])))
+    else:
+        forest_size = hyperparameters['n_estimators']
+        logger.info('Base forest training with fixed forest size of {}'.format(forest_size))
+        sub_models_dir = models_dir + os.sep + 'forest_size' + os.sep + str(forest_size)
 
+        # Check if the result file already exists
+        already_exists = False
+        if os.path.isdir(sub_models_dir):
+            sub_models_dir_files = os.listdir(sub_models_dir)
+            for file_name in sub_models_dir_files:
+                if '.pickle' != os.path.splitext(file_name)[1]:
+                    continue
+                else:
+                    already_exists = os.path.getsize(os.path.join(sub_models_dir, file_name)) > 0
+                    break
+        if already_exists:
+            logger.info('Base forest result already exists. Skipping...')
+        else:
+            pathlib.Path(sub_models_dir).mkdir(parents=True, exist_ok=True)
             model_parameters = ModelParameters(
-                extracted_forest_size=extracted_forest_size,
+                extracted_forest_size=forest_size,
                 normalize_D=parameters['normalize_D'],
                 subsets_used=parameters['subsets_used'],
                 normalize_weights=parameters['normalize_weights'],
@@ -76,14 +93,36 @@ def process_job(seed, parameters, experiment_id, hyperparameters):
             trainer.init(model, subsets_used=parameters['subsets_used'])
             trainer.train(model)
             trainer.compute_results(model, sub_models_dir)
+    logger.info(f'Training done for seed {seed_str}')
+    seed_job_pb.update(1)
+
+def extracted_forest_size_job(extracted_forest_size_job_pb, extracted_forest_size, models_dir,
+    seed, parameters, dataset, hyperparameters, experiment_id, trainer):
+
+    logger = LoggerFactory.create(LOG_PATH, 'training_seed{}_extracted_forest_size{}_ti{}'.format(
+        seed, extracted_forest_size, threading.get_ident()))
+    logger.info('extracted_forest_size={}'.format(extracted_forest_size))
+
+    sub_models_dir = models_dir + os.sep + 'extracted_forest_sizes' + os.sep + str(extracted_forest_size)
+
+    # Check if the result file already exists
+    already_exists = False
+    if os.path.isdir(sub_models_dir):
+        sub_models_dir_files = os.listdir(sub_models_dir)
+        for file_name in sub_models_dir_files:
+            if '.pickle' != os.path.splitext(file_name)[1]:
+                return
             else:
-        forest_size = hyperparameters['n_estimators']
-        logger.info('Base forest training with fixed forest size of {}'.format(forest_size))
-        sub_models_dir = models_dir + os.sep + 'forest_size' + os.sep + str(forest_size)
+                already_exists = os.path.getsize(os.path.join(sub_models_dir, file_name)) > 0
+                break
+    if already_exists:
+        logger.info(f'Extracted forest {extracted_forest_size} result already exists. Skipping...')
+        return
+
     pathlib.Path(sub_models_dir).mkdir(parents=True, exist_ok=True)
 
     model_parameters = ModelParameters(
-            extracted_forest_size=forest_size,
+        extracted_forest_size=extracted_forest_size,
         normalize_D=parameters['normalize_D'],
         subsets_used=parameters['subsets_used'],
         normalize_weights=parameters['normalize_weights'],
@@ -98,7 +137,6 @@ def process_job(seed, parameters, experiment_id, hyperparameters):
     trainer.init(model, subsets_used=parameters['subsets_used'])
     trainer.train(model)
     trainer.compute_results(model, sub_models_dir)
-    logger.info('Training done')
 
 """
 Command lines example for stage 1:
@@ -138,6 +176,7 @@ if __name__ == "__main__":
     DEFAULT_SKIP_BEST_HYPERPARAMS = False
     DEFAULT_JOB_NUMBER = -1
     DEFAULT_EXTRACTION_STRATEGY = 'omp'
+    DEFAULT_OVERWRITE = False
 
     begin_random_seed_range = 1
     end_random_seed_range = 2000
@@ -163,7 +202,8 @@ if __name__ == "__main__":
     parser.add_argument('--skip_best_hyperparams', action='store_true', default=DEFAULT_SKIP_BEST_HYPERPARAMS, help='Do not use the best hyperparameters if there exist.')
     parser.add_argument('--save_experiment_configuration', nargs='+', default=None, help='Save the experiment parameters specified in the command line in a file. Args: {{stage_num}} {{name}}')
     parser.add_argument('--job_number', nargs='?', type=int, default=DEFAULT_JOB_NUMBER, help='Specify the number of job used during the parallelisation across seeds.')
-    parser.add_argument('--extraction_strategy', nargs='?', type=str, default=DEFAULT_EXTRACTION_STRATEGY, help='Specify the strategy to apply to extract the trees from the forest. Either omp, random, none or similarity.')
+    parser.add_argument('--extraction_strategy', nargs='?', type=str, default=DEFAULT_EXTRACTION_STRATEGY, help='Specify the strategy to apply to extract the trees from the forest. Either omp, random, none, similarity, kmeans.')
+    parser.add_argument('--overwrite', action='store_true', default=DEFAULT_OVERWRITE, help='Overwrite the experiment id')
     args = parser.parse_args()
 
     if args.experiment_configuration:
@@ -173,7 +213,7 @@ if __name__ == "__main__":
     else:
         parameters = args.__dict__
 
-    if parameters['extraction_strategy'] not in ['omp', 'random', 'none', 'similarity']:
+    if parameters['extraction_strategy'] not in ['omp', 'random', 'none', 'similarity', 'kmeans']:
         raise ValueError('Specified extraction strategy {} is not supported.'.format(parameters.extraction_strategy))
 
     pathlib.Path(parameters['models_dir']).mkdir(parents=True, exist_ok=True)
@@ -220,6 +260,7 @@ if __name__ == "__main__":
 
     if args.experiment_id:
         experiment_id = args.experiment_id
+        if args.overwrite:
             shutil.rmtree(os.path.join(parameters['models_dir'], str(experiment_id)), ignore_errors=True)
     else:
         # Resolve the next experiment id number (last id + 1)
@@ -255,6 +296,6 @@ if __name__ == "__main__":
             )
 
     # Run as much job as there are seeds
-    with tqdm_joblib(tqdm(total=len(seeds), disable=not args.verbose)) as progress_bar:
-        Parallel(n_jobs=args.job_number)(delayed(process_job)(seeds[i],
-            parameters, experiment_id, hyperparameters) for i in range(len(seeds)))
+    with tqdm_joblib(tqdm(total=len(seeds), disable=not args.verbose)) as seed_job_pb:
+        Parallel(n_jobs=args.job_number)(delayed(seed_job)(seed_job_pb, seeds[i],
+            parameters, experiment_id, hyperparameters, args.verbose) for i in range(len(seeds)))