Merge pull request #25 from denBruneBarone/grid_search_cv

Grid search cv

machine_learning/config.py

@@ -4,6 +4,12 @@ class HPConfig:
  max_features = None # samme som paper
  max_leaf_nodes = 500 # værdien fra paperet om modeller er 10
 
+class BestHPConfig:
+ criterion = None
+ max_depth = None
+ max_features = None
+ max_leaf_nodes = None
+
 class GridSearchConfig:
  param_grid = {
  'criterion': ['friedman_mse', 'squared_error'],

machine_learning/training.py

@@ -1,32 +1,84 @@
 import numpy as np
+from sklearn.model_selection import KFold, GridSearchCV
 from sklearn.tree import DecisionTreeRegressor
-from sklearn.metrics import mean_squared_error
-from machine_learning.config import HPConfig
 from machine_learning.prepare_for_training import TrainingDataset
+from sklearn.metrics import mean_squared_error
+from machine_learning.config import HPConfig, GridSearchConfig
+from sklearn.metrics import make_scorer
 
 
-def train_model(train_data):
- training_dataset = TrainingDataset(train_data)
+def rmse_cum_power(true_labels, predicted_labels):
+ time_diff = np.diff(true_labels[:, 0], prepend=0)
+ true_cumulative_power = np.cumsum(predicted_labels[:, 0] * predicted_labels[:, 1] * time_diff)
+ test_targets_cumulative_power = np.cumsum(true_labels[:, 0] * true_labels[:, 1] * time_diff)
 
- # Instantiate the decision tree model with specified hyperparameters
- model = DecisionTreeRegressor(criterion=HPConfig.criterion, max_depth=HPConfig.max_depth,
- max_features=HPConfig.max_features, max_leaf_nodes=HPConfig.max_leaf_nodes,
- random_state=42)
+ rmse = np.sqrt(mean_squared_error(test_targets_cumulative_power, true_cumulative_power))
+ return rmse
 
- # Extract features and targets from the training dataset
- train_features = []
- train_targets = []
- for index in range(len(training_dataset)):
- input_array, target_array = training_dataset[index]
- train_features.append(input_array)
- train_targets.append(target_array)
 
- # Concatenate the lists along the appropriate axis
- train_features_np = np.concatenate(train_features, axis=0)
- train_targets_np = np.concatenate(train_targets, axis=0)
+def custom_scoring(true_labels, predicted_labels):
+ return rmse_cum_power(true_labels, predicted_labels)
+
+custom_scoring = make_scorer(custom_scoring)
+
+
+def train_model(train_data, grid_search_cv=True):
+ if grid_search_cv:
+ training_dataset = TrainingDataset(train_data)
+
+ # Instantiate the decision tree model with specified hyperparameters
+ model = DecisionTreeRegressor()
+ # splits the train-test data into n_splits number of subsets for cross validation
+ cv = KFold(n_splits=5, shuffle=True, random_state=42) # TODO best n_split?
+ grid_search = GridSearchCV(estimator=model, param_grid=GridSearchConfig.param_grid,
+ cv=cv, scoring=custom_scoring, verbose=2)
+
+ train_features = []
+ train_targets = []
+
+ for index in range(len(training_dataset)):
+ input_array, target_array = training_dataset[index]
+ train_features.append(input_array)
+ train_targets.append(target_array)
 
- # Fit the decision tree model
- model.fit(train_features_np, train_targets_np)
+ # Concatenate the lists along the appropriate axis
+ train_features_np = np.concatenate(train_features, axis=0)
+ train_targets_np = np.concatenate(train_targets, axis=0)
+
+ # Fit the decision tree model
+ grid_search.fit(train_features_np, train_targets_np)
+ best_params = grid_search.best_params_
+ best_score = grid_search.best_score_
+
+ print("Best Params: ", best_params)
+ print("Best score: ", best_score)
+
+ best_dt_model = grid_search.best_estimator_
+ best_dt_model.fit(train_features_np, train_targets_np)
+ # TODO: Tilføj print detaljer
+
+ else:
+ training_dataset = TrainingDataset(train_data)
+
+ # Instantiate the decision tree model with specified hyperparameters
+ model = DecisionTreeRegressor(criterion=HPConfig.criterion, max_depth=HPConfig.max_depth,
+ max_features=HPConfig.max_features, max_leaf_nodes=HPConfig.max_leaf_nodes,
+ random_state=42)
+
+ # Extract features and targets from the training dataset
+ train_features = []
+ train_targets = []
+ for index in range(len(training_dataset)):
+ input_array, target_array = training_dataset[index]
+ train_features.append(input_array)
+ train_targets.append(target_array)
+
+ # Concatenate the lists along the appropriate axis
+ train_features_np = np.concatenate(train_features, axis=0)
+ train_targets_np = np.concatenate(train_targets, axis=0)
+
+ # Fit the decision tree model
+ model.fit(train_features_np, train_targets_np)
 
  return model
 
@@ -47,6 +99,8 @@ def evaluate_model(model, test_data):
  test_features_np = np.concatenate(test_features, axis=0)
  test_targets_np = np.concatenate(test_targets, axis=0)
 
+ model.fit(test_features_np, test_targets_np)
+
  # Predict on the test set
  test_predictions = model.predict(test_features_np)