def training(grid_param, X_train, X_test, y_train, y_test):
#Create cluster/client
cluster = make_cluster()
cluster
client = Client(cluster)
client
#Construct Dask DataFrame
X_train = dd.from_pandas(X_train, npartitions=4)
y_train = dd.from_pandas(y_train, npartitions=4)
X_test = dd.from_pandas(X_test, npartitions=4)
y_test = dd.from_pandas(y_test, npartitions=4)
estimator = RandomForestRegressor()
#Train model
train_time = time.time()
grid_search = GridSearchCV_dask(estimator, grid_param, cv=2, n_jobs=-1)
with joblib.parallel_backend("dask", scatter=[X_train, y_train]):
grid_search.fit(X_train, y_train)
grid_search.score(X_test, y_test)
train_time = time.time() - train_time
#Predictions
acc_r2 = grid_search.best_estimator_.score(X_test, y_test)
acc_mse = mean_squared_error(grid_search.best_estimator_.predict(X_test), y_test)
return acc_r2, acc_mse, train_time
def RandomForestDask(param_grid, X_train, X_test, y_train, y_test):
cluster = make_cluster()
cluster
client = Client(cluster)
client
dask_X_train = dd.from_pandas(X_train,
npartitions=3) # preprocess data
dask_y_train = dd.from_pandas(y_train, npartitions=3)
dask_X_test = dd.from_pandas(X_test, npartitions=3)
dask_y_test = dd.from_pandas(y_test, npartitions=3)
estimator = RandomForestRegressor()
param_grid = param_grid
grid_search_dask = GridSearchCV_dask(estimator,
param_grid,
cv=2,
n_jobs=-1)
with joblib.parallel_backend("dask",
scatter=[dask_X_train, dask_y_train]):
grid_search_dask.fit(dask_X_train, dask_y_train)
grid_search_dask.score(dask_X_test, dask_y_test)
r_2 = grid_search_dask.best_estimator_.score(dask_X_test, dask_y_test)
mse = mean_squared_error(
grid_search_dask.best_estimator_.predict(X_test), y_test)
return r_2, mse,
def train_evaluate(job_dir, training_dataset_path, search_space, scoring_measure):
"""Runs the training pipeline."""
# Load and prepare training data
df_train = pd.read_csv(training_dataset_path)
num_features_type_map = (
{feature: 'float64' for feature in df_train.columns[NUMERIC_FEATURE_INDEXES]})
df_train = df_train.astype(num_features_type_map)
X_train = df_train.drop('Cover_Type', axis=1)
y_train = df_train['Cover_Type']
# Define the training pipeline
preprocessor = ColumnTransformer(
transformers=[
('num', StandardScaler(), NUMERIC_FEATURE_INDEXES),
('cat', OneHotEncoder(), CATEGORICAL_FEATURE_INDEXES)
])
pipeline = Pipeline([
('preprocessor', preprocessor),
('classifier', SGDClassifier())
])
# Configure hyperparameter tuning
grid = GridSearchCV(pipeline, cv=5,
param_grid=search_space,
scoring=scoring_measure)
# Start training
grid.fit(X_train, y_train)
return grid
def run(self):
self.load_data()
self.split_data()
# nulls = X_train.isnull().sum()
# total_nulls = nulls.sum()
# if total_nulls > 0:
# with pd.option_context('display.max_rows', None, 'display.max_columns', None):
# print(nulls[nulls > 0], "total: ", total_nulls)
reg = self._create_model()
gs = GridSearchCV(reg,
self.model_desc["params_grid"],
cv=self.model_desc["num_folds"],
n_jobs=-1,
refit=False)
start_time = time.monotonic()
with joblib.parallel_backend("dask",
scatter=[self.X_train, self.y_train]):
gs.fit(self.X_train, self.y_train)
finish_time = time.monotonic()
gs_time = finish_time - start_time
print("Searching for marameters for [{}]".format(
self.model_desc["algorithm_name"]))
print("GridSearchCV time: {}".format(gs_time))
cv_res, best_params, gs_score = _get_best_params_score(gs)
print("GridSearchCV score: {}".format(gs_score))
print("Best params: {}".format(best_params))
start_time = time.monotonic()
regr_best = self._create_model(**best_params).fit(
self.X_train, self.y_train)
finish_time = time.monotonic()
test_time = finish_time - start_time
print("Final training time: {}".format(test_time))
test_score = float(regr_best.score(self.X_test, self.y_test))
print("Test score: {}".format(test_score))
ans = {
"algorithm_name": self.model_desc["algorithm_name"],
"gs_time": gs_time,
"gs_score": gs_score,
"test_time": test_time,
"test_score": test_score
}
if "output_path" in self.model_desc:
with open(self.model_desc["output_path"], "w") as fp:
json.dump(ans, fp)
return ans
def grid_search_hyperparams(self,
model,
data,
feature_names,
hyperparam_grid,
n_leave_out=1):
cv_splits = LeavePGroupsOut(n_groups=n_leave_out).split(
data[feature_names],
np.ravel(data[self.target]),
groups=data['group_id'])
gs_models = GridSearchCV(model,
hyperparam_grid,
cv=cv_splits,
scoring=self.metric,
n_jobs=-1)
gs_models.fit(data[feature_names], np.ravel(data[self.target]))
return gs_models.best_params_
def run(self):
print("Running dataset: " + self.name)
for exp_param in self.param_set:
param = common_param.copy()
param.update(exp_param["xgb_params"])
cv_param = exp_param["search parameter"]
print("Running cv grid for: " + str(param))
grid_param = cv_param
# param["reg_alpha"] = 0.001
# model = xgb.XGBRegressor(**param)
# model.fit(X,y)
# print(model.get_booster().get_dump()[0])
# exit(0)
model = xgb.XGBRegressor(
**param
) if self.objective == "reg:linear" else xgb.XGBClassifier(**param)
clf = GridSearchCV(model, grid_param, cv=5, n_jobs=-1)
# with parallel_backend('dask'):
clf.fit(self.X_train, self.y_train)
# clf.fit(X_train, y_train)
param.update(clf.best_params_)
model = xgb.XGBRegressor(
**param
) if self.objective == "reg:linear" else xgb.XGBClassifier(**param)
model.fit(self.X_train, self.y_train)
pred = model.predict(self.X_test)
if isinstance(model, xgb.XGBRegressor):
score = np.sqrt(metrics.mean_squared_error(self.y_test, pred))
else:
score = 1.0 - metrics.accuracy_score(self.y_test, pred)
df_cv_results.at[exp_param['name'],
(self.name,
self.metric)] = "{0:.4g}".format(score)
best_param_value = list(clf.best_params_.values())[0]
best_param_string = "{0:.4g}".format(
best_param_value
) if best_param_value > 0.1 or best_param_value == 0.0 else "{0:.4e}".format(
best_param_value)
df_cv_results.at[exp_param['name'],
(self.name, "param")] = best_param_string
print(df_cv_results.to_latex())
def do_ada_tuning(X,
y,
model,
grid,
n_cores,
random_state=123,
save_path=None):
start_time = timeit.default_timer()
print('Finding best params with 10-fold CV')
rf = DaskGridSearchCV(AdaBoostClassifier(random_state=random_state),
param_grid=grid,
n_jobs=n_cores,
cv=10)
rf.fit(X, y)
means = rf.cv_results_['mean_test_score']
stds = rf.cv_results_['std_test_score']
for mean, std, params in zip(means, stds, rf.cv_results_['params']):
print("%0.3f (+/-%0.03f) for %r" % (mean, std * 2, params))
elapsed_time = timeit.default_timer() - start_time
print('CV time: ' + str(elapsed_time))
print()
best_params = rf.best_params_
best_params.update({'score': rf.best_score_})
forest_performance = {
'score': rf.best_score_,
'fitting_time': elapsed_time
}
print(best_params)
best_params.update({'base_estimator': str(best_params['base_estimator'])})
save_tuning_results(save_path,
random_state,
best_params,
forest_performance,
model=model)
return (best_params, forest_performance)
def do_gbm_tuning(X,
y,
model,
grid,
n_cores,
random_state=123,
save_path=None,
verbose=True):
start_time = timeit.default_timer()
o_print('Finding best params with 10-fold CV', verbose)
rf = DaskGridSearchCV(
GradientBoostingClassifier(random_state=random_state),
param_grid=grid,
n_jobs=n_cores,
cv=10)
rf.fit(X, y)
means = rf.cv_results_['mean_test_score']
stds = rf.cv_results_['std_test_score']
for mean, std, params in zip(means, stds, rf.cv_results_['params']):
o_print("%0.3f (+/-%0.03f) for %r" % (mean, std * 2, params), verbose)
elapsed_time = timeit.default_timer() - start_time
o_print('CV time: ' + str(elapsed_time), verbose)
o_print('', verbose)
best_params = rf.best_params_
best_params.update({'score': rf.best_score_})
forest_performance = {
'score': rf.best_score_,
'fitting_time': elapsed_time
}
save_tuning_results(save_path,
random_state,
best_params,
forest_performance,
model=model)
return (best_params, forest_performance)
def _fit_and_score(self, X, y, train, test):
X_train = X.iloc[train]
X_test = X.iloc[test]
y_train = y.iloc[train]
y_test = y.iloc[test]
scores_by_params = dict.fromkeys(self.strategy_grid.keys())
for i, (strategy,
strategy_values) in enumerate(self.strategy_grid.items()):
print(f'{strategy} ...')
# inner-cv model selection
gscv = DaskGridSearchCV(self.pipeline,
param_grid=strategy_values,
scoring=self.inner_scoring,
cv=self.inner_cv,
refit=True)
gscv.fit(X_train, y_train)
# outer-cv model evaluation
scores_by_params[strategy] = self._score(gscv, X_test, y_test)
return scores_by_params
def subcount_forecast(data, feature):
"""
Creates a new a column that is the predicted value of the input feature
Essentially an abstraction for 'prediction_forecasts'
:param data: a pandas dataframe where each row is an hour
:param feature: a String containing the feature that should be forecasted (one of: casual, registered)
:return: a pandas dataframe containing the new column
"""
var_name = feature + "_forecast"
print("\tAdding {} variable...".format(var_name))
df = dd.get_dummies(data.copy().drop("cnt", axis=1))
to_predict = dd.read_csv(PATH)[feature]
df[feature] = to_predict
train = get_train(df)
model = RandomForestRegressor(random_state=SEED)
model_params = {"n_estimators": list(range(10, 110, 10))}
#tscv = TimeSeriesSplit(n_splits=5)
grid_search = GridSearchCV(estimator=model,
param_grid=model_params,
scoring="r2",
cv=None,
refit=True)
grid_search.fit(train.drop(feature, axis=1), train[feature])
print("\t\tPredictions for GridSearchCV on {}: {:.5f} +/- {:.5f}".format(
feature, grid_search.best_score_,
grid_search.cv_results_["std_test_score"][da.argmax(
grid_search.cv_results_["mean_test_score"])]))
data[var_name] = grid_search.best_estimator_.predict(
dd.get_dummies(data.drop("cnt", axis=1)))
return data
def __fit_and_score(x_train, y_train, x_test, y_test, clf, cv_repeat=None, grid_search_k=None, param_grid=None, grid_scorer=None, scorers=None):
"""Fitting and scoring a classifier through crossvalidated gridsearch over
a parameter grid.
"""
# copy classifier
clf = clone(clf)
# grid search
grid_search_clf = GridSearchCV(
estimator=clf,
param_grid=param_grid,
scoring=scorers,
cv=grid_search_k,
refit=grid_scorer,
)
grid_search_clf = grid_search_clf.fit(x_train, y_train)
# save to results
result = {}
result['gridsearch'] = grid_search_clf.cv_results_
result['best_params'] = grid_search_clf.best_params_
result['scores'] = _multimetric_score(grid_search_clf, x_test, y_test, scorers)
result['cv_repeat'] = cv_repeat
result['clf'] = clf.__class__.__name__
return result
n_sig = y_train[y_train == 1].shape[0]
n_bkg = y_train[y_train == 0].shape[0]
spw = n_bkg / n_sig
n_sig = y[y == 1].shape[0]
n_bkg = y[y == 0].shape[0]
spw = n_bkg / n_sig
print(spw)
search_parameters = {
"learning_rate": [0.02, 0.05, 0.1],
"num_leaves": [20, 50, 150, 200],
"min_child_samples": [40, 60, 100, 160, 240],
"max_depth": [3, 4, 5, 6, 7, 8],
}
clf = lgbm.LGBMClassifier(boosting_type="gbdt",
scale_pos_weight=spw,
n_estimators=1000)
fit_params = {
"early_stopping_rounds": 15,
"eval_metric": "auc",
"eval_set": [(X_test, y_test)],
"eval_sample_weight": [w_test],
}
search = GridSearchCV(clf, param_grid=search_parameters, cv=2)
search.fit(df, y, **fit_params)
def test_gridsearch():
X, y = make_classification(n_samples=100, n_features=5, chunks=50)
grid = {"logisticregression__C": [1000, 100, 10, 2]}
pipe = make_pipeline(DoNothingTransformer(), LogisticRegression())
search = GridSearchCV(pipe, grid, cv=3)
search.fit(X, y)
def rbf_svr_tuning(c = [0.001, 0.01, 0.1, 1, 10], gamma = [0.001, 0.01, 0.1, 1, 10], k = 5,
train_data_path = '../data/training_data.csv', save_model = False, tracking_uri = "http://0.0.0.0:5000"):
# Log the parameters with mlflow
mlflow.log_param("c", c)
mlflow.set_tag("k", k)
# Set random seed for reproducibility
np.random.seed(RANDOM_SEED)
random.seed(RANDOM_SEED)
# Get data shuffled and split into training and test sets
mdr = MiningDataReader(path = train_data_path)
(variable_names, X_train, X_test, y_train, y_test) = mdr.get_splitted_data()
pipeline = Pipeline(steps = [('scaling', StandardScaler()),
('regression', SVR(kernel = 'rbf'))])
### TRAINING ###
################
# Generate grid search for hyperparam tuning
hyperparams = {}
hyperparams['regression__C'] = c
hyperparams['regression__gamma'] = gamma
print("Training started...\n")
# Create an instance of Random Forest Regressor and fit the data for the grid parameters using all processors
modelCV = GridSearchCV(estimator = pipeline,
param_grid = hyperparams,
cv = k,
scoring = 'neg_mean_squared_error',
n_jobs = -1)
with ProgressBar():
modelCV.fit(X_train, y_train)
# Iterate over the results storing training error for each hyperparameter combination
results = modelCV.cv_results_
param_list, training_err_list, training_dev_list = [], [], []
for i in range(len(results['params'])):
param = results['params'][i]
score = (-1) * results['mean_test_score'][i] # NEGATIVE MSE
std = results['std_test_score'][i]
param_list.append(param)
training_err_list.append(score)
training_dev_list.append(std)
print(f"\nBest parameter set found for the training set:\n{modelCV.best_params_}")
# Store the index of the best combination
best_index = param_list.index(modelCV.best_params_)
# Get the best values for hyperparams
best_c = modelCV.best_params_['regression__C']
best_gamma = modelCV.best_params_['regression__gamma']
print("\nTraining finished. Evaluating model...\n")
### EVALUATION ###
##################
# Criteria is C
criteria = 'c'
mlflow.set_tag("criteria", criteria)
param_values = c
# Predict test data variying criteria param and evaluate the models
training_err_by_criteria, training_dev_by_criteria, test_err_list = [], [], []
rmse_score, mae_score, r2_score = -1, -1, -1
feature_names, feature_importances = [], []
for param_value in tqdm(param_values):
model = Pipeline(steps = [('scaler', StandardScaler()),
('regression', SVR(
C = param_value,
gamma = best_gamma,
kernel = 'rbf'))])
param = {'regression__C': param_value, 'regression__gamma': best_gamma}
# Fit model and evaluate results
model.fit(X_train, y_train)
prediction = model.predict(X_test)
index = param_list.index(param)
training_err = training_err_list[index]
training_dev = training_dev_list[index]
(training_mse, test_mse, rmse, mae, r2) = get_test_metrics(training_err, y_test, prediction)
# Store metrics
training_err_by_criteria.append(training_mse)
training_dev_by_criteria.append(training_dev)
test_err_list.append(test_mse)
# Set aditional metrics for the best combination
if index == best_index:
rmse_score = rmse
mae_score = mae
r2_score = r2
# Generate the plots
empty_img_folder()
plot_errors(criteria, param_values, training_err_by_criteria, training_dev_by_criteria, test_err_list)
# Once hyperparameters are selected, train and save the best model
if save_model:
print("\nEvaluation finished. Training final model with train + test data with the best hyperparameters...")
final_model = Pipeline(steps = [('scaler', StandardScaler()),
('regression', SVR(
C = param_list[best_index]['regression__C'],
gamma = best_gamma,
kernel = 'rbf'))])
# Train the best model with all the data (training + test)
full_X = np.vstack((X_train, X_test))
full_y = np.concatenate((y_train, y_test))
final_model.fit(full_X, full_y)
# Log plots and model with mlflow
mlflow.log_artifacts('./img')
mlflow.sklearn.log_model(final_model, 'model')
# Log results with mlflow
mlflow.log_metric("train_mse", training_err_list[best_index])
mlflow.log_metric("test_mse", min(test_err_list))
mlflow.log_metric("rmse", rmse_score)
mlflow.log_metric("mae", mae_score)
mlflow.log_metric("r2", r2_score)
mlflow.set_tag("best_params", param_list[best_index])
# Output the results
print(f'''
-----------------------------------------------------------------------------------------------------------------------
RESULTS
-----------------------------------------------------------------------------------------------------------------------
Best params: {param_list[best_index]}
Training MSE: {training_err_list[best_index]}
Test MSE: {min(test_err_list)}
RMSE: {rmse_score}
MAE: {mae_score}
R2: {r2_score}
-----------------------------------------------------------------------------------------------------------------------
''')
def xgb_reg(n_estimators=100,
max_depth=6,
learning_rate=0.05,
k=5,
train_data_path='../data/training_data.csv',
save_model=False,
tracking_uri="http://0.0.0.0:5000"):
# Log the parameters with mlflow
mlflow.log_param("n_estimators", n_estimators)
mlflow.log_param("max_depth", max_depth)
mlflow.log_param("learning_rate", learning_rate)
mlflow.set_tag("k", k)
# Set random seed for reproducibility
np.random.seed(RANDOM_SEED)
random.seed(RANDOM_SEED)
# Get data shuffled and split into training and test sets
mdr = MiningDataReader(path=train_data_path)
(variable_names, X_train, X_test, y_train,
y_test) = mdr.get_splitted_data()
pipeline = Pipeline(steps=[('scaling', StandardScaler()),
('regression',
xgb.XGBRegressor(objective="reg:squarederror",
seed=RANDOM_SEED))])
### TRAINING ###
################
# Generate grid search for hyperparam tuning
hyperparams = {}
hyperparams['regression__n_estimators'] = np.arange(
n_estimators[0], n_estimators[1], n_estimators[2])
hyperparams['regression__max_depth'] = np.arange(max_depth[0],
max_depth[1],
max_depth[2])
hyperparams['regression__learning_rate'] = learning_rate
print("Training started...\n")
# Create an instance of Random Forest Regressor and fit the data for the grid parameters using all processors
modelCV = GridSearchCV(estimator=pipeline,
param_grid=hyperparams,
cv=k,
scoring='neg_mean_squared_error',
n_jobs=-1)
with ProgressBar():
modelCV.fit(X_train, y_train)
# Iterate over the results storing training error for each hyperparameter combination
results = modelCV.cv_results_
param_list, training_err_list, training_dev_list = [], [], []
for i in range(len(results['params'])):
param = results['params'][i]
score = (-1) * results['mean_test_score'][i] # NEGATIVE MSE
std = results['std_test_score'][i]
param_list.append(param)
training_err_list.append(score)
training_dev_list.append(std)
best_params = modelCV.best_params_
print(f"\nBest parameter set found for the training set:\n{best_params}")
# Store the index of the best combination
best_index = param_list.index(best_params)
# Get the best values for hyperparams
best_n_estimators = best_params['regression__n_estimators']
best_max_depth = best_params['regression__max_depth']
best_learning_rate = best_params['regression__learning_rate']
print("\nTraining finished. Evaluating model...\n")
### EVALUATION ###
##################
# Criteria is C
criteria = 'n_estimators'
mlflow.set_tag("criteria", criteria)
param_values = hyperparams['regression__n_estimators']
# Predict test data variying criteria param and evaluate the models
training_err_by_criteria, training_dev_by_criteria, test_err_list = [], [], []
rmse_score, mae_score, r2_score = -1, -1, -1
feature_names, feature_importances = [], []
for param_value in tqdm(param_values):
model = Pipeline(
steps=[('scaler', StandardScaler()),
('regression',
xgb.XGBRegressor(objective="reg:squarederror",
n_estimators=param_value,
max_depth=best_max_depth,
learning_rate=best_learning_rate))])
param = {
'regression__n_estimators': param_value,
'regression__max_depth': best_max_depth,
'regression__learning_rate': best_learning_rate
}
# Fit model and evaluate results
model.fit(X_train, y_train)
prediction = model.predict(X_test)
index = param_list.index(param)
training_err = training_err_list[index]
training_dev = training_dev_list[index]
(training_mse, test_mse, rmse, mae,
r2) = get_test_metrics(training_err, y_test, prediction)
# Store metrics
training_err_by_criteria.append(training_mse)
training_dev_by_criteria.append(training_dev)
test_err_list.append(test_mse)
# Set aditional metrics for the best combination
if index == best_index:
rmse_score = rmse
mae_score = mae
r2_score = r2
# Generate the plots
empty_img_folder()
plot_errors(criteria, param_values, training_err_by_criteria,
training_dev_by_criteria, test_err_list)
# Once hyperparameters are selected, train and save the best model
if save_model:
print(
"\nEvaluation finished. Training final model with train + test data with the best hyperparameters..."
)
final_model = Pipeline(
steps=[('scaler', StandardScaler()),
('regression',
xgb.XGBRegressor(objective="reg:squarederror",
n_estimators=best_n_estimators,
max_depth=best_max_depth,
learning_rate=best_learning_rate))])
# Train the best model with all the data (training + test)
full_X = np.vstack((X_train, X_test))
full_y = np.concatenate((y_train, y_test))
final_model.fit(full_X, full_y)
# Plot importances and final tree
ax = xgb.plot_importance(final_model.named_steps['regression'])
fig = ax.figure
fig.savefig('./img/importances.png', bbox_inches='tight')
plt.close(fig)
ax = xgb.plot_tree(final_model.named_steps['regression'], rankdir='LR')
fig = ax.figure
fig.set_size_inches(30, 15)
fig.savefig('./img/tree.png', dpi=400, bbox_inches='tight')
plt.close(fig)
# Log plots and model with mlflow
mlflow.log_artifacts('./img')
mlflow.sklearn.log_model(final_model, 'model')
# Log results with mlflow
mlflow.log_metric("train_mse", training_err_list[best_index])
mlflow.log_metric("test_mse", min(test_err_list))
mlflow.log_metric("rmse", rmse_score)
mlflow.log_metric("mae", mae_score)
mlflow.log_metric("r2", r2_score)
mlflow.set_tag("best_params", param_list[best_index])
# Output the results
print(f'''
-----------------------------------------------------------------------------------------------------------------------
RESULTS
-----------------------------------------------------------------------------------------------------------------------
Best params: {param_list[best_index]}
Training MSE: {training_err_list[best_index]}
Test MSE: {min(test_err_list)}
RMSE: {rmse_score}
MAE: {mae_score}
R2: {r2_score}
-----------------------------------------------------------------------------------------------------------------------
''')
print('\n\nNaive Bayes Classifier Score : ', parallel_nb.score(X_test, y_test))
##### OUTPUT --------> Naive Bayes Classifier Score : 0.65
######################################################################################
# Performing GridSearch on the Logistic Regression Classifier
from dask_ml.model_selection import GridSearchCV
parameters = {'penalty': ['l1', 'l2'], 'C': [0.5, 1, 2]}
lr = LogisticRegression()
tuned_lr = GridSearchCV(lr, parameters)
with ProgressBar():
tuned_lr.fit(X_train, y_train)
print('\n\nGrid Search Results for Logistic Regression')
print(pd.DataFrame(tuned_lr.cv_results_)[['params', 'mean_test_score']])
#### OUTPUT
#### Grid Search Results for Logistic Regression
#### params mean_test_score
#### 0 {'C': 0.5, 'penalty': 'l1'} 0.700778
#### 1 {'C': 0.5, 'penalty': 'l2'} 0.700306
#### 2 {'C': 1, 'penalty': 'l1'} 0.700806
#### 3 {'C': 1, 'penalty': 'l2'} 0.700500
#### 4 {'C': 2, 'penalty': 'l1'} 0.700972
#### 5 {'C': 2, 'penalty': 'l2'} 0.700944
# ####################################################################################
def cart_tuning(max_depth=None,
min_samples_leaf=[1, 2, 1],
min_samples_split=[2, 3, 1],
k=5,
train_data_path='../data/training_data.csv',
save_model=False,
tracking_uri="http://0.0.0.0:5000"):
mlflow.log_param('max_depth', max_depth)
mlflow.log_param('min_samples_leaf', min_samples_leaf)
mlflow.log_param('min_samples_split', min_samples_split)
mlflow.set_tag("k", k)
# Set random seed for reproducibility
np.random.seed(RANDOM_SEED)
random.seed(RANDOM_SEED)
# Get data shuffled and split into training and test sets
mdr = MiningDataReader(path=train_data_path)
(variable_names, X_train, X_test, y_train,
y_test) = mdr.get_splitted_data()
pipeline = Pipeline(steps=[('scaler', StandardScaler(
)), ('regression', DecisionTreeRegressor(random_state=RANDOM_SEED))])
### TRAINING ###
################
# Generate grid search for hyperparam tuning
hyperparams = {}
hyperparams['regression__max_depth'] = [
None
] if max_depth is None else np.arange(max_depth[0], max_depth[1],
max_depth[2])
hyperparams['regression__min_samples_leaf'] = np.arange(
min_samples_leaf[0], min_samples_leaf[1], min_samples_leaf[2])
hyperparams['regression__min_samples_split'] = np.arange(
min_samples_split[0], min_samples_split[1], min_samples_split[2])
print("Training started...\n")
# Create an instance of Decision Tree Regressor and fit the data for the grid parameters using all processors
modelCV = GridSearchCV(estimator=pipeline,
param_grid=hyperparams,
cv=k,
scoring='neg_mean_squared_error',
n_jobs=-1)
with ProgressBar():
modelCV.fit(X_train, y_train)
# Iterate over the results storing training error for each hyperparameter combination
results = modelCV.cv_results_
param_list, training_err_list, training_dev_list = [], [], []
for i in range(len(results['params'])):
param = results['params'][i]
score = (-1) * results['mean_test_score'][i] # NEGATIVE MSE
std = results['std_test_score'][i]
param_list.append(param)
training_err_list.append(score)
training_dev_list.append(std)
print(
f"\nBest parameters set found for the training set:\n{modelCV.best_params_}"
)
# Store the index of the best combination
best_index = param_list.index(modelCV.best_params_)
# Get the best values for hyperparams
best_depth = modelCV.best_params_['regression__max_depth']
best_samples_leaf = modelCV.best_params_['regression__min_samples_leaf']
best_samples_split = modelCV.best_params_['regression__min_samples_split']
print("\nTraining finished. Evaluating model...\n")
### EVALUATION ###
##################
# Select the hyperparam with most values as the criteria for the study and calculate test error with the best value
# obtained for the other hyperparameters so the individual effect of this parameter can be studied
criteria = [('max_depth', len(hyperparams['regression__max_depth'])),
('min_samples_leaf',
len(hyperparams['regression__min_samples_leaf'])),
('min_samples_split',
len(hyperparams['regression__min_samples_split']))]
criteria = sorted(criteria, key=lambda x: x[1], reverse=True)[0][0]
mlflow.set_tag("criteria", criteria)
param_values = []
if criteria == 'max_depth':
if max_depth is None:
param_values = [None]
else:
param_values = range(max_depth[0], max_depth[1], max_depth[2])
elif criteria == 'min_samples_leaf':
param_values = range(min_samples_leaf[0], min_samples_leaf[1],
min_samples_leaf[2])
else:
param_values = range(min_samples_split[0], min_samples_split[1],
min_samples_split[2])
# Predict test data variying criteria param and evaluate the models
training_err_by_criteria, training_dev_by_criteria, test_err_list = [], [], []
rmse_score, mae_score, r2_score = -1, -1, -1
feature_names, feature_importances = [], []
for param_value in tqdm(param_values):
if criteria == 'max_depth':
model = Pipeline(steps=[('scaler', StandardScaler()),
('regression',
DecisionTreeRegressor(
max_depth=param_value,
min_samples_leaf=best_samples_leaf,
min_samples_split=best_samples_split,
random_state=RANDOM_SEED))])
param = {
'regression__max_depth': param_value,
'regression__min_samples_leaf': best_samples_leaf,
'regression__min_samples_split': best_samples_split
}
elif criteria == 'min_samples_leaf':
model = Pipeline(steps=[('scaler', StandardScaler()),
('regression',
DecisionTreeRegressor(
max_depth=best_depth,
min_samples_leaf=param_value,
min_samples_split=best_samples_split,
random_state=RANDOM_SEED))])
param = {
'regression__max_depth': best_depth,
'regression__min_samples_leaf': param_value,
'regression__min_samples_split': best_samples_split
}
else:
model = Pipeline(steps=[('scaler', StandardScaler()),
('regression',
DecisionTreeRegressor(
max_depth=best_depth,
min_samples_leaf=best_samples_leaf,
min_samples_split=param_value,
random_state=RANDOM_SEED))])
param = {
'regression__max_depth': best_depth,
'regression__min_samples_leaf': best_samples_leaf,
'regression__min_samples_split': param_value
}
# Fit model and evaluate results
model.fit(X_train, y_train)
prediction = model.predict(X_test)
index = param_list.index(param)
training_err = training_err_list[index]
training_dev = training_dev_list[index]
(training_mse, test_mse, rmse, mae,
r2) = get_test_metrics(training_err, y_test, prediction)
# Store metrics
training_err_by_criteria.append(training_mse)
training_dev_by_criteria.append(training_dev)
test_err_list.append(test_mse)
# Set aditional metrics for the best combination
if index == best_index:
rmse_score = rmse
mae_score = mae
r2_score = r2
# Generate the plots
empty_img_folder()
plot_errors(criteria, param_values, training_err_by_criteria,
training_dev_by_criteria, test_err_list)
# Once hyperparameters are selected, train and save the best model
if save_model:
print(
"\nEvaluation finished. Training final model with train + test data with the best hyperparameters..."
)
final_model = Pipeline(
steps=[('scaler', StandardScaler()),
('regression',
DecisionTreeRegressor(
max_depth=param_list[best_index]
['regression__max_depth'],
min_samples_leaf=param_list[best_index]
['regression__min_samples_leaf'],
min_samples_split=param_list[best_index]
['regression__min_samples_split']))])
# Train the best model with all the data (training + test)
full_X = np.vstack((X_train, X_test))
full_y = np.concatenate((y_train, y_test))
final_model.fit(full_X, full_y)
# Get a barplot with feature importances
feature_importances = final_model.named_steps[
'regression'].feature_importances_
plot_feature_importances(feature_importances, variable_names)
# Create a visual representation of the tree and convert it to PNG
tree_graph = tree.export_graphviz(
final_model.named_steps['regression'],
out_file='/tmp/tree.dot',
max_depth=4)
(graph, ) = pydot.graph_from_dot_file('/tmp/tree.dot')
graph.write_png('./img/tree.png')
# Log plots and model with mlflow
mlflow.log_artifacts('./img')
mlflow.sklearn.log_model(final_model, 'model')
# Log results with mlflow
mlflow.log_metric("training_mse", training_err_list[best_index])
mlflow.log_metric("test_mse", min(test_err_list))
mlflow.log_metric("rmse", rmse_score)
mlflow.log_metric("mae", mae_score)
mlflow.log_metric("r2", r2_score)
mlflow.set_tag("best_params", param_list[best_index])
# Output the results
print(f'''
-----------------------------------------------------------------------------------------------------------------------
RESULTS
-----------------------------------------------------------------------------------------------------------------------
Best params: {param_list[best_index]}
Training MSE: {training_err_list[best_index]}
Test MSE: {min(test_err_list)}
RMSE: {rmse_score}
MAE: {mae_score}
R2: {r2_score}
-----------------------------------------------------------------------------------------------------------------------
''')
def train_model(x_train,
x_test,
y_train,
alphas,
l1_ratios,
n_folds=5,
max_iter=1000):
"""
Build the logic and sklearn pipelines to train x matrix based on input y
Arguments:
x_train - pandas DataFrame of feature matrix for training data
x_test - pandas DataFrame of feature matrix for testing data
y_train - pandas DataFrame of processed y matrix (output from align_matrices())
alphas - list of alphas to perform cross validation over
l1_ratios - list of l1 mixing parameters to perform cross validation over
n_folds - int of how many folds of cross validation to perform
max_iter - the maximum number of iterations to test until convergence
Output:
The full pipeline sklearn object and y matrix predictions for training, testing,
and cross validation
"""
# Setup the classifier parameters
clf_parameters = {
"classify__loss": ["log"],
"classify__penalty": ["elasticnet"],
"classify__alpha": alphas,
"classify__l1_ratio": l1_ratios,
}
estimator = Pipeline(steps=[(
"classify",
SGDClassifier(
random_state=0,
class_weight="balanced",
loss="log",
max_iter=max_iter,
tol=1e-3,
),
)])
cv_pipeline = GridSearchCV(
estimator=estimator,
param_grid=clf_parameters,
n_jobs=-1,
cv=n_folds,
scoring="roc_auc",
return_train_score=True,
)
# Fit the model
cv_pipeline.fit(X=x_train, y=y_train.status)
# Obtain cross validation results
y_cv = cross_val_predict(
cv_pipeline.best_estimator_,
X=x_train,
y=y_train.status,
cv=n_folds,
method="decision_function",
)
# Get all performance results
y_predict_train = cv_pipeline.decision_function(x_train)
y_predict_test = cv_pipeline.decision_function(x_test)
return cv_pipeline, y_predict_train, y_predict_test, y_cv
# modified from https://github.com/amueller/scipy-2018-sklearn/blob/master/notebooks/15.Pipelining_Estimators.ipynb
from pathlib import Path
import pandas as pd
from sklearn.model_selection import train_test_split
from dask_ml.model_selection import GridSearchCV
from dask.distributed import Client
from sklearn.pipeline import make_pipeline
from dask_ml.preprocessing import StandardScaler
from dask_ml.linear_model import LogisticRegression
if __name__ == "__main__":
client = Client()
data = Path('./data')
df = pd.read_csv(data / "01_heights_weights_genders.csv")
y = 1 * (df.Gender == "Male").values
X = df[['Height', 'Weight']].values
X_train, X_test, y_train, y_test = train_test_split(X, y)
pipeline = make_pipeline(StandardScaler(), LogisticRegression())
grid = GridSearchCV(pipeline,
param_grid={'logisticregression__C': [.1, 1, 10, 100]},
cv=5)
grid.fit(X_train, y_train)
print("Score", grid.score(X_test, y_test))
from keras.models import Sequential
from keras.layers import Dense
from keras.wrappers.scikit_learn import KerasClassifier
from sklearn.datasets import load_breast_cancer
from sklearn.model_selection import train_test_split
from dask_ml.model_selection import GridSearchCV
from dask.distributed import Client
from sklearn.externals import joblib
def simple_nn(hidden_neurons):
model = Sequential()
model.add(Dense(hidden_neurons, activation='relu', input_dim=30))
model.add(Dense(1, activation='sigmoid'))
model.compile(loss='binary_crossentropy', optimizer='rmsprop', metrics=['accuracy'])
return model
param_grid = {'hidden_neurons': [100, 200, 300]}
if __name__=='__main__':
client = Client()
cv = GridSearchCV(KerasClassifier(build_fn=simple_nn, epochs=100), param_grid)
X, y = load_breast_cancer(return_X_y=True)
X_train, X_test, y_train, y_test = train_test_split(X, y)
with joblib.parallel_backend("dask", scatter=[X_train, y_train]):
cv.fit(X_train, y_train)
print(f'Best Accuracy for {cv.best_score_:.4} using {cv.best_params_}')
from sklearn.datasets import load_digits
import sklearn.metrics as skmetrics
from sklearn.svm import SVC
run = Run.get_context()
# For a bigger dataset, try
# x, y = make_classification(5000, n_classes=16, n_informative=13)
digits = load_digits()
x = digits.data
y = digits.target
# Obtain cross validation performance for single parameter setting:
param_space = {'C': [1]}
model = SVC(kernel='rbf')
if args.cv:
search = GridSearchCV(model,
param_space,
scoring=['accuracy'],
refit=False,
cv=args.cv)
search.fit(x, y)
run.log('accuracy_mean', search.cv_results_['mean_test_accuracy'][0])
run.log('accuracy_std', search.cv_results_['std_test_accuracy'][0])
else:
x_train, x_test, y_train, y_test = train_test_split(x, y)
model.fit(x_train, y_train)
y_test_pred = model.predict(x_test)
run.log('accuracy', skmetrics.accuracy_score(y_test, y_test_pred))