def parameter_tuning(grid_search: bool = True):
# We need to define the parameter grid in which we will exhaustively search for the best combination
# In order to do so, we need to understand what the available hyperparameters SVC() has.
clf = SVC()
images, labels = utils.load_image_data()
images = preprocess.preprocess(images)
start = time.time()
if grid_search:
search_cv = get_grid_search_cv(clf)
else:
# TODO: implement the `get_randomized_search_cv()` functions above for this part
pass
search_cv.fit(images, labels)
logger.log_info(
"GridSearchCV took %.2f seconds for %d candidate parameter settings." %
(time.time() - start, len(search_cv.cv_results_['params'])))
report(search_cv.cv_results_)
# By default, sklearn will refit the model with the best parameters on the entire dataset
# This refitted model is accessible by calling `grid_search.best_estimator_`
best_clf = search_cv.best_estimator_
best_score = search_cv.best_score_
logger.log_info('Best clf (%.3f validation score):' % best_score)
logger.log_info(best_clf)
# let's save this best classifier for later use
model_path = os.path.join(config.MODELS_SUBDIR,
f'svm_grid_search={grid_search}.clf')
logger.log_info(f'Saving fitted model to {model_path}')
utils.save_binary(best_clf, model_path)
def full_train():
# Load images data
logger.log_info('>>>>>>>>>>>>>>>>Run svm.py')
images, labels = utils.load_image_data()
# Separate
images_train, images_test, y_train, y_test = utils.prepare_data(X=images, y=labels)
# Preprocess training and test messages
prev_time = time.time()
logger.log_info("Start transforming data...")
X_train, X_test = preprocess.process_data(images_train, images_test)
logger.log_info(f'Finished transforming data in {round(time.time() - prev_time, 2)} seconds')
# Fit a linear SVM classifier with all hyperparameters set to their default values
prev_time = time.time()
logger.log_info("Start fitting SVM classifier...")
clf = SVC(kernel='linear')
clf.fit(X_train, y_train)
logger.log_info(f'Finished training in {round(time.time() - prev_time, 2)} seconds')
predictions = clf.predict(X_test)
utils.evaluate_prediction(predictions=predictions, y_test=y_test)
utils.error_analysis(images_test, predictions, y_test)
logger.log_info('>>>>>>>>>>>>>>>>End of svm.py')
def predict_images(grid_search: bool = True):
images, labels = utils.load_image_data()
images_train, images_test, y_train, y_test = utils.prepare_data(X=images,
y=labels)
_, X_test = preprocess.process_data(images_train, images_test)
# assume the model is already trained and saved using the parameter_tuning() step above
model_path = os.path.join(config.MODELS_SUBDIR,
f'svm_grid_search={grid_search}.clf')
clf = utils.load_binary(model_path)
# tp see how our best model generalizes to unseen data
# let's corrupt the test images by some noise such that they are not exactly the same as the ones
# the model has seen already
noise_intensity = 2
noise = numpy.random.normal(0, 1, X_test.shape) * noise_intensity
X_test = np.clip(X_test + noise, 0, 255)
predictions = clf.predict(X_test)
utils.evaluate_prediction(predictions, y_test)
def cross_validation():
# Load images data
images, labels = utils.load_image_data()
# Separate
fold_data = utils.prepare_cross_validation_data(X=images, y=labels)
logger.log_info('>>>>>>>>>>>>>>>>Run svm_cross_validation.py')
accuracies = []
for (i, (images_train, images_test, y_train,
y_test)) in enumerate(fold_data):
logger.log_info(f'Training and evaluating fold {i+1}...')
X_train, X_test = preprocess.process_data(images_train, images_test)
# Fit a linear SVM classifier with all hyperparameters set to their default values
prev_time = time.time()
logger.log_info("Start fitting SVM classifier...")
clf = SVC(kernel='linear')
clf.fit(X_train, y_train)
logger.log_info(
f'Finished training in {round(time.time() - prev_time, 2)} seconds'
)
predictions = clf.predict(X_test)
acc = utils.evaluate_prediction(predictions=predictions, y_test=y_test)
accuracies.append(acc)
logger.log_info(f'mean accuracy: %.3f, min: %.3f, max: %.3f, std: %.3f' %
(np.mean(accuracies), np.min(accuracies),
np.max(accuracies), np.std(accuracies)))
logger.log_info('>>>>>>>>>>>>>>>>End of svm_cross_validation.py')