def balance_data_frame(data_frame: np.ndarray) -> np.ndarray:
"""
Balances the given data frame to the class label with the least occurrence; Label must be at last column.
:param data_frame: must be pre-filtered of NA labels
:return: data frame with equal rows of each class; shuffled.
"""
print("Balancing data frame")
util.start_timer()
labels = np.transpose(data_frame)[-1]
unique_labels, unique_label_counts = np.unique(labels, return_counts=True)
min_count = min(unique_label_counts)
# min_count = 3038
balanced_labelled_data = []
for idx, unique_label in enumerate(unique_labels):
unique_label_data = np.array([
x for x in filter(lambda row: row[-1] == unique_label, data_frame)
])
# Random re-balance if there is excess data of this label
if unique_label_counts[idx] != min_count:
np.random.shuffle(unique_label_data) # in-place shuffle
unique_label_data = np.delete(
unique_label_data,
[x for x in range(unique_label_counts[idx] - min_count)],
axis=0)
balanced_labelled_data.extend(unique_label_data)
# Shuffle data frame before returning
balanced_labelled_data = np.array(balanced_labelled_data)
np.random.shuffle(balanced_labelled_data)
util.end_timer_print_duration()
return balanced_labelled_data
def extract_feature_array_from_product(
file_path: str) -> (np.ndarray, int, int, int, list):
"""
Extract bands from an ESA data product
:param file_path: path to a ESA SNAP dim file
:return: (numpy array of shape (bands, pixels), image width, image height, number of pixels)
"""
print('Extracting feature array from product', file_path)
util.start_timer()
from snappy import ProductIO
p = ProductIO.readProduct(file_path)
bands = [p.getBand(x) for x in p.getBandNames()]
if len(bands) == 0:
raise Exception("No bands found in product")
image_width = bands[0].getRasterWidth()
image_height = bands[0].getRasterHeight()
number_of_pixels = image_width * image_height
feature_array = np.array([
band.readPixels(0, 0, image_width, image_height,
np.zeros(number_of_pixels, np.float32))
for band in bands
])
band_names = [band.getName() for band in bands]
util.end_timer_print_duration()
return feature_array, image_width, image_height, number_of_pixels, band_names
def get_labels_npy(self, image_width: int, number_of_pixels: int) -> []:
print("Label npy generation from ground truth instance")
util.start_timer()
labels = []
for i in range(number_of_pixels):
x = i % image_width
y = i // image_width
labels.append(self.check_pixel_class(x, y).value)
util.end_timer_print_duration()
return labels
def experiment_cross_validated_recursive_feature_elimination(
data_frame, band_names, results_folder):
print("Experiment Started: RFECV")
util.start_timer()
label_array = np.transpose(data_frame)[-1]
feature_data = np.delete(data_frame, 0, axis=1)
feature_data = np.delete(feature_data, -1, axis=1)
sc = StandardScaler()
feature_data = sc.fit_transform(feature_data)
X = feature_data
Y = label_array
lr_model = LogisticRegression(multi_class='ovr', solver='liblinear')
rfecv = RFECV(lr_model, step=1, cv=3)
rfecv = rfecv.fit(X, Y)
util.end_timer_print_duration()
print("Feature Names:", band_names)
print("RFECV RESULTS")
print("Optimal Number of features:", rfecv.n_features_)
print("Selected Features:", rfecv.support_)
print("Feature Ranking:", rfecv.ranking_)
print("Grid Scores:",
(range(1,
len(rfecv.grid_scores_) + 1), rfecv.grid_scores_))
plt.figure()
plt.xlabel("Number of features selected")
plt.ylabel("3-Fold Cross validation score (mean accuracy)")
plt.plot(range(1, len(rfecv.grid_scores_) + 1), rfecv.grid_scores_)
plt.savefig(results_folder + 'rfecv.png')
print("RFE RESULTS ")
for i in range(1, len(band_names) + 1):
lr_model_2 = LogisticRegression(multi_class='ovr', solver='liblinear')
rfe = RFE(lr_model_2, i)
rfe = rfe.fit(X, Y)
print("Selected Number of features:", rfe.n_features_)
print("Selected Features:", rfe.support_)
print("Feature Ranking:", rfe.ranking_)
def get_labels_npy_exclude_some_polygons(self, image_width: int,
number_of_pixels: int,
polygons_to_exclude: [()]) -> []:
print("Label npy generation from ground truth instance")
util.start_timer()
labels = []
for i in range(number_of_pixels):
x = i % image_width
y = i // image_width
try:
pixel_parent_polygon = self.polygon_lookup(x, y)
except Exception:
labels.append(self.check_pixel_class(
x, y).value) # Can just assign na
else:
labels.append(
self.check_pixel_class(x, y).value if polygons_to_exclude.
count(self.polygon_lookup(x, y)) ==
0 else self.pixel_class_enum.__members__['na'].value)
util.end_timer_print_duration()
return labels
def __init__(self, pins_file_path, pixel_class_enum):
print("Generating ground truth instance from pins file:",
pins_file_path, "enum:", pixel_class_enum)
util.start_timer()
self.pixel_class_enum = pixel_class_enum
self.boundaries = {
x: {}
for x in filter(lambda y: y != 'na', pixel_class_enum.__members__)
}
with open(pins_file_path, 'r') as fd:
text = fd.read()
rows = text.split('\n')[
6:] # First 6 rows are comments of the file
for row in rows:
fields = row.split(
'\t') # Each field is separated by a Tab character
if len(fields) != 8:
continue
name_fields = fields[PinsFileRowIndex.name.value].split('_')
class_name = name_fields[0]
polygon_index = name_fields[1]
pin_type = name_fields[2]
if class_name not in list(pixel_class_enum.__members__):
continue
if polygon_index not in self.boundaries[class_name].keys():
self.boundaries[class_name][polygon_index] = {}
self.boundaries[class_name][polygon_index][
'x_min' if pin_type == 'TL' else 'x_max'] = fields[
PinsFileRowIndex.x.value]
self.boundaries[class_name][polygon_index][
'y_min' if pin_type == 'TL' else 'y_max'] = fields[
PinsFileRowIndex.y.value]
util.end_timer_print_duration()
def generate_filtered_data_frame_with_pixel_index_and_labels(
number_of_pixels: int, feature_array: np.ndarray,
label_array: np.ndarray) -> np.ndarray:
"""
Adds a pixel index row, and a label row, before transposing to a data-frame. Then filters unlabelled pixels.
\n(WARNING: PIXEL INDEX MUST BE REMOVED BEFORE USING DATA TO FIT MODEL)
:param number_of_pixels:
:param feature_array:
:param label_array:
:return: data_frame of shape (number of pixels, number of features + 2), 1 extra column for pixel index, 1 for label
"""
print('Generating data frame with pixel index and labels')
util.start_timer()
data_frame = np.insert(feature_array,
0, [x for x in range(number_of_pixels)],
axis=0) # Insert Index Row
data_frame = np.append(data_frame, [label_array], axis=0)
data_frame = np.transpose(data_frame)
data_frame = np.array(
list(
filter(lambda row: row[-1] != RosebelPixelClass3.na.value,
data_frame)))
util.end_timer_print_duration()
return data_frame
def experiment_original_multi_model_lr(data_frame):
# Multi-model Methodology
print("\n@@@@@@@@@@@@@@@@@@@@@@@@\n", "MULTI-MODEL CLASSIFICATION",
"\n@@@@@@@@@@@@@@@@@@@@@@@@\n")
# Generate train, test sets
print("Generating train-test split")
util.start_timer()
data_frame_labels = np.transpose(data_frame)[-1]
data_frame_features_with_pixel_index = np.delete(data_frame, -1, axis=1)
x_train, x_test, y_train, y_test = train_test_split(
data_frame_features_with_pixel_index,
data_frame_labels,
test_size=0.33)
util.end_timer_print_duration()
x_train_scaled, x_test_scaled, fitted_scaler = scale_but_ignore_index_column(
x_train, x_test, StandardScaler)
print("Starting multi-model training")
util.start_timer()
# Forest classifier
x_train_forest_clf = np.copy(x_train_scaled)
y_train_forest_clf = np.copy(y_train)
# Relabel
for idx, original_label in enumerate(y_train_forest_clf):
if original_label != RosebelPixelClass3.forest.value:
# The value used is arbitrary as long as its consistent -- use lower class value for the complement
y_train_forest_clf[idx] = RosebelPixelClass3.na.value
# Re-balance
forest_clf_frame = np.transpose(
np.append(np.transpose(x_train_forest_clf),
np.array([y_train_forest_clf]),
axis=0))
balanced_data_frame = balance_data_frame(forest_clf_frame)
x_train_forest_clf = np.delete(balanced_data_frame, -1, axis=1)
x_train_forest_clf_no_pixel_index = np.delete(x_train_forest_clf,
0,
axis=1)
y_train_forest_clf = np.transpose(balanced_data_frame)[-1]
# Fit forest classifier
forest_lr_classifier = LogisticRegression(multi_class='ovr',
solver='liblinear')
forest_lr_classifier.fit(x_train_forest_clf_no_pixel_index,
y_train_forest_clf)
# Water classifier
x_train_water_clf = np.copy(x_train_scaled)
y_train_water_clf = np.copy(y_train)
# relabel
for idx, original_label in enumerate(y_train_water_clf):
if original_label != RosebelPixelClass3.water.value:
# The value used is arbitrary as long as its consistent -- use lower class number for the complement
y_train_water_clf[idx] = RosebelPixelClass3.na.value
# Re-balance
water_clf_frame = np.transpose(
np.append(np.transpose(x_train_water_clf),
np.array([y_train_water_clf]),
axis=0))
balanced_data_frame = balance_data_frame(water_clf_frame)
x_train_water_clf = np.delete(balanced_data_frame, -1, axis=1)
x_train_water_clf_no_pixel_index = np.delete(x_train_water_clf, 0, axis=1)
y_train_water_clf = np.transpose(balanced_data_frame)[-1]
# Fit water classifier
water_lr_classifier = LogisticRegression(multi_class='ovr',
solver='liblinear')
water_lr_classifier.fit(x_train_water_clf_no_pixel_index,
y_train_water_clf)
util.end_timer_print_duration()
# Use both classifiers to generate a prediction array
combined_predictor_result = [] # use normal list first
x_test_no_pixel_index = np.delete(x_test_scaled, 0, axis=1)
for test_pixel in x_test_no_pixel_index:
forest_predictor_result = forest_lr_classifier.predict([test_pixel])[0]
water_predictor_result = water_lr_classifier.predict([test_pixel])[0]
# Neither forest or water
if (forest_predictor_result != RosebelPixelClass3.forest.value) and (
water_predictor_result != RosebelPixelClass3.water.value):
combined_predictor_result.append(RosebelPixelClass3.mines.value)
# Forest and not water
elif (forest_predictor_result == RosebelPixelClass3.forest.value) and (
water_predictor_result != RosebelPixelClass3.water.value):
combined_predictor_result.append(RosebelPixelClass3.forest.value)
# Water and not forest
elif (water_predictor_result == RosebelPixelClass3.water.value) and (
forest_predictor_result != RosebelPixelClass3.forest.value):
combined_predictor_result.append(RosebelPixelClass3.water.value)
# Conflict -- forest predictor says forest, water predictor says water, choose 1 and we'll be right 50% of the time
else:
combined_predictor_result.append(RosebelPixelClass3.water.value)
# Evaluate our results
accuracy_score = metrics.accuracy_score(
y_test, np.array(combined_predictor_result))
confusion_matrix = metrics.confusion_matrix(y_test,
combined_predictor_result)
print("Accuracy:\n", accuracy_score)
print("\nConfusion Matrix:\n", confusion_matrix)
for row_index in range(confusion_matrix.shape[0]):
# Print class prediction accuracy first
value = confusion_matrix[row_index][row_index]
print(
'\n' + RosebelPixelClass3(row_index +
1).name, # Account for na class in enum
'pixel prediction accuracy: %.2f%%' %
(value * 100 / sum(confusion_matrix[row_index])))
# Print out incorrectness
for col_index in range(confusion_matrix.shape[1]):
if row_index != col_index:
value = confusion_matrix[row_index][col_index]
print(
RosebelPixelClass3(row_index + 1).name,
'pixels mis-predicted as',
RosebelPixelClass3(col_index + 1).name + ': %.2f%%' %
(value * 100 / sum(confusion_matrix[row_index])))
def experiment_original_single_model_lr(result_folder,
gt,
image_width,
enum,
data_frame,
error_discover=False,
save_directory=None):
print("Single Model LR Experiment")
# Generate train, test sets
print("Generating train-test split")
util.start_timer()
data_frame_labels = np.transpose(data_frame)[-1]
data_frame_features_with_pixel_index = np.delete(data_frame, -1, axis=1)
x_train, x_test, y_train, y_test = train_test_split(
data_frame_features_with_pixel_index,
data_frame_labels,
test_size=0.33)
util.end_timer_print_duration()
# SCALING
x_train_scaled, x_test_scaled, fitted_scaler = scale_but_ignore_index_column(
x_train, x_test, StandardScaler)
# x_train_scaled = x_train
# x_test_scaled = x_test
x_train_scaled_no_index = np.delete(x_train_scaled, 0, axis=1)
x_test_scaled_no_index = np.delete(x_test_scaled, 0, axis=1)
# Fit one LR model
print("Fitting Logistic Regression Model with Training Data")
util.start_timer()
lr_model = LogisticRegression(multi_class='ovr', solver='liblinear')
lr_model.fit(x_train_scaled_no_index, y_train)
util.end_timer_print_duration()
# Evaluate model
print("Evaluating Logistic Regression Model with Test Data")
util.start_timer()
accuracy_score = lr_model.score(x_test_scaled_no_index, y_test)
test_predictions = lr_model.predict(x_test_scaled_no_index)
confusion_matrix = metrics.confusion_matrix(y_test, test_predictions)
util.end_timer_print_duration()
# Summarize Results
print("\n@@@@@@@@@@@@@@@@@@@@@@@@\n" + "CLASSIFICATION SUMMARY" +
"\n@@@@@@@@@@@@@@@@@@@@@@@@\n")
util.print_train_test_pixel_summary(y_train, y_test, RosebelPixelClass3)
print("Accuracy:\n" + str(accuracy_score))
print("\nConfusion Matrix:\n", confusion_matrix)
util.print_translate_confusion_matrix(confusion_matrix, RosebelPixelClass3,
lambda x: x + 1)
# Error Discovery
if error_discover:
print("\n@@@@@@@@@@@@@@@@@@@@@@@@\n", "ERROR DISCOVERY",
"\n@@@@@@@@@@@@@@@@@@@@@@@@\n")
util.print_error_discovery(confusion_matrix, enum, lr_model,
x_test_scaled, y_test, gt, image_width)
# Model Persistence TODO pipeline the scaler and predictor into a single estimator
if save_directory is not None:
joblib.dump(lr_model, save_directory + "/lr_model.joblib")
joblib.dump(fitted_scaler, save_directory + "/std_scaler.joblib")
x_train, x_test, y_train, y_test = train_test_split(
data_frame_features_with_pixel_index, data_frame_labels, test_size=0.33)
x_train_scaled, x_test_scaled, fitted_scaler = m.scale_but_ignore_index_column(x_train, x_test, StandardScaler)
x_train_scaled_no_index = np.delete(x_train_scaled, 0, axis=1)
x_test_scaled_no_index = np.delete(x_test_scaled, 0, axis=1)
x_train_no_index = np.delete(x_train, 0, axis=1)
x_test_no_index = np.delete(x_test, 0, axis=1)
util.start_timer()
# print("Fitting Logistic Regression Model with Training Data")
# lr_model = LogisticRegression(multi_class='ovr', solver='liblinear')
# lr_model.fit(x_train_scaled_no_index, y_train)
print("Fitting Random Forest Model with Training Data")
rf_model = RandomForestClassifier(n_estimators=50, max_depth=50, min_samples_leaf=22)
rf_model.fit(x_train_no_index, y_train)
util.end_timer_print_duration()
# # Evaluate model
# print("Evaluating Logistic Regression Model with Test Data")
# util.start_timer()
# accuracy_score = lr_model.score(x_test_scaled_no_index, y_test)
# test_predictions = lr_model.predict(x_test_scaled_no_index)
# confusion_matrix = metrics.confusion_matrix(y_test, test_predictions)
# util.end_timer_print_duration()
#
# # Summarize Results
# print("\n@@@@@@@@@@@@@@@@@@@@@@@@\n" + "CLASSIFICATION SUMMARY" + "\n@@@@@@@@@@@@@@@@@@@@@@@@\n")
# util.print_train_test_pixel_summary(y_train, y_test, RosebelPixelClass3)
# print("Accuracy:\n" + str(accuracy_score))
# print("\nConfusion Matrix:\n", confusion_matrix)
# util.print_translate_confusion_matrix(confusion_matrix, RosebelPixelClass3, lambda x: x + 1)