def save(model, model_name):
print("\nSaving model...")
path = common.MODEL_DIR + model_name
model.save(path)
print("Model saved to " + cf.skyBlue(path))
def show_summary():
print("\n" + cf.skyBlue("Training set"))
print(get_summary_matrix(common.TRAINING_DIR))
print("\n" + cf.salmon("Validation set"))
print(get_summary_matrix(common.VALIDATION_DIR))
print("\n" + cf.lightGreen("Testing set"))
print(get_summary_matrix(common.TESTING_DIR))
def show_settings(batch_size):
print("Classification model training application started.\n")
settings = pd.DataFrame(index=('max_epochs', 'patience', 'batch_size'),
columns=('value', ))
settings['value']['max_epochs'] = MAX_EPOCHS
settings['value']['patience'] = PATIENCE
settings['value']['batch_size'] = batch_size
print(cf.skyBlue("Settings"))
print(settings)
def labeled_value(self, key: str, msg: str, *args: Any, **kwargs: Any):
"""Displays a key-value pair with special formatting.
Args:
key (str): Label that is prepended to the message.
For other arguments, see `_format_msg`.
"""
if self.old_style:
return
self._print(
cf.skyBlue(key) + ": " +
_format_msg(cf.bold(msg), *args, **kwargs))
def main(batch_size, model_name):
show_settings(batch_size)
model = make_model()
training, validation = make_generators(batch_size)
fit(model, training, validation, batch_size)
save(model, model_name)
print("\nFinished!")
print("Run " + cf.skyBlue("classify.py") +
" to test the model and get information about its performance.")
print("More information available with " + cf.orange("Tensorboard") + ".")
def main(src_subdir, verbose, model_name):
common.create_directories()
src_dir = common.DATA_DIR + src_subdir + "/"
model = tensorflow.keras.models.load_model(common.MODEL_DIR + model_name)
# real (index) x predicted (column)
confusion_matrix = pd.DataFrame(np.zeros((4, 4), dtype=np.int32),
index=('down', 'left', 'right', 'up'),
columns=('down', 'left', 'right', 'up'))
classification_matrix = pd.DataFrame(np.zeros((4, 3)),
index=('down', 'left', 'right', 'up'),
columns=('precision', 'recall', 'f1'))
type_matrix = pd.DataFrame(np.zeros((4, 2), dtype=np.int32),
index=('round', 'wide', 'narrow', 'total'),
columns=('correct', 'incorrect'))
images = common.get_files(src_dir)
print("Processing {} file(s) in {}/...\n".format(len(images), src_subdir))
for path, filename in images:
img = cv2.imread(path, cv2.IMREAD_GRAYSCALE)
data = np.reshape(img, (1, ) + common.INPUT_SHAPE)
prediction = model.predict(data)
class_index = np.argmax(prediction)
predicted_class = common.CLASSES[class_index]
real_class, arrow_type = common.arrow_labels(filename)
if verbose and real_class != predicted_class:
print(path)
print("Expected {} but got {}: {}\n".format(
cf.lightGreen(real_class), cf.lightCoral(predicted_class),
str(prediction[0])))
confusion_matrix[predicted_class][real_class] += 1
if real_class == predicted_class:
type_matrix['correct'][arrow_type] += 1
type_matrix['correct']['total'] += 1
else:
type_matrix['incorrect'][arrow_type] += 1
type_matrix['incorrect']['total'] += 1
print("\n" + cf.sandyBrown("Confusion matrix"))
print(confusion_matrix)
classification_matrix['precision'] = confusion_matrix.apply(precision)
classification_matrix['recall'] = confusion_matrix.apply(recall, axis=1)
classification_matrix['f1'] = classification_matrix.apply(f1, axis=1)
print("\n" + cf.skyBlue("Classification summary"))
print(classification_matrix)
type_matrix['accuracy'] = type_matrix.apply(type_accuracy, axis=1)
print("\n" + cf.plum("Accuracy by type"))
print(type_matrix)
print("\nFinished!")
def main(inspection, mode, automatic):
print(" SPACE = approve")
print("OTHER KEYS = skip")
print(" Q = quit\n")
labeled_imgs = common.get_files(common.LABELED_DIR)
approved = 0
for path, filename in labeled_imgs:
arrows = []
display = cv2.imread(path)
height, width, _ = display.shape
# Manually tuned values
search_x, search_y = width // 5 + 35, height // 4
search_width, search_height = SEARCH_REGION_WIDTH, height // 2 - search_y
for _ in range(4):
x0 = search_x
x1 = x0 + search_width
y0 = search_y
y1 = y0 + search_height
img = display[y0:y1, x0:x1]
(cx, cy), arrow_box = process_arrow(img, mode)
search_x += int(cx + ARROW_BOX_DIST - SEARCH_REGION_WIDTH / 2)
search_y += int(cy - SEARCH_REGION_HEIGHT / 2)
search_width = SEARCH_REGION_WIDTH
search_height = SEARCH_REGION_HEIGHT
arrows.append(arrow_box)
if not automatic:
arrow_type, directions, _ = re.split('_', filename)
reference = get_reference_arrows(directions, arrows[0].shape)
cv2.imshow(arrow_type, np.vstack([np.hstack(arrows), reference]))
key = cv2.waitKey()
cv2.destroyAllWindows()
else:
key = APPROVE_KEY
if key == APPROVE_KEY:
if not inspection:
save_arrow_imgs(arrows, filename)
approved += 1
elif key == EXIT_KEY:
break
else:
print("Skipped " + cf.skyBlue(path))
if len(labeled_imgs) > 0:
print("\nApproved {} out of {} images ({}%).\n".format(
approved, len(labeled_imgs), 100 * approved // len(labeled_imgs)))
else:
print("There are no images to preprocess.\n")
show_summary()
print("Finished!")