def process_images(input_path, batch_size=64, crop=True):
model = load_model(
"I:\\pythonProject\\RotNet\\rotnet_models\\rotnet_street_view_resnet50_keras2.hdf5",
custom_objects={'angle_error': angle_error},
compile=False)
extensions = ['.jpg', '.jpeg', '.bmp', '.png']
if os.path.isfile(input_path) or input_path[:4].lower() == "http":
image_paths = [input_path]
else:
image_paths = [
os.path.join(input_path, f) for f in os.listdir(input_path)
if os.path.splitext(f)[1].lower() in extensions
]
predictions = model.predict_generator(RotNetDataGenerator(
image_paths,
input_shape=(224, 224, 3),
batch_size=batch_size,
one_hot=True,
preprocess_func=preprocess_input,
rotate=False,
crop_largest_rect=True,
crop_center=True),
val_samples=len(image_paths))
predicted_angles = np.argmax(predictions, axis=1)
print(predicted_angles)
return predicted_angles
def train(self):
train_generator = RotNetDataGenerator(
self.train_data,
input_shape=self.input_shape,
batch_size=self.batch_size,
preprocess_func=preprocess_input,
crop_center=False,
crop_largest_rect=True,
shuffle=True,
is_hw_ratio=self.is_hw_ratio,
random_angle=self.random_angle,
is_random_crop=self.is_random_crop)
test_generator = RotNetDataGenerator(
self.test_data,
input_shape=self.input_shape,
batch_size=self.batch_size,
preprocess_func=preprocess_input,
crop_center=False,
crop_largest_rect=True,
is_train=False, # 关闭训练参数
is_hw_ratio=self.is_hw_ratio,
random_angle=self.random_angle,
is_random_crop=self.is_random_crop)
steps_per_epoch = min(len(self.train_data) // self.batch_size, 50000)
validation_steps = min(len(self.test_data) // self.batch_size, 5000)
monitor = 'val_acc'
checkpointer = ModelCheckpoint(filepath=os.path.join(
self.output_dir, self.model_name),
monitor=monitor,
save_best_only=True)
n_workers = 3
print('[Info] n_workers: {}'.format(n_workers))
# training loop
self.model.fit(
train_generator,
steps_per_epoch=steps_per_epoch,
epochs=self.nb_epoch,
validation_data=test_generator,
validation_steps=validation_steps,
callbacks=[checkpointer],
workers=n_workers,
)
def process_images(model, input_path, output_path,
batch_size=64, crop=True):
extensions = ['.jpg', '.jpeg', '.bmp', '.png']
output_is_image = False
if os.path.isfile(input_path):
image_paths = [input_path]
if os.path.splitext(output_path)[1].lower() in extensions:
output_is_image = True
output_filename = output_path
output_path = os.path.dirname(output_filename)
else:
image_paths = [os.path.join(input_path, f)
for f in os.listdir(input_path)
if os.path.splitext(f)[1].lower() in extensions]
if os.path.splitext(output_path)[1].lower() in extensions:
print('Output must be a directory!')
predictions = model.predict_generator(
RotNetDataGenerator(
image_paths,
input_shape=(224, 224, 3),
batch_size=64,
one_hot=True,
preprocess_func=preprocess_input,
rotate=False,
crop_largest_rect=True,
crop_center=True
),
val_samples=len(image_paths)
)
predicted_angles = np.argmax(predictions, axis=1)
if output_path == '':
output_path = '.'
if not os.path.exists(output_path):
os.makedirs(output_path)
for path, predicted_angle in zip(image_paths, predicted_angles):
image = cv2.imread(path)
rotated_image = rotate(image, -predicted_angle)
if crop:
size = (image.shape[0], image.shape[1])
rotated_image = crop_largest_rectangle(rotated_image, -predicted_angle, *size)
if not output_is_image:
output_filename = os.path.join(output_path, os.path.basename(path))
cv2.imwrite(output_filename, rotated_image)
def predict(model, input_path, batch_size=64, crop=True):
predictions = model.predict_generator(RotNetDataGenerator(
[input_path],
input_shape=(224, 224, 3),
batch_size=64,
one_hot=True,
preprocess_func=preprocess_input,
rotate=False,
crop_largest_rect=True,
crop_center=True),
val_samples=1)
predicted_angle = np.argmax(predictions, axis=1)[0]
image = cv2.imread(input_path)
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
rotated_image = rotate(image, -predicted_angle)
if crop:
size = (image.shape[0], image.shape[1])
rotated_image = crop_largest_rectangle(rotated_image, -predicted_angle,
*size)
return rotated_image, predicted_angle
# callbacks
monitor = 'val_angle_error'
checkpointer = ModelCheckpoint(filepath=os.path.join(output_folder,
model_name + '.hdf5'),
monitor=monitor,
save_best_only=True)
reduce_lr = ReduceLROnPlateau(monitor=monitor, patience=3)
early_stopping = EarlyStopping(monitor=monitor, patience=5)
tensorboard = TensorBoard()
# training loop
model.fit_generator(
RotNetDataGenerator(train_filenames,
input_shape=input_shape,
batch_size=batch_size,
preprocess_func=preprocess_input,
crop_center=True,
crop_largest_rect=True,
shuffle=True),
steps_per_epoch=len(train_filenames) / batch_size,
epochs=nb_epoch,
validation_data=RotNetDataGenerator(test_filenames,
input_shape=input_shape,
batch_size=batch_size,
preprocess_func=preprocess_input,
crop_center=True,
crop_largest_rect=True),
validation_steps=len(test_filenames) / batch_size,
callbacks=[checkpointer, reduce_lr, early_stopping, tensorboard],
workers=10)
copyfile(os.path.basename(__file__),
output_folder + os.path.basename(__file__))
tensorboard = TensorBoard(log_dir=output_log)
checkpointer = CustomModelCheckpoint(
model_for_saving=model,
filepath=output_weight + "weights_{epoch:02d}_{val_loss:.2f}.h5",
save_best_only=True,
monitor='val_loss',
save_weights_only=True)
generator_training = RotNetDataGenerator(input_shape=input_shape,
batch_size=batch_size,
one_hot=True,
preprocess_func=preprocess_input,
shuffle=True).generate(
paths_train, labels_train,
len(classes_focal))
generator_valid = RotNetDataGenerator(input_shape=input_shape,
batch_size=batch_size,
one_hot=True,
preprocess_func=preprocess_input,
shuffle=True).generate(
paths_valid, labels_valid,
len(classes_focal))
# training loop
model.fit_generator(
generator=generator_training,
steps_per_epoch=(len(paths_train) // batch_size), # 29977
batch_size = 128
nb_epoch = 50
output_folder = 'models'
if not os.path.exists(output_folder):
os.makedirs(output_folder)
# callbacks
checkpointer = ModelCheckpoint(filepath=os.path.join(output_folder,
model_name + '.hdf5'),
save_best_only=True)
early_stopping = EarlyStopping(patience=2)
tensorboard = TensorBoard()
# training loop
model.fit_generator(RotNetDataGenerator(X_train,
batch_size=batch_size,
one_hot=False,
preprocess_func=binarize_images,
shuffle=True),
steps_per_epoch=nb_train_samples / batch_size,
epochs=nb_epoch,
validation_data=RotNetDataGenerator(
X_test,
one_hot=False,
preprocess_func=binarize_images,
batch_size=batch_size),
validation_steps=nb_test_samples / batch_size,
verbose=1,
callbacks=[checkpointer, early_stopping, tensorboard])
# training parameters
batch_size = 128
nb_epoch = 50
output_folder = 'models'
if not os.path.exists(output_folder):
os.makedirs(output_folder)
# callbacks
checkpointer = ModelCheckpoint(filepath=os.path.join(output_folder,
model_name + '.hdf5'),
save_best_only=True)
early_stopping = EarlyStopping(patience=2)
tensorboard = TensorBoard()
# training loop
model.fit_generator(
RotNetDataGenerator(X_train,
batch_size=batch_size,
preprocess_func=binarize_images,
shuffle=True),
steps_per_epoch=nb_train_samples / batch_size,
epochs=nb_epoch,
validation_data=RotNetDataGenerator(X_test,
batch_size=batch_size,
preprocess_func=binarize_images),
validation_steps=nb_test_samples / batch_size,
verbose=1 #,
# callbacks=[checkpointer, early_stopping, tensorboard]
)
def process_images(model, input_path, output_path, batch_size=64, log=False):
extensions = ['.jpg', '.jpeg', '.bmp', '.png']
output_is_image = False
if os.path.isfile(input_path):
image_paths = [input_path]
if os.path.splitext(output_path)[1].lower() in extensions:
output_is_image = True
output_filename = output_path
output_path = os.path.dirname(output_filename)
else:
image_paths = [
os.path.join(input_path, f) for f in os.listdir(input_path)
if os.path.splitext(f)[1].lower() in extensions
]
if os.path.splitext(output_path)[1].lower() in extensions:
print('Output must be a directory!')
predictions = model.predict_generator(
RotNetDataGenerator(image_paths,
input_shape=(224, 224, 3),
batch_size=64,
one_hot=True,
preprocess_func=preprocess_input,
rotate=False,
crop_largest_rect=True,
crop_center=True))
predicted_angles = np.argmax(predictions, axis=1)
rotated_angles = [
] # We will be using this to compile what angles we actually rotate by for logging
if output_path == '':
output_path = '.'
if not os.path.exists(output_path):
os.makedirs(output_path)
for path, predicted_angle in zip(image_paths, predicted_angles):
image = cv2.imread(path)
isLandscape = image.shape[1] >= image.shape[
0] # image.shape = [height, width, channels]
if 45 <= predicted_angle < 315:
"""
Angle rotation logic:
if 45 < pred_angle < 315:
if isLandscape: rotate by +-90
else: rotate by 0/180
"""
if 45 <= predicted_angle and predicted_angle < 135 and not isLandscape:
angle_cv2 = cv2.ROTATE_90_CLOCKWISE
rotated_angle = 90
elif 135 <= predicted_angle and predicted_angle < 225 and isLandscape:
angle_cv2 = cv2.ROTATE_180
rotated_angle = 180
elif 225 <= predicted_angle and predicted_angle < 315 and not isLandscape:
angle_cv2 = cv2.ROTATE_90_COUNTERCLOCKWISE
rotated_angle = 270
else:
rotated_angles.append(0)
continue
rotated_image = cv2.rotate(image, angle_cv2)
rotated_angles.append(rotated_angle)
print("Rotated {} {}°".format(path, rotated_angle))
if not output_is_image:
output_filename = os.path.join(output_path,
os.path.basename(path))
cv2.imwrite(output_filename, rotated_image)
else:
rotated_angles.append(0)
if log:
"""
Create a CSV file and dump the logs in the format
"image_filename, rotation_angle"
"""
with open(
"log-{}.csv".format(
datetime.datetime.now().strftime('%Y%m%dT%H%M%S')),
'a') as f:
f.write("Filename, Angle\n")
for path, rotated_angle in zip(image_paths, rotated_angles):
f.write("{}, {}\n".format(os.path.basename(path),
rotated_angle))