def flow(self, mode='train'):
while True:
if mode == 'train':
print self.train_keys[0]
# shuffle(self.train_keys)
keys = self.train_keys
elif mode == 'val' or mode == 'demo':
shuffle(self.validation_keys)
keys = self.validation_keys
else:
raise Exception('invalid mode: %s' % mode)
inputs = []
targets = []
for key in keys:
image_path = self.path_prefix + key
if not os.path.exists(image_path):
continue
image_array = imread(image_path)
image_array = imresize(image_array, self.image_size)
num_image_channels = len(image_array.shape)
if num_image_channels != 3:
continue
ground_truth = self.ground_truth_data[key]
if self.do_random_crop:
image_array = self._do_random_crop(image_array)
image_array = image_array.astype('float32')
if mode == 'train' or mode == 'demo':
if self.ground_truth_transformer != None:
image_array, ground_truth = self.transform(
image_array, ground_truth)
ground_truth = (self.ground_truth_transformer.
assign_boxes(ground_truth))
else:
image_array = self.transform(image_array)[0]
if self.grayscale:
image_array = cv2.cvtColor(
image_array.astype('uint8'),
cv2.COLOR_RGB2GRAY).astype('float32')
image_array = np.expand_dims(image_array, -1)
inputs.append(image_array)
targets.append(ground_truth)
if len(inputs) == self.batch_size:
#print len(inputs)
inputs = np.asarray(inputs)
targets = np.asarray(targets)
#print targets
targets = to_categorical(targets, self.classes)
if mode == 'train' or mode == 'val':
inputs = self.preprocess_images(inputs)
#print inputs.shape
#print targets.shape
image_data = {'input': inputs}
targets = {'label': targets}
yield [image_data, targets]
inputs = []
targets = []
def flow(self, mode='train'):
while True:
if mode == 'train':
shuffle(self.train_keys)
keys = self.train_keys
elif mode == 'val' or mode == 'demo':
shuffle(self.validation_keys)
keys = self.validation_keys
else:
raise Exception('invalid mode: %s' % mode)
inputs = []
targets = []
for key in keys:
image_path = self.path_prefix + key
image_array = imread(image_path)
image_array = imresize(image_array, self.image_size)
num_image_channels = len(image_array.shape)
if num_image_channels != 3:
continue
ground_truth = self.ground_truth_data[key]
image_array = image_array.astype('float32')
if mode == 'train' or mode == 'demo':
image_array = self.transform(image_array)[0]
if self.grayscale:
image_array = cv2.cvtColor(
image_array.astype('uint8'),
cv2.COLOR_RGB2GRAY).astype('float32')
image_array = np.expand_dims(image_array, -1)
inputs.append(image_array)
targets.append(ground_truth)
if len(targets) == self.batch_size:
inputs = np.asarray(inputs)
targets = np.asarray(targets)
gender = np_utils.to_categorical(targets[:, 0], 2)
# Quantizing the age
age_bins = np.linspace(0, 100, 21)
age_step = np.digitize(targets[:, 1], age_bins)
age_quantized = np_utils.to_categorical(age_step, 21)
if mode == 'train' or mode == 'val':
inputs = self.preprocess_images(inputs)
yield self._wrap_in_dictionary(inputs, gender,
age_quantized)
if mode == 'demo':
yield self._wrap_in_dictionary(inputs, gender,
age_quantized)
inputs = []
targets = []
def flow(self, mode='train'):
while True:
if mode == 'train':
shuffle(self.train_keys)
keys = self.train_keys
elif mode == 'val':
keys = self.validation_keys
else:
raise Exception('invalid mode: %s' % mode)
inputs = []
targets = []
for key in keys:
image_path = os.path.join(self.path_prefix, key)
image_array = imread(image_path)
image_array = imresize(image_array, self.image_size)
num_image_channels = len(image_array.shape)
if num_image_channels != 3:
continue
ground_truth = self.ground_truth_data[key]
image_array = image_array.astype('float32')
if mode == 'train':
image_array = self.transform(image_array)[0]
if self.grayscale:
image_array = cv2.cvtColor(
image_array.astype('uint8'),
cv2.COLOR_RGB2GRAY).astype('float32')
image_array = np.expand_dims(image_array, -1)
inputs.append(image_array)
targets.append(ground_truth)
if len(targets) == self.batch_size:
inputs = np.asarray(inputs)
targets = np.asarray(targets)
# gender = np_utils.to_categorical(targets[:, 0], 2) # softmax
gender = targets[:, 0] # sigmoid
age = np_utils.to_categorical(targets[:, 1], self.bins)
emotion = np_utils.to_categorical(targets[:, 2], 7)
if mode == 'train' or mode == 'val':
inputs = self.preprocess_images(inputs)
yield self._wrap_in_dictionary(inputs, gender, age,
emotion)
inputs = []
targets = []
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--model', type=str, help='Save model path')
parser.add_argument('--epsilon',
type=float,
help='Epsilon for adversarial perturbation')
args = parser.parse_args()
input_shape = (IMG_SIZE, IMG_SIZE, 3)
alpha = 1
n_age_bins = 21
model = MobileNetDeepEstimator(input_shape[0],
alpha,
n_age_bins,
weights='imagenet')()
model.load_weights(args.model)
image_path, gender_one_hot, age_quantized = random_image()
image = imread(image_path)
image = preprocess_image(image, input_shape)
predictions = predict(model, image)
gender_orig_img, age_orig_img = predictions
x_adv = FGSM(image,
np.expand_dims(gender_one_hot, axis=0),
model,
eps=args.epsilon)
predictions = predict(model, x_adv)
gender_adv_img, age_adv_img = predictions
age_result = 'Actual age bin: {}, Predicted age bin for original image: {}, Predicted age bin for adversarial image: {}'
gender_result = 'Actual gender: {}, Predicted gender for original image: {}, Predicted gender for adversarial image: {}'
print(
age_result.format(np.argmax(age_quantized), np.argmax(age_orig_img),
np.argmax(age_adv_img)))
print(
gender_result.format(np.argmax(gender_one_hot),
np.argmax(gender_orig_img),
np.argmax(gender_adv_img)))
plot_adversarial(image, x_adv)
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--model', type=str, help='Save model path')
parser.add_argument('--epsilon', type=float, help='Epsilon for adversarial perturbation')
args = parser.parse_args()
input_shape = (IM_WIDTH, IM_HEIGHT, 3)
model = load_model(args.model)
img_path, age, gender_one_hot, race_one_hot = random_image()
print(age)
print(gender_one_hot)
print(race_one_hot)
image = imread(img_path)
image = preprocess_image(image, input_shape)
predictions = predict(model, image)
age_pred, race_pred, gender_pred = predictions
race_pred, gender_pred = race_pred.argmax(axis=-1), gender_pred.argmax(axis=-1)
print(age_pred)
print(gender_pred)
print(race_pred)
x_adv1, x_adv2, x_adv3, x_adv4, x_adv5 = FGSM(image, np.expand_dims(race_one_hot, axis=0), model, eps=args.epsilon)
predictions = predict(model, x_adv1)
age_adv_img, race_adv_img, gender_adv_img = predictions
race_adv_img, gender_adv_img = race_adv_img.argmax(axis=-1), gender_adv_img.argmax(axis=-1)
print(age_adv_img)
print(gender_adv_img)
print(race_adv_img)
predictions = predict(model, x_adv5)
age_adv_img, race_adv_img, gender_adv_img = predictions
race_adv_img, gender_adv_img = race_adv_img.argmax(axis=-1), gender_adv_img.argmax(axis=-1)
print(age_adv_img)
print(gender_adv_img)
print(race_adv_img)
img_list = [image, x_adv1, x_adv2, x_adv3, x_adv4, x_adv5]
plot_adversarial(img_list)