def reshape_decode(self, data, shape):
if self.float_data: # @TODO(tzaman): this is LMDB specific - Make generic!
data = tf.reshape(data, shape)
data = digits.chw_to_hwc(data)
else:
# Decode image of any time option might come: https://github.com/tensorflow/tensorflow/issues/4009
# Distinguish between mime types
if self.data_encoded:
if self.data_mime == 'image/png':
data = tf.image.decode_png(data, dtype=self.image_dtype, name='image_decoder')
elif self.data_mime == 'image/jpeg':
data = tf.image.decode_jpeg(data, name='image_decoder')
else:
logging.error('Unsupported mime type (%s); cannot be decoded' % (self.data_mime))
exit(-1)
else:
if self.backend == 'lmdb':
data = tf.decode_raw(data, self.image_dtype, name='raw_decoder')
# if data is in CHW, set the shape and convert to HWC
if self.unencoded_data_format == 'chw':
data = tf.reshape(data, [shape[0], shape[1], shape[2]])
data = digits.chw_to_hwc(data)
else: # 'hwc'
data = tf.reshape(data, shape)
if (self.channels == 3) and self.unencoded_channel_scheme == 'bgr':
data = digits.bgr_to_rgb(data)
# Convert to float
data = tf.to_float(data)
# data = tf.image.convert_image_dtype(data, tf.float32) # normalize to [0:1) range
return data
def embedding(image_path, detection_method='hog'):
'''
Get an paht to image of SINGLE person and return his embedding
:param image_path: path to SINGLE person image
:param detection_method: face detection model to use: either hog or cnn
'''
# read the image
image = cv.imread(image_path)
if image is None:
print("Cannot load current image")
return None
# convert images to RGB format
rgb_frame = utils.bgr_to_rgb(image)
# face_locations is a list of bounding boxes for each face location in a the image - list of [tuple(top, right, bottom, left).....]
face_locations = fc.face_locations(rgb_frame, model=detection_method)
# embedding is a list of 128-dimensional face encodings.
# NOTE: Each embedding in the list is: <class 'numpy.ndarray'> with shape of (128,) That store <class 'numpy.float64'>
embedding = fc.face_encodings(rgb_frame, face_locations)[0]
return embedding
from utils import bgr_to_rgb
import numpy as np
from utils import get_model, rgb_to_bgr
from utils import get_style_gram_matrices
from utils import ShowCombination
from keras.preprocessing import image
from keras.applications import vgg19
from keras.optimizers import Adam
IMAGE_SIZE_REDUCE_FACTOR = .5
CAMERA_NUMBER = 0
CAMERA = cv2.VideoCapture(0)
frame = bgr_to_rgb(CAMERA.read()[1])
IMG_HEIGHT, IMG_WIDTH, N_CHANNELS = frame.shape
IMG_HEIGHT = int(IMG_HEIGHT * IMAGE_SIZE_REDUCE_FACTOR)
IMG_WIDTH = int(IMG_WIDTH * IMAGE_SIZE_REDUCE_FACTOR)
print((IMG_HEIGHT, IMG_WIDTH, N_CHANNELS))
content_img = cv2.resize(frame, (IMG_WIDTH, IMG_HEIGHT))
style_img = rgb_to_bgr(
np.array(
image.load_img('the_scream.jpg',
target_size=(IMG_HEIGHT, IMG_WIDTH, N_CHANNELS))))
style_gram_matrices = get_style_gram_matrices(style_img)
LEARNING_RATE = 30
DECAY_RATE = 0.01