def face_to_cartoon(img):
num_styles = 1
scale = 1.0
aligned=False
original_shape = (img.shape[1], img.shape[0])
if not aligned:
img, tfm = detect_align(img)
img2 = cv2.warpAffine(img, tfm, original_shape, flags=cv2.WARP_INVERSE_MAP)
img = (img - 127.5) / 128.0
images = np.tile(img[None], [num_styles, 1, 1, 1])
scales = scale * np.ones((num_styles))
styles = np.zeros((num_styles, network.input_style.shape[1]))
output = network.generate_BA(images, scales, 16, styles=styles)
output = 0.5*output + 0.5
mask_size = (int(output[0].shape[0]), int(output[0].shape[1]))
alpha_mask = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, mask_size)
alpha_mask = cv2.blur(alpha_mask.astype(np.float64),(25,25), borderType=cv2.BORDER_CONSTANT)
face = cv2.warpAffine(output[0], tfm, original_shape, flags=cv2.WARP_INVERSE_MAP)
alpha_mask = cv2.warpAffine(alpha_mask, tfm, original_shape, flags=cv2.WARP_INVERSE_MAP)
return face, alpha_mask
def process(network, img, aligned=False, num_styles=5, styles=None, scale=1.):
if not aligned:
img = detect_align(img)
if img is None:
return None
img = (img - 127.5) / 128.0
images = np.tile(img[None], [num_styles, 1, 1, 1])
scales = scale * np.ones((num_styles))
if styles is None:
styles = get_styles(network, num_styles)
output = network.generate_BA(images, scales, 16, styles=styles)
output = 0.5 * output + 0.5
return output
def process(inputs, ctx, **kwargs):
LOG.info("process incoming")
frame, is_streaming = helpers.load_image(inputs, 'input')
if frame is None:
raise RuntimeError("Unable to read frame")
LOG.info("input frame size: {}".format(frame.shape))
network = ctx.global_ctx[0]
styles = ctx.global_ctx[1]
scale = 1.0
frame = detect_align(frame)
LOG.info("aligned frame size: {}".format(frame.shape))
if frame is None:
output = np.zeros((256, 256, 1), dtype="uint8")
LOG.info("no aligned image")
else:
frame = (frame - 127.5) / 128.0
LOG.info("aligned image exists")
images = np.tile(frame[None], [1, 1, 1, 1])
scales = scale * np.ones(1)
output = network.generate_BA(images, scales, 16, styles=styles)
output = 0.5*output + 0.5
output = (output[0] * 256).astype('uint8')
LOG.info("output frame size: {}".format(output.shape))
if not is_streaming:
_, buf = cv2.imencode('.jpg', output[:, :, ::-1])
output = buf.tostring()
return {'output': output}
def generate_cartoon(self, img):
if not self.isAligned:
s = time.time()
img = detect_align(img)
e = time.time()
print("detect time cost ", e - s, " s")
if img is None:
print("no face in img ******")
return
img = (img - 127.5) / 128.0
images = np.tile(img[None], [self.num_styles, 1, 1, 1])
scales = 1.0 * np.ones((self.num_styles))
styles = np.random.normal(
0., 1., (self.num_styles, self.warpGAN.input_style.shape[1].value))
start = time.time()
output = self.warpGAN.generate_BA(images, scales, 16, styles=styles)
output = 0.5 * output + 0.5
end = time.time()
print("generate caricatue time cost: ", end - start, " s.")
return output
type=float,
default=1.0)
parser.add_argument("--aligned",
help="Set true if the input face is already normalized",
action='store_true')
args = parser.parse_args()
if __name__ == '__main__':
network = WarpGAN()
network.load_model(args.model_dir)
img = misc.imread(args.input, mode='RGB')
if not args.aligned:
from align.detect_align import detect_align
img = detect_align(img)
img = (img - 127.5) / 128.0
images = np.tile(img[None], [args.num_styles, 1, 1, 1])
scales = args.scale * np.ones((args.num_styles))
styles = np.random.normal(
0., 1., (args.num_styles, network.input_style.shape[1].value))
output = network.generate_BA(images, scales, 16, styles=styles)
output = 0.5 * output + 0.5
for i in range(args.num_styles):
misc.imsave(args.output + '_{}.jpg'.format(i), output[i])
