def reconstruct_faces(
self,
out_path,
imgs_to_reconstruct,
transformer=lambda x: img_ops.unbound_image_values(x)):
out_original_path = fs_utils.add_suffix_to_path(out_path, "original")
self.__save_original_images(imgs_to_reconstruct, out_original_path,
transformer)
img_size = imgs_to_reconstruct.shape[1]
sess = tf.InteractiveSession()
with tf.device(device_utils.get_device()):
input_images = tf.placeholder(dtype=tf.float32,
shape=[1, img_size, img_size, 3])
_, _, z_mean, z_cov_log_sq = self.load_enc_with_weights_func(
sess, input_images)
eps = tf.random_normal(shape=(1, self.z_dim), mean=0.0, stddev=1.0)
z = z_mean + tf.multiply(tf.sqrt(tf.exp(z_cov_log_sq)), eps)
gen_out_layer, _ = self.load_gen_with_weights_func(sess, z)
x_reconstr = gen_out_layer.outputs
reconstr_images = []
for i in xrange(imgs_to_reconstruct.shape[0]):
reconstr_img = sess.run(
x_reconstr,
feed_dict={input_images: [imgs_to_reconstruct[i, :, :, :]]})
reconstr_images.append(reconstr_img[0])
self.logger.info("Reconstructed %d/%d images" %
(i, imgs_to_reconstruct.shape[0]))
img_ops.save_gen_images(reconstr_images, out_path, transformer)
def main():
args = parse_args()
celeb_faces_generator = GanImgGenerator(args.gen_weights_path,
cifar10_gan.load_gen_with_weights)
out_vis_path = args.out_vis_path
celeb_faces_generator.generate_images(
args.how_many_samples,
out_vis_path,
transformer=lambda x: img_ops.unbound_image_values(x[:, :, ::-1]))
def __save_gen_images(gen_images_list, out_path, transformer):
imgs = [
transformer(img_ops.unbound_image_values(gen_img))
for gen_img in gen_images_list
]
imgs = np.array(imgs)
grid_size = int(np.sqrt(imgs.shape[0]))
big_img = img_ops.merge_images_to_grid(imgs, (grid_size, grid_size))
cv2.imwrite(out_path, big_img)
def main():
args = parse_args()
reconstructor = VaeImgReconstructor(load_enc_weights(args.enc_params_path),
load_gen_weights(args.gen_params_path),
get_latent_dim(args.gen_params_path))
cifar_imgs = np.array(
get_cifar_samples(args.cifar_ds_path, args.how_many_samples))
reconstructor.reconstruct_faces(
args.out_vis_path,
cifar_imgs,
transformer=lambda x: img_ops.unbound_image_values(x[:, :, ::-1]))
def generate_images(self, samples_num, out_path, transformer=lambda x: img_ops.unbound_image_values(x)):
sess = tf.InteractiveSession()
with tf.device(device_utils.get_device()):
z = tf.random_normal(shape=[1, self.z_dim], mean=0., stddev=1., dtype=tf.float32)
gen_out_layer, _ = self.load_gen_with_weights_func(sess, z, self.gen_params_path)
x_generated = gen_out_layer.outputs
gen_images = []
for i in xrange(samples_num):
x_gen = sess.run(x_generated)[0]
gen_images.append(x_gen)
self.logger.info("Generated %d/%d image" % (i, samples_num))
img_ops.save_gen_images(gen_images, out_path, transformer)
def inv_transform(transformed_img, scale_img_opt):
if scale_img_opt == '0_to_1':
return img_ops.unbound_images_values_01(transformed_img).astype(
np.uint8)
else:
return img_ops.unbound_image_values(transformed_img).astype(np.uint8)