def get_batch(self, batch_size):
x_imgs, y_imgs = self.data.get_batch(batch_size)
select_rg = random.choice(['0', '90'])
select_axis = random.choice(['0', '1', '2', '-1'])
x_en_imgs = utils.enhance_imgs(x_imgs, rotate_rg=int(select_rg), flip_axis=int(select_axis), is_swirl=self.is_swirl)
y_en_imgs = utils.enhance_imgs(y_imgs, rotate_rg=int(select_rg), flip_axis=int(select_axis))
return x_en_imgs,y_en_imgs
def enhance_predict(lr_imgs, network=None):
outs_list = []
for _, flip_axis in enumerate([0, 1, 2, -1]):
for _, rotate_rg in enumerate([0, 90]):
en_imgs = utils.enhance_imgs(lr_imgs, rotate_rg, flip_axis)
outs = network.predict(en_imgs)
anti_outs = utils.anti_enhance_imgs(outs, rotate_rg, flip_axis)
outs_list.append(anti_outs)
return np.mean(outs_list, axis=0)
def ensem_predict(input_imgs, network):
# ensembling
outs_list = []
for _, flip_axis in enumerate([0, 1, 2, -1]):
for _, rotate_rg in enumerate([0, 90]):
en_imgs = utils.enhance_imgs(input_imgs, rotate_rg, flip_axis)
outs = network.predict(en_imgs)
composed_img = utils.compose_dwt_images(outs, FLAGS.wavelet)
anti_outs = utils.anti_enhance_imgs(composed_img, rotate_rg,
flip_axis)
outs_list.append(anti_outs[0])
output = np.mean(outs_list, axis=0)
return [output]