from multiprocessing import Pool
def filter_img(img):
gauss = np.zeros(img1.shape, img1.dtype)
for j in range(3):
gauss[:, :, j] = filters.gaussian_filter(img1[:, :, j], 10)
return gauss
def bilateral_filter(old_path):
print(old_path)
out_path = old_path.replace("IJCAI_fgsm_output", "holdEdge_output")
#out_path = old_path
out_path_dir = os.path.split(out_path)[0] + "/"
if not os.path.exists(out_path_dir):
os.makedirs(out_path_dir)
input_img = Image.open(old_path)
input_img = np.array(input_img)
output_img = np.zeros(input_img.shape, np.float32)
output_img = filter_img(input_img)
Image.fromarray(output_img.astype(np.uint8)).save(out_path)
train, _ = get_filename('./fgsm_val_list.txt')
p = Pool(8)
p.map(bilateral_filter, train)
p.close()
p.join()
# tensor-scalar operation
images = (images / 255.0) * 2.0 - 1.0
return images
if 'resnet' in model_type.lower() or 'vgg' in model_type.lower():
_R_MEAN = 123.68
_G_MEAN = 116.78
_B_MEAN = 103.94
images = tf.image.resize_bilinear(images, [224,224],align_corners=False)
tmp_0 = images[:,:,:,0] - _R_MEAN
tmp_1 = images[:,:,:,1] - _G_MEAN
tmp_2 = images[:,:,:,2] - _B_MEAN
images = tf.stack([tmp_0,tmp_1,tmp_2],3)
return images
train_path, train_label = get_filename('./train_list.txt')
val_path, val_label = get_filename('./val_list.txt')
def load_input_images(batch_shape):
filenames = []
output_filenames = []
input_images = np.zeros(batch_shape)
idx = 0
batch_size = batch_shape[0]
for i in range(len(val_path)):
path = val_path[i]
output_path = path.replace("IJCAI_2019_AAAC_train", "IJCAI_fgsm_val")
out_path_dir = os.path.split(output_path)[0] + "/"
if not os.path.exists(out_path_dir):
os.makedirs(out_path_dir)