数据增强可以使用其他库函数吗？

[oyoanan]

作者您好，我在其他项目中加入您的车道线训练标签过程中，似乎因为DALI不支持多进程导致数据增强失败。所以，想问下您的车道线标签数据，是只能使用DALI增强，还是可以使用其他类似的库吗？最重要的是这里的labels要如何跟随图像的变换而变换呢？希望可以得到您的回复，非常感谢！
def ExternalSourceTrainPipeline(batch_size, num_threads, device_id, external_data, train_width, train_height, top_crop, normalize_image_scale = False, nscale_w = None, nscale_h = None): pipe = Pipeline(batch_size, num_threads, device_id) with pipe: jpegs, seg_images, labels = fn.external_source(source=external_data, num_outputs=3) images = fn.decoders.image(jpegs, device="mixed") seg_images = fn.decoders.image(seg_images, device="mixed") if normalize_image_scale: images = fn.resize(images, resize_x=nscale_w, resize_y=nscale_h) seg_images = fn.resize(seg_images, resize_x=nscale_w, resize_y=nscale_h, interp_type=types.INTERP_NN) # make all images at the same size size = encoded_images_sizes(jpegs) center = size / 2 mt = fn.transforms.scale(scale = fn.random.uniform(range=(0.8, 1.2), shape=[2]), center = center) mt = fn.transforms.rotation(mt, angle = fn.random.uniform(range=(-6, 6)), center = center) off = fn.cat(fn.random.uniform(range=(-200, 200), shape = [1]), fn.random.uniform(range=(-100, 100), shape = [1])) mt = fn.transforms.translation(mt, offset = off) images = fn.warp_affine(images, matrix = mt, fill_value=0, inverse_map=False) seg_images = fn.warp_affine(seg_images, matrix = mt, fill_value=0, inverse_map=False) labels = fn.coord_transform(labels.gpu(), MT = mt) images = fn.resize(images, resize_x=train_width, resize_y=int(train_height/top_crop)) seg_images = fn.resize(seg_images, resize_x=train_width, resize_y=int(train_height/top_crop), interp_type=types.INTERP_NN) images = fn.crop_mirror_normalize(images, dtype=types.FLOAT, mean = [0.485 * 255, 0.456 * 255, 0.406 * 255], std = [0.229 * 255, 0.224 * 255, 0.225 * 255], crop = (train_height, train_width), crop_pos_x = 0., crop_pos_y = 1.) seg_images = fn.crop_mirror_normalize(seg_images, dtype=types.FLOAT, mean = [0., 0., 0.], std = [1., 1., 1.], crop = (train_height, train_width), crop_pos_x = 0., crop_pos_y = 1.) pipe.set_outputs(images, seg_images, labels) return pipe