stylepro_artistic API内存不足

[60999]

欢迎您反馈PaddleHub使用问题，非常感谢您对PaddleHub的贡献！
在留下您的问题时，辛苦您同步提供如下信息：

• 版本、环境信息
  1）PaddleHub和PaddlePaddle版本：
  PaddleHub1.6.1，paddlepaddle-gpu 1.7.1.post97
  2）系统环境：

aistudio平台GPU环境

stylepro_artistic 使用GPU出错,在转换多张照片之后.

代码:
`def stylepro_artistic(fore_image_path,paddlehub_out_path):

1.指定待预测图片路径，遍历循环抠图 保存到 paddlehub_out_path

g = os.walk(fore_image_path)
for path,dir_list,file_list in g: 
    file_list = natsorted(file_list)
    for file_name in file_list:
        #取出一个前景文件
        front_pic = os.path.join(path, file_name)
        # print (front_pic)
        # 2.加载预训练模型
        module = paddlehub.Module(name="stylepro_artistic")
        input_dict = {"content": front_pic,"styles":["/home/aistudio/work/Picasso_pic.jpg"]}
        #print (input_dict)
        #GPU开启下面一行,如果用下面一行，必须安装paddlepaddle-gpu
        results = module.style_transfer(paths=[input_dict],use_gpu = True,visualization=True,output_dir=paddlehub_out_path)
        #以下还可以优化吗
        #for result in results: 
        #print(result)

stylepro_artistic(front_path,HUMANSEG_SPACE)
`

[nepeplwu]

请问下你所用GPU的显存是多大？是否有其他进程也占用了显存？

[60999]

aistudio平台 16/32 G

[nepeplwu]

@60999 是否可以提供下aistudio的链接？我们fork复现看。此外，下面的代码建议放在for循环外，没有必要每次都加载一次module

module = paddlehub.Module(name="stylepro_artistic")

[60999]

怀疑就是多次载入模型导致的。调用其他模型时候也遇到过这种情况，在升级其他模型的更新版本后没出现过问题，stylepro_artistic 模型后来没有再测试过。
建议其他模型都载入几百次测试下，没准其他模型也有这种问题。

[InternetMaster1]

@nepeplwu

This can be used on Android?

[Steffy-zxf]

@InternetMaster1 As for mobile-end usage， please refer to PaddleLite: https://github.com/PaddlePaddle/Paddle-Lite

[InternetMaster1]

@Steffy-zxf @nepeplwu

1. Any guide on how to convert "stylepro_artistic" to Android using PaddleLite. I don't see any demo present like other projects such as segmentation, etc have.
   https://github.com/PaddlePaddle/Paddle-Lite-Demo/tree/master/PaddleLite-android-demo

2. What is the model size and the processing time for style transfer? If you check Tensorflow document, they have mentioned the model file size and time taken.
   https://blog.tensorflow.org/2020/04/optimizing-style-transfer-to-run-on-mobile-with-tflite.html
   

What are the similar size and latency values for Paddle StyleTransfer module?

3. What are the difference in the below two?
   https://github.com/PaddlePaddle/hapi/tree/b337063c7ecf44c60944801dbd32453f1481d987/examples/style-transfer
   https://github.com/PaddlePaddle/PaddleHub/tree/dbca09ae78b5387ebe3b49f37ce88de45d41d26a/hub_module/modules/image/style_transfer/stylepro_artistic

[InternetMaster1]

@Steffy-zxf @nepeplwu

Is it possible to run "stylepro_artistic" on our own server, or on the mobile phone as offline code?

On PaddleHub, I only see the API.

We wish to provide stylepro_artistic for offline purpose on mobile phone similar to other libraries like PaddleSeg. Is this possible?