I went through the same issue. You need to do save_inference_model instead of save_persistables in the official model library of paddlepaddle SSD.

The target_vars and feeded_var_names can be obtained from build_program function.

I discussed this in length here ->

https://github.com/PaddlePaddle/Paddle/issues/19343

Hope this helps

great appreciation for your reply. your suggestion works so well and i sucessfully build a model whose output fits score.py perfectly  . it works well when i trained a new model and save it , but i still meets some problem when i want to load the model which saved by fluid.io.save_persistables . i make sure that i use the same feeded_var_names and target_vars. 

I am not sure I understood your exact concern. Anyway, here is a reply to what I think you are asking.

Whenever a model is saved using the function save_inference_model, it created a __model__ file, which dictates how the inference should run. score.py uses load_inference_model, which looks for this file. But, save_persistables does not create any such file. Thus, any model saved using save_persistables cannot be directly used for inference. That's why the model saved using save_persistables will not work with score.py

So you have 2 options now,

1. Train a new model and this time use save_inference_model instead of save_persistables

2. If you already have a trained model which you want to use, but it was saved using save_persistables, write a different code where you can load that model using load_persistables and then save it again using save_inference_model. Special case -> If this model is MobileNet-SSD, you can simply use the pretrained_model flag in paddlepaddle official code and then instead of save_persistables in the save_model function, use save_inference_model

Hope this helps.

看看，顺便学习有一下。

看看