Created by: wojtuss

When training a MobileNet-SSD model, the train.py script allows for saving persistables using

fluid.io.save_persistables(exe, model_path, main_program=main_prog)

But how can we save the whole model? In order to save the whole model for inference using:

fluid.io.save_inference_model("abc", feed_dict, [target], exe)

we need to provide feed_dict and target.

1. What should the feed_dict be? Usually, the feed argument of exe.run() method suggest the answer, but in MobileNet-SSD, the feed is not defined. The py_reader is used there.
2. What should the target be? Several attempts to put anything into target failed with message

Traceback (most recent call last):
...
  File "/home/wojtuss/repos/PaddlePaddle/Paddle/build/python/paddle/fluid/io.py", line 654, in save_inference_model
    main_program = main_program._prune(targets=target_vars)
  File "/home/wojtuss/repos/PaddlePaddle/Paddle/build/python/paddle/fluid/framework.py", line 1739, in _prune
    "The target variable must have an "
ValueError: The target variable must have an associated operator that generates it.

3. Is it possible to save the model with the DetectionMAP evaluator?