Mobilenet SSD architecture: Downloaded vs trained

问题:

The model zoo of Tensorflow's object detection API provides a bunch of pre-trained models that are ready to be downloaded here. I'm currently looking at ssd_mobilenet_v1_coco.

I can quickly obtain the computational graph of the model by running Tensorboard on the event file (see figure below)

However, if I train the model from scratch on the coco dataset and run Tensorboard on the event file obtained from the checkpoint, I get a computational graph that looks very different (although it has some similarities):

1) the entire graph appears to have been expanded by default,

2) the variable scope is not within "import" (like it is in the downloaded version),

3) there are many small discrepancies between the two graphs.

My goal is to obtain a similar frozen_inference_model.pb as that contained in the downloadable (discussed above).

Should the two computational graphs be expected to look different? If not, how would I obtain an identical one?