node.extend | A port of jQuery.extend that actually works on node.js | Frontend Framework library

My understanding of this code is:

• It creates a graph with 2 placeholders 'Postprocessor/convert_scores' and 'Postprocessor/ExpandDims_1'.

• Converts it to a graph_def and keeps the nodes corresponding to the placeholders.

  • Those 2 nodes correspond to the 1917 boxes the model outputs, the first one is the class probabilities the second one is the box coordinates.

• It creates a second graph and loads the trained model.

• It lists all the nodes in the graph and how they are connected to each other.
• Lists all the nodes connected to 'Postprocessor/convert_scores' and 'Postprocessor/ExpandDims_1', and store them in the keep list.

• Lists all the nodes not in nodes_to_keep_list and stores them in nodes_to_remove_list.

• It then creates a graph def and fills it with a copy of all the nodes_to_keep_list nodes.

• Then a second graph def with a copy of all the nodes_to_remove_list nodes.

• Finally it loads both graph defs, the first one with device '/gpu:0', the second '/cpu:0'.

The aim of this, as stated by the author, is to run the CNN on the GPU, and the postprocessing on the CPU because it is much faster there. If you look at mobilenet+SSD you'll see that the model outputs a bunch of boxes (1917) and then there is a rather complex (from a graph standpoint at least) post processing done on those boxes to provide the final output (detection_boxes, detection_scores, detection_classes, num_detections).

It's not visible in this piece of code, but the placeholders are later used to insert the output of the keep graph in the remove graph. The execution is performed in 2 steps (2 calls to sess.run())