I want to backpropagate the model that is already compile to the graph.
I will programming backpropagation code in python, but how can I update the weight?
Is there any way to access the graph?
@skim The Movidius graph file is a binary graph file format that isn't open-sourced at the moment. Normally the mvNCCompile tool is used to compile models that are deployment-ready. If you need to make changes to your model, you will have to make the changes with your own model before using the mvNCCompile tool. Once the tool is used to compile a Movidius graph file, there isn't a way to make edits to the Movidius graph representation of the model.
@Tome_at_Intel Thank you for the answer. I want to ask you one more question.
According to NCSDK documentation, there is a size limitations. (Compiled Movidius™ “graph” file < 320 MB; Intermediate layer buffer size < 100 MB)
I would like to know if the graph file is 320MB and attach the additional fully connected layer, (developed in python or something) is it possible to run?
Additional custom layer(not graph, python code) is just up to NCS memory spec?
@skim If I understand your question correctly, you are talking about the possibility of running a maximum size graph file through the NCS, and then take the result and run it through some fully connected layers on the CPU as a way of working around the NCS memory size limit? If this is what you are referring to, theoretically it should be possible to do this.