Hello to all,
I looked in the existing discussions but I did not find discussions approaching my problem where then I did not see it and I apologize.
So here is my problem, I am trying to make the image classification in real time with a hybrid approach, I take a CNN, I cut it in the penultimate layer and I replace the back layer with an SVM.
However, when I use the get_tnsor_by_name () function I always find myself with a tensor or operation error that does not exist in the graph.
I followed the following tutorial:
But nothing works, so I wanted to know if it was possible to cut the graph when using the key Movidius (because I have no problem to do that without the key).
Have a good day.
@KitsuneM Are you trying to cut the network by taking the output and placing the output in a SVM running as a separate network or cut and add the SVM into the network itself? If it is the latter, it would likely require you to do it during training. I haven't done this myself, but let us know if you get it working. Thanks.
My program is divided into two parts, the first is to recover the features and labels of my images. I write the features in one file and the labels in another. To recover the features I cut the graph at the penultimate layer and I recover what it gives me.
Then when I have these two files, I give them as a parameter to an SVM that will take care of doing the training, I do not try to add the SVM to the graph.
In addition I noticed that I can not use all types of graph, which can be a disadvantage, like the graph ".pb".
For now it still does not work.