How to upload prediction data from the FCNN-AlexNet neural network?

Question

Python

The essence of the question is as follows: on a very old hardware (Nvidia Tegra TK20), it was necessary to launch object detection using neural networks. The choice fell on the Caffe framework and the FCNN-AlexNet network (the prototype is here ). Having written a simple layer in Python using examples, I figured out a little about network training (the input is an image + a byte map with zones describing objects of certain classes, the maximum number of classes is 255). It remains unclear how it would be possible to start the network itself for recognition without introducing an additional byte map and unload the result of the prediction - a new byte map with the expected objects, if at all possible. Please help and examples!