Demonstration of Dynamic Flow CNN
This is a tool for the implementation of dynamic CNN networks
A dynamic fature extraction module, featuring a tuneable complexity accuracy characteristic.
This module facilitates the execution of both a static and a dynamic CiFar10 CNN over the CiFar10 test set of 10K images taken and processed accordingly. The dynamic implementation not only is faster, but also its execution speed and accuracy can be tuned accordingly via changing a set of numerical values.
Results from the intermediate Pool3 layer, which can be used as a feature vector of 1024 values, are extracted and stored in an output CCEXP format. Furthermore the application provides results from the final layer which are used for validation purposes. In validation timing and accuracy can be evaluated via a python script which is also provided.