Cycle-spinning (CS) method employement into densenet-pytorch application
The pytorch densenet re-implement of the DenseNet classification with Cycle-Spinning (CS) method employement performance in real time.
Extra information for cycle-spinning please visit: https://github.com/UlkuUZUN/Cycle-Spinning-NN
In the original code, the Cycle-Spinning (CS) method has been applied to the model.py file to work with 1 and 2 shifts in the first convolution process. In our experiments it is seen that CS method employement gets good results if it works with first convolution and small shifts up to 3 shifts.
densenet_cycle_spinning_max.py file an example of employment of cycle-spinning. max, mean or median aggregation functions have been tried.