where can I start for building a simple CNN by myself?

Dear all,
I would like to test a modified back-propagation algorithm for a CNN, but some existed frameworks such as TensorFlow didn’t allow me to achieve it because my bp algorithm is not the same the tf’s, so perhaps I need to implement the algorithm using cuDNN by myself. But where can I start? Are there any simple samples or tips?
thanks advance!

What kind of backprop are you talking about? I’ve never used tensor flow or any other framework. I’ve also went out of my way to not program in Python. So, I don’t know the limits of using those doodads. The best place to start is by building a cnn from scratch.