Finish conv2d about tf-encrypted

[jvmncs]

I think pushing forward on the backwards pass for conv2d as in #156 would be super helpful to us!

[kamathhrishi]

Could you elaborate more on what's missing?

[justin1121]

I believe the backward pass is what is missing. There is a branch open with some initial work here https://github.com/mortendahl/tf-encrypted/pull/156, might need to coordinate with @koenvanderveen to find out more!

[justin1121]

There is also some work being done related to conv1d and might require some changes to the current conv2d implementation that could effect the backward pass. @jvmancuso or @yanndupis might have more information regarding this.

[jvmncs]

Any changes for conv1d can adapt to what's done in the meantime. We're not entirely sure we'll need to change conv2d, but there's a lot to be done before we would need to modify conv2d, if at all.

[kamathhrishi]

@jvmancuso while there is a PR for conv2d backward pass could I know what else could be done?

[koenvanderveen]

Hey guys, sorry for not being responsive, I am holiday, back this saturday. While trying to reimplement pond conv2d backward code from the numpy code i got stuck on finding a tf equivalent for the np.add.at function

[vincehong]

ABY3 protocol now has fully functional NN training.