Idea for sample\demo: ONNX networks reuse

[sergey-tihon]

up-for-grabs issue

Would be cool to have a CNTK.FSharp sample that

1. Load convolution base from NN trained with another framework (let's say Caffe2) and saved to ONNX
2. Add small NN head on top
3. Freeze base (to avoid weight's update)
4. Train custom head on new data
5. Profit

The idea is to show how to combine NNs and reuse NNs trained by tech giants on big clusters

[mathias-brandewinder]

That's a good idea, I like it!