[REVIEW] Adding RAPIDS <-> DLFrameworks Jupyter Notebook

[awthomp]

This notebook shows how to pass data between __cuda_array_interface__ supporting libraries (CuPy and Numba, for this demonstration) and both PyTorch and TensorFlow. When using PyTorch, we can simply call:

torch.as_tensor(foo) on an existing array that accepts the __cuda_array_interface__ but for TensorFlow, we leverage the nascent tfdlpack package described in this RFC.

There is a bug in tfdlpack.to_dlpack() that is documented here: VoVAllen/tf-dlpack#12 and also in this notebook.

[awthomp]

The blog for this PR hasn't been written yet, and its dependent on both tfdlpack and PyTorch 1.4. Do these notebooks assume that the user has configured libraries appropriately?

[taureandyernv]

The blog for this PR hasn't been written yet, and its dependent on both tfdlpack and PyTorch 1.4. Do these notebooks assume that the user has configured libraries appropriately?

@awthomp we should not assume that the user has configured these libraries properly. Also, to pass CI, we need the notebooks to be able to run unattended. I had spent some time trying to get tfdlpack-gpu to install and work in the docker container in an unattended, CI compliant way last Friday, but am yet to be successful.

We can also move this to the advanced notebooks section and ask to delay CI on it. Thoughts?

[taureandyernv]

tfdlpack-gpu install needs to work unattended to pass CI. Want to work together on this?

[awthomp]

@awthomp we should not assume that the user has configured these libraries properly. Also, to pass CI, we need the notebooks to be able to run unattended. I had spent some time trying to get tfdlpack-gpu to install and work in the docker container in an unattended, CI compliant way last Friday, but am yet to be successful.

We can also move this to the advanced notebooks section and ask to delay CI on it. Thoughts?

@taureandyernv I'd be happy to help on the tfdlpack-gpu install issues in your docker container. Are you able to share the specific error? Was the package able to be found? I noticed with Python 3.8 and Conda that a pip install tfdlpack-gpu would not work (i.e. could not find the package), but if you downgrade to 3.7, there were no issues.

Let's work on addressing the tfdlpack-gpu issue before making a decision on moving the notebook and pausing CI/CD. Sound good?