David Slater

Regarding specific scenarios, I did not have a specific set in mind - but potentially all of the ones we provide configs for. I do not see this "results" directory...

Regarding containerization: 1) Portability/replicability/integration. For the program, it is extremely important that we can replicate all that is being done with a defensive technique. Once we start adding a lot...

From above, on "some subset of samples", I would just use the first 10 or 100 samples (depending on computation time) of a specific data split (typically "test") with file...

Breaking changes are listed here: https://github.com/tensorflow/datasets/releases/tag/v4.0.0

> Some Notes/Questions: > > * Is there a reason why we have `adversarial_datasets.py` split out from `datasets`? Yes, adversarial_datasets typically return something like `(x, x_adv), y` where standard datasets...

I wouldn't want to go to an output format like that as adversarial datasets are not commonly used, and that format would be weird in all other contexts.

No, I would like to keep the base datasets as standard as possible, which is to make them return `(x, y)` and only have the adversarial ones do something different....

Since TFDS v4 is merged in, I wanted to update the name

I can replicate. It doesn't impact mnist because that uses `tensorflow.compat.v1.placeholder`.

Would this be just for training or for inference time as well?