David Slater

Unfortunately, we don't have much experience with userland containers, and haven't been able to make much progress here. Have you tried using Docker2Singularity? If so, what sorts of problems have...

Not presently in the scenarios. However the ART wrappers enable saving (https://github.com/IBM/adversarial-robustness-toolbox/blob/master/art/classifiers/pytorch.py#L474), so it shouldn't be difficult to modify a scenario to do so locally.

Some of those mentioned above are actually using the config wrapper_kwargs, if they exist. For example, in the resnet18 one: https://github.com/twosixlabs/armory/blob/53a7ccbd122ce3e05118416ba0d5442dee737fb9/armory/baseline_models/pytorch/resnet18.py#L55-L57 ``` input_shape=wrapper_kwargs.pop( "input_shape", (224, 224, 3) ), # default...

It shouldn't be downloading to `/tmp/.art/data/` if ART_DATA_PATH is set in armory/__init__.py. Are you importing armory before calling it?

Yeah, reading more fully, I think you're right - you'll still get the same error.

Yeah, I think using it should point to the subdir `checkpoint` instead of an individual file. Also, it looks like this method of loading is deprecated in TF2: https://www.tensorflow.org/api_docs/python/tf/compat/v1/train/init_from_checkpoint

Should be in ART 1.11

Good discussion of issues: https://jtreminio.com/blog/running-docker-containers-as-current-host-user/

Intel/Georgia Tech request.

We currently provide object detection per class and per-class accuracy. Confusion matrices would be nice, though I will take this off the current milestone for now.