Speech commands explanatory model

[swsuggs]

Adds explanatory model for poisoning fairness metrics.

Also corrects for an ART bug where triggered audio samples go out of range.

[swsuggs]

Most of the non-trivial changes here were needed because the code for the fairness metrics assumed that the explanatory model was a pytorch model, but the speech commands one is tensorflow. It also assumed image data, while speech commands is audio.

This should work for eval6. It makes no pretense to general applicability, meaning It is possible that future model architectures might require additional changes; but that's hard to predict.

[swsuggs]

@lcadalzo could you review when you get a chance

[swsuggs]

@lcadalzo the latest commits should address your points, let me know if there's anything else

[lcadalzo]

I had to change "compute_fairness_metric" in the config but now that I'm running the code affected by this PR, I'm seeing the following issue where the script is being killed by the self.explanatory_model() call:

On laptop, using breakpoint():

> /workspace/armory/metrics/poisoning.py(156)get_activations()
-> activation = self.explanatory_model(x_batch, training=False)
(Pdb) n
Killed
I have no name!@1d23884bdaeb:/workspace$

And on remote cluster using a GPU:

2023-02-01 17:57:03 49s INFO     armory.scenarios.poison:compute_explanatory:630 Computing fairness metrics
2023-02-01 17:57:09.852144: F ./tensorflow/core/util/gpu_launch_config.h:160] Check failed: work_element_count > 0 (0 vs. -1591109476)
Aborted (core dumped)

Have you been able to run this code successfully? Are you able to reproduce this?

Note: to make this run much faster, I'm also commenting out the fit() call to avoid waiting on it.

[swsuggs]

If you are running an audio config you may have to use fit_generator:true to avoid OOM issues

    ...
    "scenario": {
        "kwargs": {
            "fit_generator": true
        },
    ...

Not sure if that's the issue you are hitting; it's not a very descriptive error.

[swsuggs]

Ok I was finally able to replicate your error by skipping training and not using fit_generator. If I train but don't use fit_generator I get OOM.

[swsuggs]

@lcadalzo In case I wasn't clear, I don't believe the error you hit is an issue because it can be avoided by using fit_generator.

[lcadalzo]

What's the rationale behind fit_generator being a configurable scenario kwarg?

[swsuggs]

It runs more slowly but uses less memory. Some users may not find it necessary. See the discussion at 1761 and the info David added to the poisoning doc

[lcadalzo]

LGTM