Adam Gleave

@levmckinney, what's outstanding here? I know you're busy now, but I can likely find someone else to finish the PR off. However, IIRC the policy ensemble didn't help performance much,...

Not immediately obvious to me why the other tests are failing, probably we're assuming pkl extension somewhere else in the code and will need to change it everywhere. This seems...

Kinda hard to say what's going on without seeing the stack trace. There are some files in `tests/testdata` that are still in `.pkl` format.

@Rocamonde do you mind taking a look at this sometime in the next few weeks? Not time sensitive, I don't think it's blocking anyone, more just causing confusion.

This normally happens if `pip` can't find compatible versions of packages, so tries to install all of them in the hope that one of them works (hint: it usually doesn't)....

> In `experiments/imit_benchmark.sh` it's just being suggested as a way to upload models, right? Seems right -- we probably want to delete that as well though if we're not supporting...

Also one test is failing on Mac OS (maybe flaky? I don't see anything Mac OS specific). And type checking failing --- I think pytype may be confused by the...

> Maybe just moving the failing data loader back to the test is reasonable ... I think this is the least bad option. I don't think it distracts from the...

For the flaky test: it doesn't seem to have flaked before, so I do wonder if it's down to changes in the expert policy? Even if it's similarly strong, small...

> Then you get a different error that I don't really understand - `unwrap_traj` ends up calling the `'rollout'` key of some object that doesn't have that key. https://github.com/HumanCompatibleAI/imitation/blob/master/src/imitation/data/wrappers.py#L160 adds...