Icevision: [Update] Albumentations 0.5.0

[lgvaz]

🐛 Bug

The new release of albumentations 0.5.0 introduced breaking changes, the error is currently being avoided by specifying a version lower than 0.5.0 in setup.cfg

How to reproduce
Update albumentations and run: pytest tests/transforms/test_albu_transform.py

Stracktrace

 internal_data_name = 'masks'

    def _check_args(self, **kwargs):
        checked_single = ["image", "mask"]
        checked_multi = ["masks"]
        # ["bboxes", "keypoints"] could be almost any type, no need to check them
        for data_name, data in kwargs.items():
            internal_data_name = self.additional_targets.get(data_name, data_name)
            if internal_data_name in checked_single:
                if not isinstance(data, np.ndarray):
                    raise TypeError("{} must be numpy array type".format(data_name))
            if internal_data_name in checked_multi:
>               if data:
E               ValueError: The truth value of an array with more than one element is ambiguous. Use a.any() or a.all()

/opt/hostedtoolcache/Python/3.6.12/x64/lib/python3.6/site-packages/albumentations/core/composition.py:223: ValueError

[FraPochetti]

I think I might benefit from learning how to effectively debug, so I can take this one if you accept.

[lgvaz]

Thanks Francesco!!! You're on fire 🔥

[FraPochetti]

I think I found the culprit.
The problem is in the tfms applied to masks.
This line should look like params["masks"] = list(masks.data).
In the albumentations docs they specify that masks needs to be a list (here).
I made the edit and now tests are passing.
Before I create a PR, does this make sense to you?

[FraPochetti]

Also, not sure how to handle the dependency.
Assuming the fix works, shall we also edit this line removing the <0.5 constraint?
I mean, in the same PR.

[FraPochetti]

I also tested the change (e.g. params["masks"] = list(masks.data)) with albumentations version 0.4.6 and it still works, so it looks backwards compatible.

[lgvaz]

This line should look like params["masks"] = list(masks.data)

Interesting, for some reason they decided not to accept a numpy array anymore?

Before I create a PR, does this make sense to you?

Yeah yeah! To get the tests passing were the only thing required on this one

I mean, in the same PR.

Yes, both changes in the same PR, but instead of removing the condition put it <0.6, so it doesn't break unexpectedly in the future

[FraPochetti]

PR created!