Port test_quantized_models.py to pytest

[NicolasHug]

I'm taking care of this one myself because it's a bit more subtle than the rest

[datumbox]

@NicolasHug Thanks for the PR.

I think you caught an issue on our previous test, see below.

[datumbox]

I think it would be best both for consistency and functionality reasons to use the same approach as in:

vision/test/test_models.py

Lines 397 to 404 in 44fefe6

	defaults = {
	'num_classes': 50,
	'input_shape': (1, 3, 224, 224),
	}
	kwargs = {**defaults, **_model_params.get(model_name, {})}
	input_shape = kwargs.pop('input_shape')
	model = models.__dict__[model_name](**kwargs)

Then InceptionV3 can overwrite its parameters as follows:

vision/test/test_models.py

Lines 179 to 182 in 44fefe6

	_model_params = {
	'inception_v3': {
	'input_shape': (1, 3, 299, 299)
	},

[NicolasHug]

Looks like we can use input_shape = (1, 3, 224, 224) for all models including inception, with no impact on computation time. So let's leave as-is, unless we must use (1, 3, 299, 299) for a reason that I ignore?

[datumbox]

It is true that most of the models do accept smaller input sizes than the canonicals. But in the case of classification where speed is not a problem, we always use the correct size (hence why the previous code defined 299). Given that we have an investigation of how we can use the actual pre-trained weights (provided we won't introduce flakiness from downloading data over the internet), I would recommend to follow the same logic as above.

The change should be very very easy to make and it will follow the example of all other models, building a stronger case of why it can be part of the test_models.py.

[NicolasHug]

OK, I'll use (1, 3, 299, 299) and move to test_models.py, no need to do that in another PR.

Given that we have an investigation of how we can use the actual pre-trained weights (provided we won't introduce flakiness from downloading data over the internet)

Could you detail what you mean by that?

[datumbox]

OK, I'll use (1, 3, 299, 299) and move to test_models.py, no need to do that in another PR.

That works.

Could you detail what you mean by that?

Currently we use random weight init for the models in the tests which means that a) we don't validate against real images and real data and b) we often see flakiness (see on detection). If we had a way to use actual weights without downloading them (maybe exploit caching of CircleCI) we could improve the quality of our tests. Happy to chat offline about that as it's a big topic unrelated to this PR.

[datumbox]

I understand that you removed this because we filtered out inception incorrectly. Let's restore this using the testing approach on the normal models, as described above.

[NicolasHug]

Thanks for the review @datumbox . BTW, do you know why this test is in a file on its own? We might as well move it to test_models.py, right?

[datumbox]

@NicolasHug I would be 100% OK to move it along with the other models. Given that other submodules such as detection and segmentation are in the main models file, I don't see why quantized is any different.

Perhaps we can move them to models on a separate follow up PR?

[datumbox]

LGTM, thx!