Adds Anchor tests

[oke-aditya]

Followup PR. Adds tests for #2960

Somehow I was not able to run tests locally. I'm unsure if this is fixed quite well.

I'm not sure what and how tests can or should be for this, somehow it simply creates an empty Anchor and no ValueErorr is generated.

[zhiqwang]

Hi, I think that something like this

vision/test/test_onnx.py

Lines 197 to 200 in 052edce

	def _init_test_rpn(self):
	anchor_sizes = ((32,), (64,), (128,), (256,), (512,))
	aspect_ratios = ((0.5, 1.0, 2.0),) * len(anchor_sizes)
	rpn_anchor_generator = AnchorGenerator(anchor_sizes, aspect_ratios)

and

vision/test/test_onnx.py

Lines 268 to 284 in 052edce

	class RPNModule(torch.nn.Module):
	def __init__(self_module):
	super(RPNModule, self_module).__init__()
	self_module.rpn = self._init_test_rpn()
	def forward(self_module, images, features):
	images = ImageList(images, [i.shape[-2:] for i in images])
	return self_module.rpn(images, features)
	images = torch.rand(2, 3, 150, 150)
	features = self.get_features(images)
	images2 = torch.rand(2, 3, 80, 80)
	test_features = self.get_features(images2)
	model = RPNModule()
	model.eval()
	model(images, features)

could be added to the unittest?

[fmassa]

Hi

Thanks for the PR!

@oke-aditya Test failures are related, can you have a look?

@zhiqwang yes, we can improve testing of AnchorGenerator to check for correctness as well, and this is just a first PR to start this effort. I think tests which would be useful would be tests for a few known values of a very small feature map size, so that we know the ground-truth.
Would you be willing to help with adding more tests to AnchorGenerator?

[oke-aditya]

Yes the failures are related. The reason is I'm not able to generate ValueError which I created with last PR.
Somehow I am not able to figure out why a simple tuple doesn't raise this.
Can someone help me ? 😓

[fmassa]

@oke-aditya might be because of

vision/torchvision/models/detection/anchor_utils.py

Lines 42 to 46 in 5a4bb19

	if not isinstance(sizes[0], (list, tuple)):
	# TODO change this
	sizes = tuple((s,) for s in sizes)
	if not isinstance(aspect_ratios[0], (list, tuple)):
	aspect_ratios = (aspect_ratios,) * len(sizes)

, which was there to help on those cases.

I think you should add instead a tuple of tuple, but with different sizes

[zhiqwang]

@zhiqwang yes, we can improve testing of AnchorGenerator to check for correctness as well, and this is just a first PR to start this effort. I think tests which would be useful would be tests for a few known values of a very small feature map size, so that we know the ground-truth.
Would you be willing to help with adding more tests to AnchorGenerator?

Hi @fmassa , Yep, It's my pleasure, I will trace this current PR, and check what can I contribute to following up this PR.

[oke-aditya]

I'm really unsure what code would generate the error. I have tried various combinations locally but no success yet.

Following is what I tried to generate the ValueError but somehow it isn't getting generated

I think there is some confusion.

assert len(sizes) == len(aspect_ratios)

This assert actually checks if sizes and aspect ratios match.

This can be easily triggered.
E.g. case

incorrect_sizes = ((128, 258, 512), (128, 32))
incorrect_aspects = ((0.5, 1.0, 2.0),),

But this

if not (len(grid_sizes) == len(strides) == len(cell_anchors)):

is a call to very internal code in the forward method which makes it quite hard to get such test cases.

Some test cases that didn't work to fix the ValueError

incorrect_sizes = ((128), (258, 32), )
incorrect_aspects = ((0.5, 2.0), (3.0))

incorrect_sizes = ((128, 128, 128), (258, 32), )
incorrect_aspects = ((0.5, 2.0), (3.0))

Passing only tuple too didn't work. I think that it needs further digging into code.

[fmassa]

@oke-aditya you need to set the features to have a different (wrong?) number of feature maps

[oke-aditya]

Ah, now I get it. I have fixed the test. 😄

[fmassa]

Thanks!