Vision: Empty expected values on test_retinanet_resnet50_fpn

[datumbox]

🐛 Bug

The Unit-test expected values of RetinaNet are empty and thus the test does not protect us from breaking the code.

To Reproduce

Steps to reproduce the behavior:

>>> import torch
>>> expected = torch.load("./test/expect/ModelTester.test_retinanet_resnet50_fpn_expect.pkl")
>>> print(expected)
[{'boxes': tensor([], size=(0, 4), requires_grad=True), 'scores': tensor([], requires_grad=True), 'labels': tensor([], dtype=torch.int64)}]

Expected behavior

Ensure we have some non-trivial expected values so that we detect it when the code breaks. @fmassa advised to try and find a way that does not involve download pretrained weights.

[vfdev-5]

@datumbox I wonder if this issue is exclusively for retinanet ?

[datumbox]

@vfdev-5 yes it is! Every other model is fine. :)