Remove _has_warned attribute from detection models

[datumbox]

🚀 Feature

Several detection models make use of an attribute _has_warned which was originally used by JIT to avoid throwing the same warning multiple times:

vision/torchvision/models/detection/retinanet.py

Lines 373 to 374 in 9778d26

	# used only on torchscript mode
	self._has_warned = False

vision/torchvision/models/detection/retinanet.py

Lines 554 to 556 in 9778d26

	if not self._has_warned:
	warnings.warn("RetinaNet always returns a (Losses, Detections) tuple in scripting")
	self._has_warned = True

JIT should now be able to handle this automatically without the need of this variable. Thus we should remove the specific variable and confirm that calling forward multiple times on a JIT-scripted model produces a single warning. If that's indeed the case, we can remove it from all places:

• RetinaNet
• GeneralizedRCNN
• SSD (pending merge at Add SSD architecture with VGG16 backbone #3403)

[prabhat00155]

If I remove the if check from line 554, I see the warning appear every time forward() is invoked. Here is what I did:

model = retinanet_resnet50_fpn(pretrained=True)
model.eval()
img = torch.rand([1, 3, 512, 512])
model(img)
for p in model.parameters():
    p.requires_grad_(False)
scripted_model = torch.jit.script(model, img)
for _ in range(10):
    img_test = torch.rand([3, 512, 512])
    scripted_model([img_test])

This prints the warning 10 times

[datumbox]

Thanks for the investigation @prabhat00155!

Sounds like it's not fixed yet on the TorchScript side and can't remove it. I'll be closing the issue and we can revisit on the future once it's patched.