Re-enable RetinaNet unit-tests on GPU

[datumbox]

Fixing #2824

When running the tests of ReminaNet on GPU, the outputs are checked twice:

vision/test/test_models.py

Line 193 in 5e33cc8

check_out(out)

vision/test/test_models.py

Lines 210 to 215 in 5e33cc8

	if dev == "cuda":
	with torch.cuda.amp.autocast():
	out = model(model_input)
	# See autocast_flaky_numerics comment at top of file.
	if name not in autocast_flaky_numerics:
	check_out(out)

On the first check, the values are very close:

expected: 
[{'boxes': tensor([[ 60.6603, 115.2130,  86.0578, 129.6661],
        [ 27.7266,  42.5258,  92.2687, 115.0305],
        [ 28.3789, 138.9524,  91.5207, 211.4824]], requires_grad=True), 'scores': tensor([0.0130, 0.0130, 0.0130], requires_grad=True), 'labels': tensor([2, 4, 4])}]

output: 
[{'boxes': tensor([[ 60.6603, 115.2130,  86.0578, 129.6661],
        [ 27.7266,  42.5258,  92.2687, 115.0306],
        [ 28.3788, 138.9524,  91.5207, 211.4824]], device='cuda:0',
       grad_fn=<StackBackward>), 'scores': tensor([0.0130, 0.0130, 0.0130], device='cuda:0', grad_fn=<CatBackward>), 'labels': tensor([2, 4, 4], device='cuda:0')}]

On the second that uses autocast() are not:

expected: 
[{'boxes': tensor([[ 60.6603, 115.2130,  86.0578, 129.6661],
        [ 27.7266,  42.5258,  92.2687, 115.0305],
        [ 28.3789, 138.9524,  91.5207, 211.4824]], requires_grad=True), 'scores': tensor([0.0130, 0.0130, 0.0130], requires_grad=True), 'labels': tensor([2, 4, 4])}]

output: 
[{'boxes': tensor([[ 60.6603, 115.2104,  86.0584, 129.6646],
        [ 54.6808,  85.3325,  80.0379,  99.6362],
        [ 28.3769, 138.9522,  91.5294, 211.4853],
        [ 27.7301,  42.4879,  92.2699, 115.0121]], device='cuda:0',
       grad_fn=<StackBackward>), 'scores': tensor([0.0130, 0.0130, 0.0130, 0.0130], device='cuda:0', dtype=torch.float16,
       grad_fn=<CatBackward>), 'labels': tensor([2, 2, 4, 4], device='cuda:0')}]

I thought it would be better to partially test the results on GPU, so I decided to re-enable it and include it in the autocast_flaky_numerics list.

[codecov]

Codecov Report

Merging #2825 into master will not change coverage.
The diff coverage is n/a.

@@           Coverage Diff           @@
##           master    #2825   +/-   ##
=======================================
  Coverage   73.26%   73.26%           
=======================================
  Files          99       99           
  Lines        8778     8778           
  Branches     1387     1387           
=======================================
  Hits         6431     6431           
  Misses       1920     1920           
  Partials      427      427           

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update adfc15c...b9c92e8. Read the comment docs.

[fmassa]

Awesome, thanks a lot for the investigation!