Support for image with no annotations in RetinaNet

[datumbox]

Fixes #2800

Merge after #3031, coauthored with @fmassa.

To ensure that the patch in this branch does not affect the performance of the model I retrained it and compared it with master.

Branch

retinanet_resnet50_fpn + no annotations support: boxAP=36.4

0: IoU metric: bbox
0:  Average Precision  (AP) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.364
0:  Average Precision  (AP) @[ IoU=0.50      | area=   all | maxDets=100 ] = 0.558
0:  Average Precision  (AP) @[ IoU=0.75      | area=   all | maxDets=100 ] = 0.384
0:  Average Precision  (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.190
0:  Average Precision  (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.401
0:  Average Precision  (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.492
0:  Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=  1 ] = 0.314
0:  Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets= 10 ] = 0.503
0:  Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.542
0:  Average Recall     (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.342
0:  Average Recall     (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.584
0:  Average Recall     (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.690

Master

retinanet_resnet50_fpn: boxAP=36.4

0: IoU metric: bbox
0:  Average Precision  (AP) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.364
0:  Average Precision  (AP) @[ IoU=0.50      | area=   all | maxDets=100 ] = 0.558
0:  Average Precision  (AP) @[ IoU=0.75      | area=   all | maxDets=100 ] = 0.381
0:  Average Precision  (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.193
0:  Average Precision  (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.400
0:  Average Precision  (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.489
0:  Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=  1 ] = 0.314
0:  Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets= 10 ] = 0.502
0:  Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.541
0:  Average Recall     (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.334
0:  Average Recall     (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.584
0:  Average Recall     (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.692

The above master metrics were taken from the recently merged PR on #2954.

[datumbox]

Rebasing this on the latest master fails an unrelated CI test:

• Windows GPU
• Linux GPU

Let's try rebasing later to see if this was the result of a bug on nightly upstream.

[fmassa]

Looks awesome, much simpler and more consistent, thanks a lot Vasilis!

[datumbox]

There seems to be something fishy with the failures on this PR. First of all the PR does not make any changes on the DeformConv2d operator which gets CI failures. Details:

• The commit c5d2868, passed all CI test on 20th of November (CI logs)
• Rebasing on latest master with commit 860669e, failed CI tests on 27th of November (CI logs)
• Undoing the rebase and force pushing the commit c5d2868, passes all the CI test on 27th of November (CI logs)

I tried rebasing multiple times and every time I push the change it fails. Every time I undo the the rebase it succeeds.

[datumbox]

I did a dummy PR #3054 to remove the test introduced and rebase and it works. It seems that the DeformConv2d test for autocast is flaky and we hit a bad seed. I'm going to try to fix this here by setting the seed to a different value.

[datumbox]

Added a couple of comments to explain the flakiness we observed for future reference.

[datumbox]

The addition of this model initialisation leads to the use of a different random seed on a separate test. See below.