Return loss information for detection models

[Delaunay]

Fix for #1574

Why:
Current code assume that loss is only useful in training mode.
This is not true, users might want validation and test loss.

How:

1. We replace current logic of checking for training mode by checking if targets are provided.
   If targets are provided then we assume users are interested in loss.

2. We always return a loss dictionary even if it is empty.
   This makes the API consistent and users do not have to change code between eval and training mode

[codecov-io]

Codecov Report

❗ No coverage uploaded for pull request base (master@5b1716a). Click here to learn what that means.
The diff coverage is 12.5%.

@@            Coverage Diff            @@
##             master    #1605   +/-   ##
=========================================
  Coverage          ?   65.67%           
=========================================
  Files             ?       90           
  Lines             ?     7083           
  Branches          ?     1074           
=========================================
  Hits              ?     4652           
  Misses            ?     2125           
  Partials          ?      306

Impacted Files	Coverage Δ
torchvision/models/detection/roi_heads.py	51.13% <0%> (ø)
torchvision/models/detection/rpn.py	82.19% <0%> (ø)
torchvision/models/detection/generalized_rcnn.py	84.61% <100%> (ø)

---

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 5b1716a...f1fe576. Read the comment docs.

[fmassa]

Hi,

I like the idea of this change.
It will even make it easier for making the model be torchscript-compatible, because torchscript can't handle dynamic return values.

This is unfortunately a BC-breaking change, so we can't merge this as is for now. We will be cutting a new release of torchvision in the beginning of January, and that version will have a deprecation warning saying that we will be changing the return values to be always loss, detections.
One more thing is that we now have a very unintuitive behavior: if we pass the targets, we don't compute the postprocessed predictions, but if we don't, then they are postprocessed. I think we should iterate a bit on that in January, maybe by adding a new argument to the model to postprocess the outputs.

In the meantime, the training scripts in references/detection/ are now broken because of this change, and should be fixed before merge.

[Delaunay]

Merged the master branch to get the torch.jit changes as well.
Failing tests are the same as the master branch.

Todos

• updated references/detection/
• detect when to postprocess predictions (if targets are passed they are not)
  • ? adding a new argument to the model to postprocess the outputs.