fix transform for rcnns so original images list is unchanged

[buoyancy99]

Maskrcnn or fastrcnn pass input image list into the transform module in torchvision/models/detection/transform.py. However, the method modifies content in the list where transformed images are pasted back to corresponding entries of the input list.

As a result, user may meet unintended behavior when they hold reference to the original input list outside the maskrcnn call. The proposed change leave user input image list to the maskrcnn call unchanged.

[codecov-io]

Codecov Report

Merging #1084 into master will increase coverage by 0.74%.
The diff coverage is 100%.

@@            Coverage Diff             @@
##           master    #1084      +/-   ##
==========================================
+ Coverage   63.92%   64.67%   +0.74%     
==========================================
  Files          68       68              
  Lines        5406     5407       +1     
  Branches      829      829              
==========================================
+ Hits         3456     3497      +41     
+ Misses       1707     1661      -46     
- Partials      243      249       +6

Impacted Files	Coverage Δ
torchvision/models/detection/transform.py	78.3% <100%> (+0.2%)	⬆️
torchvision/datasets/cityscapes.py	70% <0%> (+50%)	⬆️

---

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 6c56029...44157a4. Read the comment docs.

[fmassa]

Thanks for the PR!

I have a few comments.

Also, could you add tests for this case? Maybe in tests/test_models.py

[buoyancy99]

Thanks for the PR!

I have a few comments.

Also, could you add tests for this case? Maybe in tests/test_models.py

fixed and added test in #1085, however #1085 will fail auto check since this is not merged yet. Please rerun the check after merging

[fmassa]

Thanks!