Remove model download from tests

[fmassa]

They can introduce flakiness due to connection errors, and slow tests down, while not being needed for the tests that we currently have.

[vfdev-5]

@fmassa I understand the issue to fix and this can help. But how about test_hub.py and test_onnx.py which also download pretrained wieghts ?

vision/test/test_hub.py

Line 32 in 02f5c1f

pretrained=True,

vision/test/test_onnx.py

Line 379 in 02f5c1f

model = models.detection.faster_rcnn.fasterrcnn_resnet50_fpn(pretrained=True, min_size=200, max_size=300)

What do you think to accept failing if we catch ConnectionResetError ?

[fmassa]

@vfdev-5 ideally they should all avoid downloading weights, and we should fix the tests for this. I've messed up with both of the tests here, I'll be fixing them in a few moments

[codecov]

Codecov Report

Merging #2806 into master will decrease coverage by 0.13%.
The diff coverage is 100.00%.

@@            Coverage Diff             @@
##           master    #2806      +/-   ##
==========================================
- Coverage   73.31%   73.17%   -0.14%     
==========================================
  Files          99       99              
  Lines        8724     8728       +4     
  Branches     1373     1373              
==========================================
- Hits         6396     6387       -9     
- Misses       1909     1925      +16     
+ Partials      419      416       -3     

Impacted Files	Coverage Δ
torchvision/models/detection/backbone_utils.py	95.55% <100.00%> (+1.26%)	⬆️
torchvision/models/detection/faster_rcnn.py	80.95% <100.00%> (-0.45%)	⬇️
torchvision/models/detection/keypoint_rcnn.py	88.23% <100.00%> (-0.34%)	⬇️
torchvision/models/detection/mask_rcnn.py	90.00% <100.00%> (-0.33%)	⬇️
torchvision/models/googlenet.py	75.13% <0.00%> (-7.19%)	⬇️

---

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 ebe3f02...3f5da41. Read the comment docs.

[datumbox]

@fmassa LGTM, thanks for the PR.

I've left only 1 comment to confirm that a corner-case is handled as intended. If it is, feel free to merge.