Modify expected value and threshold for retinanet unit-test.

[datumbox]

Updating the score_thresh=0.013 for RetinaNet returns non-empty results. This fixes #2810.

[codecov]

Codecov Report

Merging #2812 into master will not change coverage.
The diff coverage is n/a.

@@           Coverage Diff           @@
##           master    #2812   +/-   ##
=======================================
  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 072d8b2...a1b077e. Read the comment docs.

[fmassa]

LGTM, thanks a lot !

Let's just wait until CircleCI gets back to normal before merging this PR.

[fmassa]

Test failures seem related (and happen only on GPU). Maybe there are duplicate detections somewhere and due to floating point errors we get a slightly off prediction?

If debugging / fixing this is hard, we can skip the numerical tests for retinanet on the GPU (while performing the other tests), and re-enable those later?

[datumbox]

@fmassa I agree. This is likely due to floating point errors, possibly made worse from the fact that all weights and input data are random. I'll have another look, hopefully we won't need to disable the GPU tests.

[datumbox]

Unfortunately there is numerical and sort instability on this unit-test across platforms. Given that currently on master we don't sufficiently test retinanet, I think it's worth putting at least some tests in. As a workaround I disable some of the cuda checks but I plan to revisit this once I have a proper GPU setup.

[fmassa]

Awesome, thanks a lot @datumbox !