Some metrics don't work on CPU using float16

[Llannelongue]

🐛 Bug

It looks like some metrics such as Precision-Recall curve don't work on CPUs when using float16, perhaps due to a missing feature in pytorch?

Please reproduce using the BoringModel

https://colab.research.google.com/drive/1xDv043rRi5WBshP4m5aoxTt2ChlfxjIk?usp=sharing

Expected behavior

the metrics should work in half precision on CPUs as well.

Environment

• CUDA:
  • GPU:
    • Tesla T4
  • available: True
  • version: 10.1
• Packages:
  • numpy: 1.19.4
  • pyTorch_debug: True
  • pyTorch_version: 1.7.0+cu101
  • pytorch-lightning: 1.1.2
  • tqdm: 4.41.1
• System:
  • OS: Linux
  • architecture:
    • 64bit
  • processor: x86_64
  • python: 3.6.9
  • version: Proposal for help pytorch-lightning#1 SMP Thu Jul 23 08:00:38 PDT 2020

[Borda]

@SkafteNicki mind have look? :]

[SkafteNicki]

Definitely an lack of support from pytorch side, but seems to be on track to be solved:
Issue: pytorch/pytorch#49889
PR with fix: pytorch/pytorch#49895
I can try to see if the code can be rewritten with operations that support float16
Probably should also add test for float16 in general for metrics.

[SkafteNicki]

Just a follow up: the following metrics does not work with float16 CPU because of lack of support from a list of pytorch operations (in parentheses)

• SSIM (torch.arange)
• PSNR (torch.pow)
• MeanSquaredLogError (torch.logp1 & torch.pow)
• MeanSquaredError (torch.pow)
• ExplainedVariance (torch.pow)
• PrecisionRecallCurve (torch.flip)
• AveragePrecision (torch.flip)
  @Borda how to move forward with this? I think some of these can be converted in operations that support float16 CPU (pow can be made to multiplication) but I am not sure about all of them

[edenlightning]

@Borda

[github-actions]

Hi! thanks for your contribution!, great first issue!

[edenlightning]

looks like this was fixed in pytorch/pytorch@5d93e2b