Fixed bug: replaced bce_loss_with_logits with bce_loss

[amisev]

What does this PR do?

Fixed bug in pl_examples/domain_templates/computer_vision_fine_tuning.py:
replaced binary_cross_entropy_with_logits with binary_cross_entropy.

See #7081

[carmocca]

Why not remove the torch.sigmoid in forward instead?

[codecov]

Codecov Report

Merging #7096 (9df85f7) into master (6be0a85) will decrease coverage by 5%.
The diff coverage is n/a.

@@           Coverage Diff           @@
##           master   #7096    +/-   ##
=======================================
- Coverage      91%     87%    -5%     
=======================================
  Files         198     198            
  Lines       12634   12790   +156     
=======================================
- Hits        11541   11067   -474     
- Misses       1093    1723   +630     

[amisev]

Some losses in torchmetrics are required probabilities rather than scores.

[awaelchli]

I would go for the more numerically stable option.
Can't we recompute sigmoid outputs where we need them?

[amisev]

I removed sigmoid activation from the forward pass, but anyway, even torchmetrics.Accuracy required probabilities, that's why I had to add torch.sigmoid two more times.

[awaelchli]

great fix!