Fix progress bar logging of metrics if compute_on_step=False

[teddykoker]

What does this PR do?

Quick fix, this issue from slack:

Hi, when using the Metrics API, I’m trying to set the compute_on_step=False when instantiating my ROC and AveragePrecision metrics, because I want to compute these over the entire epoch (as these metrics aren’t very useful batch-wise).
So in the training/val/test steps, I do something like self.roc(preds, truth), but I don’t log this in the step. Instead, at the epoch end, I compute auc() using the output of self.roc.compute(), which I then log, etc.
However, if I set compute_on_step=False , then PL throws the following error at the start of Epoch 0 (validation sanity epoch runs fine):
result[dl_key] = self[k]._forward_cache.detach()
AttributeError: 'NoneType' object has no attribute 'detach'
When looking at PL’s source code, it appears that if compute_on_step=False , then the _forward_cache property is set to and remains None, and is only updated to a non-None value if compute_on_step=True . Is this the intended behavior, or is this a bug? If this is intended, then how would the compute_on_step flag ever be able to set to False?
Thanks for the help!

cc, @tchaton I'm not sure how I feel about using the _forward_cache attribute
of the metric; I never really intended for that to be used outside the metric
class so there could be unintended consequences down the road. We might want to
consider a more robust solution here.

[codecov]

Codecov Report

Merging #5895 (28e8350) into release/1.2-dev (a96137c) will decrease coverage by 0%.
The diff coverage is 100%.

@@               Coverage Diff                @@
##           release/1.2-dev   #5895    +/-   ##
================================================
- Coverage               88%     88%    -0%     
================================================
  Files                  181     181            
  Lines                12929   12760   -169     
================================================
- Hits                 11426   11243   -183     
- Misses                1503    1517    +14     

[tchaton]

What does this PR do?

Quick fix, this issue from slack:

Hi, when using the Metrics API, I’m trying to set the compute_on_step=False when instantiating my ROC and AveragePrecision metrics, because I want to compute these over the entire epoch (as these metrics aren’t very useful batch-wise).
So in the training/val/test steps, I do something like self.roc(preds, truth), but I don’t log this in the step. Instead, at the epoch end, I compute auc() using the output of self.roc.compute(), which I then log, etc.
However, if I set compute_on_step=False , then PL throws the following error at the start of Epoch 0 (validation sanity epoch runs fine):
result[dl_key] = self[k]._forward_cache.detach()
AttributeError: 'NoneType' object has no attribute 'detach'
When looking at PL’s source code, it appears that if compute_on_step=False , then the _forward_cache property is set to and remains None, and is only updated to a non-None value if compute_on_step=True . Is this the intended behavior, or is this a bug? If this is intended, then how would the compute_on_step flag ever be able to set to False?
Thanks for the help!

cc, @tchaton I'm not sure how I feel about using the _forward_cache attribute
of the metric; I never really intended for that to be used outside the metric
class so there could be unintended consequences down the road. We might want to
consider a more robust solution here.

I agree. This solution is prone to errors in the future.