Support metrics with NaNs in them on IPU

[DomInvivo]

When working with IPU's, NaNs cannot be filtered out as is typically done on GPU/CPU. We need fancier ways of masking NaNs on the targets. I have started implementing a few metrics/losses such as BCE, MSE, and L1 in the ipu_metrics.py file, but other metrics need to be implemented.

Classification metrics

• [x] Accuracy (See ignore_index here to create a fake class that takes the NaNs as a new class, but ignores them.
• [x] Precision (See ignore_index)
• [x] Recall (See ignore_index)
• [x] Auroc (See sample_weights here to set the nan-target as anything, but give a 0-weight to it)
• [x] AveragePrecision (See sample_weights here to set the nan-target as anything, but give a 0-weight to it)
• [x] F1_score here
• [x] F_beta score

Regression metrics

• [x] PearsonR
• [x] SpearmanR: Cannot work due to indexing, unless we rewrite the function. Will not implement.
• [x] R2 score
• [x] MSE here. Simply check for NaNs in the target, and put both the target and preds = 0 where there is NaN. Then multiply by N/N_not_Nans
• [x] MAE here. Same as MSE

Unit-tests

• [x] Don't forget to unit-test them to make sure that they give the same results as ignoring the NaNs

[DomInvivo]

Currently working on that with @odymov

[DomInvivo]

A better way to make metrics invariant to NaNs for Accuracy, Precision and Recall!

Accuracy and Recall metrics are invariant to the number of False positive. So everywhere there's a NaN, convert it to a False positive!

# Make all NaNs into false positives for **Accuracy** and **Recall**
nans = torch.isnan(targets)
targets[nans] = 0 # Make it a negative target
preds[nans] = 1 # Make it a positive pred

For Precision, convert all to false negative

# Make all NaNs into false positives for **Precision**
nans = torch.isnan(targets)
targets[nans] = 1 # Make it a negative target
preds[nans] = 0 # Make it a positive pred