Deep-Learning-with-PyTorch

Graphing the positives and negatives
327
High bark Threshold, high Precision
Preston Mostly SlEeps
Figure 12.9 Preston’s choice of threshold prioritizes minimizing false
positives. Cats get left alone; only burglars are barked at!
12.3.3 Implementing precision and recall in logMetrics
Both precision and recall are valuable metrics to be able to track during training,
since they provide important insight into how the model is behaving. If either of them
drops to zero (as we saw in chapter 11!), it’s likely that our model has started to
behave in a degenerate manner. We can use the exact details of the behavior to guide
where to investigate and experiment with getting training back on track. We’d like to
update the logMetrics function to add precision and recall to the output we see for
each epoch, to complement the loss and correctness metrics we already have.
We’ve been defining precision and recall in terms of “true positives” and the like
thus far, so we will continue to do so in the code. It turns out that we are already computing
some of the values we need, though we had named them differently.
Listing 12.1
training.py:315, LunaTrainingApp.logMetrics
neg_count = int(negLabel_mask.sum())
pos_count = int(posLabel_mask.sum())
trueNeg_count = neg_correct = int((negLabel_mask & negPred_mask).sum())
truePos_count = pos_correct = int((posLabel_mask & posPred_mask).sum())
falsePos_count = neg_count - neg_correct
falseNeg_count = pos_count - pos_correct