Test with an image whose label has 3 classes including background class (i.e. class_0, class_1, class_2) in which only class_0 and class_1 are available.
Expected behavior
The metric calculation should handle division by zero and set a value (i.e., 0) when a class is absent.
Screenshots
Example of the confusion matrix for a batch:
Then, the metrics for class_2, based on the following formula becomes NaN:
Additional context
Small epsilon value can be added in the denominator of each metric calculation formua to avoid devision by zero. Something like the following:
Describe the bug When a class does not show up in a batch, the corresponding row in the confusing matrix becomes zero. As a result, division by zero happens and performance metrics for that class becomes NaN. https://github.com/NRCan/geo-deep-learning/blob/8c8250c2184606919be7c086e56e51a82df6cd39/utils/metrics.py#L130
To Reproduce Steps to reproduce the behavior:
Expected behavior The metric calculation should handle division by zero and set a value (i.e., 0) when a class is absent.
Screenshots Example of the confusion matrix for a batch:
Then, the metrics for class_2, based on the following formula becomes NaN:
Additional context Small epsilon value can be added in the denominator of each metric calculation formua to avoid devision by zero. Something like the following: