Major Description
Currently, both ZooImageClassificationPredictor and CustomImageClassificationPredictor transform the resulted Classifications objects into maps of Map<String, Float>. However, the data type of Classifications is a map from the class name to probability with String -> Double, thus the precision is lost in the mapping.
It doesn't seem to be a good transformation in the following points:
- Sometimes rounding from double to float loses not only precision but also accuracy of the value
- The Softmax function used in the classification problems expects the probabilities to sum up to 1.0, but due to the loss of accuracy above the premise seems to be broken to a large extent.
- Finally, probably it's just better to simply return the Classifications object rather than converting it to a Map because the class already provides useful methods such as best(), topK(), etc. so that downstream components can easily extract the conclusions from the output out of the box; if we just provide it as a map, downstream is required to craft some manual calculations upon the result every time.
In conclusion, I think it's better for ZooImageClassificationPredictor and CustomImageClassificationPredictor to simply return the output as Classifications.
Attachments
Issue Links
- links to